Changelog

Upgrade

npx mewkit upgrade

Fresh install: npx mewkit init. See Releasing for the full release process. Section schema: each version uses only the relevant sections from Highlights, New Skills, New Agents, New Commands, CLI, Features, Improvements, Removals, Bug Fixes, Beta.

2.14.5 (2026-07-20) — High-Assurance PR Review

The mewkit CLI ships alongside this kit as 1.19.1 (see CLI below).

Highlights

A new evidence-gated PR-review path. mk:review-pr --assured runs a deterministic ReviewSession: an isolated, SHA-bound, detached review worktree (the user's checkout is provably untouched), one immutable hash-pinned diff, a diff-scoped impact map with targeted scout escalation, a scope-driven reviewer roster, session-observed coverage, and a mechanical verdict gate that cannot emit a PASS/Approve without complete, hook-corroborated coverage. GitHub writes happen only through --reply plus an immediate user confirmation bound to the reviewed revision.

New Commands

• mewkit review prepare <pr-url | owner/repo#n | n> [--remote <name>] — provision an isolated SHA-bound worktree from a PR (fork-safe: pull/N/head fetched from the base-repo remote), capture one immutable hash-pinned diff + PR metadata/CI as untrusted DATA, and write a deterministic impact map + reviewer roster/briefs under tasks/reviews/<session>/.
• mewkit review read --session <id> --as <role> <path> — path-confined, evidence-recording read wrapper (assigned reviewer reads flow through it so coverage is observable).
• mewkit review coverage --session <id> — roster ∩ evidence gap gate; reports evidenceLevel (session-observed vs attested) and exits non-zero on any gap.
• mewkit review compose --session <id> — the mechanical gate: verify the diff hash, re-run coverage, apply the deterministic verdict cap table, resolve inline anchors by snippet, and emit a verdict-gate-compatible proof bundle + SubmitPayload.
• mewkit review submit --session <id> --reply --confirm <payload-hash> — the sole GitHub write path; re-checks the PR head SHA (aborts without posting if it changed) and is idempotent.
• mewkit review cleanup --session <id> — remove the ephemeral review worktree (manifest-nonce-owned); the session audit trail is kept.
• mk:worktree review-pr / review-pr-cleanup — detached, nonce-owned review worktrees; cleanup refuses anything but a matching manifest.

Features

• mk:review-pr gains an opt-in --assured lane (full ReviewSession pipeline) and keeps its fast, read-only default lane unchanged.
• Scope-driven review topology (small / medium / large tiers) with whole-diff roles a per-chunk territory reviewer cannot own, plus heavy-file invariant slices.

Improvements

• mk:review PR mode consumes the prepared session; step-03 runs the coverage gate before verdict composition, and step-04 composes the verdict through mewkit review compose.
• Gate 2 (gate2-check.sh) gains a narrow review-session extension that validates a review verdict's embedded coverage block (hash-clean, no gaps, session-observed required for PASS); inert for non-review verdicts.

Security

• All git/gh invocations in the review pipeline use array-argv execFileSync (no shell). Untrusted PR content is quarantined under untrusted/ and never treated as instructions; the read wrapper is realpath-confined against symlink exfiltration.
• Honest capability limit: subagent tool calls do not reach the parent PostToolUse hook, so session-observed is anti-accidental corroboration (not unforgeable) and proves session-level — not individual-reviewer — access; subagent-driven reviews are attested and cannot earn Approve.

CLI

• The mewkit CLI is bumped to 1.19.1 for the review command group (prepare / read / coverage / compose / submit / cleanup).

2.14.4 (2026-07-19) — The Orientation & Transition Spine

The mewkit CLI ships alongside this kit as 1.19.0 (see CLI below).

Highlights

A fresh session can now recover what it was doing from durable task state alone — no plan or wiki scanning. Explicit activation records one active task plus a canonical pointer, mewkit orient returns a bounded, honest resume envelope (active / none / ambiguous / corrupt-only), and the resume hook injects it through a trust-gated installed runtime with a labeled checkpoint fallback. Trace logging is now one shared append primitive across the shell and CLI, task work is queryable by id, and wiki recall verifies its own index before research trusts it.

CLI

• mewkit orient [--json] — safe resume orientation built only from durable task state; reports active, none, ambiguous, or corrupt-only, and never scans plans or the wiki.
• mewkit plan approve now durably activates the plan by default (one active record + canonical task pointer); --no-activate opts out, and a failed activation exits non-zero rather than claiming success.
• mewkit task new --activate — durably activate no-plan work.
• mewkit task-state update gained --blocker, --verification, --evidence-ref, and --capability-decision, and emits a task-transition trace after a successful write.
• mewkit query --task <id> — task-joined evidence and plan linkage; mewkit query --presets — recovery-measurement metrics (outcome distribution, stale-warning frequency, transitions missing task context, verification re-runs).
• mewkit wiki verify [--json] — canonical-vs-index consistency check (page set, per-page body hash, provenance, FTS parity); research refuses to run against an inconsistent index, and wiki context warns when the index looks stale.

Improvements

• One shared trace-append primitive now backs every writer — the shell hook and the CLI contend on the same sidecar lock, so concurrent appends can no longer interleave or corrupt a line.
• The resume/clear/compact orientation hook runs through a trust-gated installed runtime (absolute, in-project, symlink-resolved executable) and falls back to a clearly labeled checkpoint summary when no runtime is available; fresh-session startup stays disabled pending its release gate.
• Providers surface an explicit, report-only Qwen entry (no support claimed), and the session-orientation instruction is documented in the capability bootstrap.

Migration Notes

• The derived index schema is now v4 (adds task_id / plan_path columns for task-joined queries). Additive — it auto-applies on the next mewkit index / mewkit wiki reindex; no user action required.

2.14.3 (2026-07-19) — Jira & Confluence Agents on Haiku

The mewkit CLI is unchanged at 1.18.2 — this release ships kit content only.

Improvements

• The 16 Jira and 5 Confluence domain agents now pin to the Haiku model tier instead of inheriting the session model, so their CLI-wrapper operations (issue and page CRUD, search, bulk, lifecycle) run on the cheapest tier.

2.14.2 (2026-07-18) — Composed Capability Recall

The mewkit CLI ships alongside this kit as 1.18.2 (see CLI below).

Highlights

When capabilities resolve selects a knowledge-sensitive capability, it now composes a bounded, read-only recall of prior project knowledge directly into the result — the agent reads one knowledgeRecall envelope instead of running a separate wiki lookup. The same activation policy reaches both Claude Code and Codex.

CLI

• capabilities resolve now attaches a knowledgeRecall envelope to a selected capability — a required-class capability gets exactly one bounded wiki recall (≤3 snippets, no page bodies); conditional, none, and every non-selected outcome perform zero automatic wiki access.
• Missing, empty, and errored wiki indexes are now reported distinctly and always fail open — recall never blocks a resolution.
• capabilities resolve --record — opt-in flag that logs one recall decision (counts and status only, never recalled content) through the trace log; default runs stay write-free.
• A capability provided by the running mewkit CLI (invoked via npx or a workspace bin) now resolves as available with "current CLI process" evidence instead of a false "not on PATH".
• mewkit --help now lists the wiki command.

Improvements

• wiki context routes through a shared read-only probe that sanitizes free-text queries into valid full-text-search terms and reads the index without creating -wal/-shm sidecar files.

2.14.1 (2026-07-18) — Harness Integrity & Provider Parity

The mewkit CLI ships alongside this kit as 1.18.1 (see CLI below).

Highlights

This release closes the 2.14.0 harness audit. Gates now prove authority rather than paperwork, the evaluator rejects forged verdicts, provider support is stated from one honest source, Codex migration is default-deny with a measured parity score and a post-migration validator, and the release now gates on its own quality checks.

CLI

• mewkit validate --target codex <dir> — a read-only post-migration gate for a generated Codex project that checks the config/hook wiring, that hook wrappers exist, are executable, and carry the deny contract, that agent TOMLs parse, and that installed skills declare a supported runtime and carry no host-bound tool tokens.
• mewkit providers codex and --lifecycle now render from a single provider-truth source, and both the migration output and README report a skill parity: N% score.
• mewkit migrate codex is default-deny — a portable skill installs, a host-bound skill installs only through a tested adapter or is skipped with a report reason; mewkit migrate codex --include-unportable overrides the policy with an EXPERIMENTAL marker.
• The cross-harness journey (J10) runs a deterministic layer in CI — migration, target validation, then route/artifact/denied-token/side-effect oracles — with no model calls, reporting structural (not semantic) parity.
• The benchmark path now enforces its cost cap: a run halts when cumulative cost reaches the effective cap (--budget or MEOWKIT_BUDGET_CAP override the tier cap).
• mewkit inventory --emit-counts regenerates documentation count tokens from the inventory; npm run release:check aggregates the release-readiness gates for contributors.

Improvements

• Gate 1 now requires a fresh approval receipt bound to the plan revision — plan-presence alone no longer opens it, and editing the plan after approval re-arms the gate.
• Every pre-ship blocked path exits with a structured deny (exit 2), and a missing gate checker fails closed for ship commands.
• The evaluator recomputes its weighted score from the rubric preset and verifies evidence provenance, so a verdict whose score does not follow from its rubrics is rejected.
• Curated memory has a single JSON write path — a write to a non-canonical view is corrected, and a failed capture surfaces a notice instead of skipping silently.
• The release gates on its own checks (tests, lint, typecheck, format, portable validation, inventory, version sync, and plugin-payload drift), versions sync from the root as the single source, and documentation counts generate from the inventory.

Migration Notes

• npx mewkit upgrade to pick up the hardened gates and provider changes — a plan now needs npx mewkit plan approve <plan-dir> after human approval (or MEOWKIT_GATE1_PRESENCE_ONLY=1 to opt out).

2.14.0 (2026-07-17) — Provider Conformance, Contract Safety & Evals

The mewkit CLI ships alongside this kit as 1.18.0 (see CLI below).

Highlights

This release hardens skill contracts and provider portability. Generic skill bodies are cleaned of provider-specific tool names and model IDs so they stay honest per harness, mewkit validate now reports operation-level conformance, and the cook workflow stops after reporting — shipping and reflection are explicit, user-invoked steps rather than automatic.

CLI

• mewkit validate now reports operation-level conformance — it flags a logical operation an installed skill relies on (via its declared tools) that an advertised provider cannot honor, so a cross-harness payload never silently advertises an unsupported capability.
• Skill metadata now carries typed dependency edges, surfaced through the generated inventory.

Improvements

• cook stops after reporting — shipping and reflection are explicit, user-invoked steps, so a run makes no unrequested commit or push.
• Canonical workflow definitions and stricter execution gates across skills, with contract repairs that remove duplicated or conflicting phase ownership.
• Generic skill bodies no longer carry provider tool names, model IDs, or brand tokens; skills genuinely tied to one harness are annotated honestly instead of claiming portability.
• The quick workflow is formalized with a tighter context budget for low-complexity tasks.
• plan-creator is slimmed — visual rendering and lifecycle tasks moved into dedicated skills.
• scale-routing gains a bilingual English + Vietnamese routing eval covering the common task clusters.

Migration Notes

• npx mewkit upgrade to pick up the new skill contracts and the CLI — a cook run that previously auto-shipped now stops after reporting, so invoke shipping explicitly.

2.13.7 (2026-07-15) — Gate Contract Safety, Portability, DX

The mewkit CLI ships separately as 1.17.0 (see CLI below); the two are versioned independently, so this kit release does not imply a new CLI release.

Highlights

Automation can no longer approve its own gates. The contract said Gate 1 and Gate 2 were human-only, but prose across modes, cook, and autobuild had drifted into granting approval authority to a script, a score, or an evaluator stamp — and autobuild's PASS path shipped with no human step at all. The invariant is now stated once, enforced by a lint that fails CI, and backed by a structural Gate 2 check at the commit/push boundary that actually blocks (exit 2) rather than printing "BLOCKED" and proceeding.

New Skills

New Agents

New Commands

CLI

The mewkit CLI (published separately as 1.17.0) adds the enforcement surface:

• mewkit plan status <plan-dir> / mewkit plan check <phase-file> — read-only plan inspection: checkbox progress across plan.md + phase files, and a required-section report for one phase. Writes nothing (the plan Markdown stays the sole authority); there is deliberately no scaffold or edit subcommand.

mewkit plan status tasks/plans/260715-my-plan     # progress across every phase
mewkit plan check  tasks/plans/260715-my-plan/phase-02-thing.md

• mewkit validate --gates — fails on any prose granting automated gate approval, in either word order, plus command-vs-skill drift for the fix / cook / plan / advise dispatchers.
• mewkit validate --parity — regenerates the plugin into a scratch dir and diffs it against the committed plugin/, catching both a forgotten regenerate and a hand-edit.
• mewkit validate now also reports stale README/index counts (previously only mewkit inventory --check).

Features

• Gate Authority Invariant (gate-rules.md) — automation executes between gates and never supplies the authority of one. Verdicts, scores, and validator exit codes are evidence presented at a gate, never a substitute for it. Holds in every mode, including --auto and --fast.
• Structural Gate 2 at the ship boundary (hooks/lib/gate2-check.sh) — on git commit / push / merge, resolves the active plan's verdict and blocks with exit 2 when a ship-capable change has none. Profile-immune: it runs above the fast-profile early-exit, because Gate 2 has no exceptions. Docs/report-only changes take an explicit N/A path rather than a silent skip.
• Logical workflow operations (ask_user, manage_plan, run_shell, delegate_agent) with per-provider conformance in mewkit providers <p> --lifecycle. Kept deliberately out of the frontmatter-reachable invocation enum — that enum's value is what it excludes — and a build-time check fails if the two sets ever intersect.
• Fable 5 advisory model profile — model: fable resolves to the verified claude-fable-5 on Claude Code, and on other providers falls back to the configured heavy tier with a disclosed warning naming both models. Tier-orthogonal: the ModelTier union is unchanged.
• Approval-receipt ADR (docs/architecture/adr/260715-gate2-approval-receipt.md) — designs what would actually prove a human approved, and states plainly why the structural check cannot.

Improvements

• /mk:fix is now a thin dispatcher. It previously described a Simple path that skipped the scout its own skill marks MANDATORY, and dual memory writes the skill forbids.
• mk:scout's Explore workers now carry secret-file guards, a scope fence, and two sections the old prompt never asked for — Risks and Unresolved Questions — plus the A1 status block, so a scout worker terminates with the same vocabulary as every other subagent. Scouting still runs on Claude Code's built-in Explore (Haiku, read-only): a custom executor would be the same model class behind the same subagent boundary, so it was evaluated and not built.
• Plan validation presents a recap — path, objective, scope, phase count, top risks, and the section each answer may change — before the first question, and every question is self-contained.
• README and index counts are regenerated from the canonical inventory rather than hand-maintained.

Bug Fixes

• pre-ship.sh sourced a hook lib without an existence guard; because . is a POSIX special builtin, a missing lib aborted the whole hook — and an aborted hook exits non-2, which is advisory, so the ship proceeded unchecked.
• The git add . && git commit pattern shipped brand-new source files ungated: at PreToolUse the add has not run, and neither a staged nor a HEAD diff shows an untracked file.

2.13.6 (2026-07-13) — Local Visual Plan Review

Highlights

Plans that carry UI can now be reviewed visually and locally. A plan directory gains a structured visual-plan/plan.json artifact — a coverage ledger plus a canvas of surface-locked wireframes, connectors, and annotations — that a deterministic CLI validates and gates, and a transient 127.0.0.1 studio renders and edits. Reviewer edits become immutable feedback batches an agent applies back through a receipted, stale-safe loop. The experimental mewkit orchviz visualizer is retired in favor of this workflow.

CLI

The mewkit CLI (published separately as 1.16.1) adds the surface behind this workflow:

• mewkit visual-plan — new subcommands to validate, gate, review, and edit a plan's visual artifact with no hosted service:

mewkit visual-plan validate <plan-dir>                # schema + coverage closure + refs + safe-HTML + hash freshness
mewkit visual-plan status <plan-dir>                  # coverage summary + review status
mewkit visual-plan approve <plan-dir> --revision <n>  # the Gate-1 transition (single writer of review.status)
mewkit visual-plan export <plan-dir> --format html    # self-contained plan.html from the approved artifact
mewkit visual-plan view <plan-dir>                    # read-only local studio
mewkit visual-plan edit <plan-dir>                    # editable local studio (single-editor lock)
mewkit visual-plan prepare-feedback <plan-dir> --ops <file>   # freeze an immutable feedback batch
mewkit visual-plan apply-feedback <plan-dir> --batch <id>     # apply a batch (--check / --receipt)

• The studio binds 127.0.0.1 only, guards the Host header against DNS rebinding, serves under a strict CSP, and sanitizes wireframe HTML at both save and render; writes use optimistic concurrency via If-Match.
• The studio renders wireframes in a hand-drawn sketch register by default — a scoped design-token theme styles the full .wf-* vocabulary, rough.js redraws each element outline hand-drawn, and a Clean toggle switches to crisp borders; the same theme is embedded in the exported plan.html so studio and export match.

Improvements

• mk:plan-creator gates the structured visual pipeline behind the --html flag — it is opt-in and off by default, so ordinary (backend/refactor) plans stay light. With --html, the plan generates a visual artifact and routes through validate and a human approve at Gate 1 before build.
• mk:visual-plan now defaults to the structured artifact + local studio (generate/validate/view/edit/export); the legacy single-file plan.html template render moves behind --static.

Removals

• Removed the experimental mewkit orchviz visualizer and its web bundle — superseded by mewkit visual-plan and the reusable local-web loopback primitives it shares. No external consumers; the token-estimation heuristic it carried is preserved for mk:context-audit.

2.13.5 (2026-07-12) — Capability Discovery + Durable Task State

Highlights

An installed kit can now tell an agent, at session start, that a capability resolver exists — a bounded, brand-neutral discovery bootstrap surfaces on the SessionStart hook so the agent resolves an intent to the right skill/agent/command instead of guessing. A new contextual rule records durable task state — status, the last completed step, the next action, and capability decisions — so a fresh session resumes a long-running task from files alone.

Features

• Capability-discovery bootstrap — a SessionStart hook emits a small, budget-capped, injection-safe pointer to the capability resolver. It carries generic functional language only; the full capability manifest is never injected into a session.
• Task-state emission rule — orchestration flows record status, the last step, the next action, and capability decisions to a durable per-task record, so a fresh session reconstructs an active task from files without replaying the transcript.

Improvements

• mk:cook, mk:plan-creator, and mk:review reference the task-state emission rule at their orchestration boundaries — emission is advisory and best-effort, so a missing CLI never blocks the workflow.

CLI

The mewkit CLI (published separately as 1.15.0) adds the surface behind these features:

• mewkit capabilities — inspect and resolve the capability manifest (list, explain, resolve --intent, view, bootstrap, projections).
• mewkit context — task-scoped repository-context evidence (resolve, check, record); each path resolves to its own owning repository, so a folder holding many repos never conflates their context.
• mewkit providers --lifecycle — per-provider lifecycle-event matrix + the capability-adapter view (support levels, acquisition, invocation shapes, enforcement gaps).
• mewkit task-state — inspect and update the durable task record (show, update).
• mewkit doctor --consolidation — the consolidation/deprecation ledger (classification, not a runtime-availability claim).

2.13.4 (2026-07-05) — Figma Gateway + Agent-Browser Hardening

Highlights

mk:figma is refactored from a "does everything" skill into an honest read-first gateway. The default path stays lean — design analysis, Figma-to-code for small screens, token extraction, and a screenshot fallback — while advanced operations (Code Connect, canvas writes, and design-system/library patterns) become gated references that load only on explicit intent with confirmed prerequisites. A capability router maps any Figma intent to the right mode or gated reference in one lookup, so advanced work is no longer routed to external skills that may not exist in your environment.

mk:agent-browser expands into a fuller browser-automation gateway — a specialized-workflow router (exploratory QA/dogfooding, Electron desktop apps, Slack automation, and cloud browser providers), a consolidated trust-boundaries safety reference, and scoped MCP tool profiles so a task exposes only the browser tools it needs.

mk:figma also gains a versioned Figma Evidence Packet — a compact, provenance-bearing handoff of design intent that mk:plan-creator turns into viewport/state acceptance criteria and a critical-action validation matrix without re-analyzing the design source. Prototype-flow artifacts make navigation flow reviewable — extracted, inferred, confirmed, or blocked — instead of silently guessed, and mk:visual-plan can render it as an offline interactive flow explorer.

Improvements

• mk:figma advanced workflows (Code Connect, canvas writes, design-system rules) are gated behind explicit intent and confirmed prerequisites instead of loading by default.
• The Figma pre-flight checklist is now per mode — read-only analysis no longer runs the full write-path checklist.
• Token extraction normalizes Figma Variables against Color Styles and surfaces the Enterprise permission caveat for local variables.
• Large or truncated designs fall back to get_metadata with targeted child fetches instead of proceeding on a partial tree.
• The Figma-to-code validation checklist now covers accessibility, interaction states, component reuse, and asset scale.
• mk:agent-browser gains a specialized-workflow router for exploratory QA/dogfooding, Electron desktop apps, Slack automation, and cloud browser providers (Browserbase, AWS AgentCore, Vercel Sandbox).
• A consolidated trust-boundaries reference gathers the safety rules for authenticated, third-party, production, and user-data browser tasks in one place.
• New Slack browser-automation workflows and a dogfood issue taxonomy ship with report templates for QA and Slack analysis.
• MCP tool profiles (core, network, react, and more) let mk:agent-browser expose only the browser tools a task needs.
• mk:figma emits an optional versioned Figma Evidence Packet — design intent, provenance, tokens, states, and a validation contract — as the stable handoff for planning and implementation.
• mk:figma Mode 2 no longer claims standalone production-ready output — production now flows through the packet, plan, browser evidence, and deterministic checks.
• mk:figma captures prototype-flow artifacts with an ambiguity ledger, distinguishing extracted from inferred and blocked navigation instead of guessing from screenshots.
• mk:plan-creator consumes a design evidence packet directly — deriving viewport/state acceptance criteria and a critical-action validation matrix — and never re-analyzes the design source.
• mk:plan-creator blocks only affected phases on unresolved high-risk flow ambiguity via an additive phase field, keeping the overall plan approvable at Gate 1.
• mk:visual-plan renders an optional offline prototype-flow explorer alongside plan.html when a flow graph is present.

Bug Fixes

• Prototype (/proto/) links are no longer sent to get_design_context — the skill asks for the /design/ editor URL, resolving a contradiction between the URL rules and the implement workflow.
• Figma page switching uses the async setCurrentPageAsync API instead of the legacy synchronous assignment.
• Figma variable creation no longer defaults to ALL_SCOPES, which previously polluted property pickers with variables visible everywhere.

2.13.3 (2026-07-04) — Context-Engineering + Prompt-Enhancer Hardening

Highlights

A new mk:context-engineering skill gives the agent a runtime front door for context decisions — what to read, when to stop, when to ask vs assume, and when to compact or delegate — routing to 25 context patterns one reference at a time so the skill that manages context does not itself bloat it. Alongside it, mk:plan-creator gains a Plan Intake Packet step that folds several upstream artifacts into one deduplicated brief before planning starts, and mk:prompt-enhancer is hardened to stay model-agnostic by default — a model-coupled data-separation default is replaced with a neutral fence, a complexity classifier tunes emphasis without changing the portable rewrite, and model-specific steering becomes an opt-in annotation shown only under --analyze when you name a target.

New Skills

Improvements

• mk:plan-creator — Plan Intake Packet. When two or more upstream artifacts already exist (office-hours, brainstorming, planning-engine, confluence-spec, or mk:intake output), a new conditional step consolidates them into a single Plan Intake Packet before research and drafting — so planning starts from one deduplicated brief instead of several overlapping documents. It clean-skips when fewer than two artifacts are present, adding no cost to ordinary plans.
• mk:context-audit now carries a reciprocal boundary note pointing runtime context decisions to mk:context-engineering, so the two context skills disambiguate in both directions.
• mk:prompt-enhancer keeps its default data-separation fix model-neutral — a plain --- DATA --- fence replaces the previous Claude-specific <context> wrapper, so the rewrite stays portable across coding agents.
• mk:prompt-enhancer adds a complexity classifier and per-type recipes — migration, planning, debugging, review, research, long-context, design, and orchestration — that tune output emphasis and length without changing the universal kernel or adding a role.
• mk:prompt-enhancer model-specific steering is now an opt-in target-notes annotation that renders only under --analyze when the input names a target and never alters the portable rewrite.

2.13.2 (2026-07-03) — Deep Research Skill + Full-Surface Codex Migration

Highlights

A new mk:research skill gives you a standalone front door for deep, multi-source technical research that returns one cited report, and mewkit migrate now converts a project's full .claude/ surface to Codex correctly — references inside fenced code blocks are rewritten instead of left as dead links, Codex commands install as skills, and a conversion report flags anything unresolved.

New Skills

Features

• mk:research --html — after the markdown report, author a self-contained editorial HTML report beside it; the markdown stays the source of truth.
• mk:research --deep — split the question across two parallel researcher subagents on independent sub-topics.

CLI

• mewkit migrate rewrites .claude/ references fence-aware — paths inside fenced code blocks are converted instead of left as dead links, and the previous fence-unaware replace that fabricated nonexistent provider paths is gone.
• Codex migration re-enables commands as skills, converts .sh hooks through generated wrappers, preserves matcher coverage where Codex supports it, and records unsupported events instead of dropping them silently.
• Codex migration emits [shell_environment_policy] scaffolds from .claude/.env key names, omits secret-like keys, and surfaces .mcp.json conversion through --include-mcp.
• Codex migration records every migrated, skipped, failed, or narrowed artifact in .codex/migration-report.json and .codex/migration-report.md.
• mewkit migrate reports project_doc_max_bytes guidance for over-budget Codex AGENTS.md output instead of truncating or relying on undocumented nested-file behavior.
• mewkit CLI package version: 1.14.1.

2.13.1 (2026-06-29) — Wiki Handoff + Context Recall

Highlights

Knowledge-producing skills can now hand a finished artifact to the wiki — it is scanned and turned into a proposed candidate with full provenance only when it clears the salience gate — and any flow can recall prior knowledge through a disciplined, DATA-bounded wiki context surface. The capture path stays gated end to end: agents still only propose, and a human approve remains the only route to a canonical page.

Features

• mewkit wiki handoff suggest|propose|profiles — turn a skill's terminal artifact into a scanned candidate, driven by a per-skill registry (class required / conditional / none). suggest is read-only; propose rejects sensitive or out-of-root artifact paths before reading.
• mewkit wiki context "<keywords>" — disciplined recall returning ranked pages with a project-root-readable path and a snippet; bodies only with --include-content, and it fails open when no index exists.
• wiki propose now accepts full provenance — --origin, --source-id (repeatable), --reuse-scope, --verification-state, --risk-score, --review-after, --novelty-delta, --salience-json — instead of hardcoding origin and salience.
• New wiki-context-rules rule marks all wiki content as DATA and bounds the Phase 0 recall probe.

CLI

• Documented the full runtime command surface in the CLI reference — wiki, index, query, inventory, trace, pack, providers / explain-support, build-plugin, verdict-gate, and budget context.

Improvements

• wiki hint and wiki search now return a project-root-readable path (tasks/wikis/<slug>/pages/…) instead of an internal page id.

Migration Notes

• The derived index schema advances to version 3 (additive wiki_handoff + wiki_candidate_source tables). It auto-applies on the next npx mewkit index or npx mewkit wiki reindex — no user action and no data loss.

2.13.0 (2026-06-29) — The Wiki Knowledge Subsystem

Highlights

MeowKit gains a gated, long-term knowledge base. Agents and reviewers can capture provenance-bearing project knowledge, search it with full-text search, and render it to a self-contained HTML page — under a security model where agents may only propose content and a human approval (which always re-scans) is the only path to a canonical page.

New Skills

CLI

• New mewkit wiki command — init · propose · approve · reject · search · hint · list · render · reindex · enqueue · research.
• The derived SQLite index is consolidated into a single rebuildable wiki-index.db (trace + cost + wiki tables + FTS5); mewkit index and mewkit query are preserved.

Features

• Anti-self-poisoning write model — agents propose candidates only; canonical pages are written solely through mewkit wiki approve, which always re-runs the scanner.
• Every write passes secret-scrub plus a multi-pass injection scan; rejected content is quarantined with an intervention and a trace event.

Beta

• mk:wiki-research external fetch is opt-in and disabled by default — sources are url-guarded against SSRF, size-capped, and redirect-re-validated per hop, and fetched content can only ever become a candidate.

2.12.3 (2026-06-28) — Visual + HTML Workflow

Highlights

Brainstorm reports and approved plans now render to self-contained HTML on demand, and five complementary visual skills join the mk: namespace — diagrams, long-form reading, shareable showcases, and text-to-UI design.

New Skills

Improvements

• mk:brainstorming --html — after the markdown report, authors a self-contained editorial HTML report beside it (opt-in; markdown stays source of truth).
• mk:plan-creator --html — after Gate 1 approval and task hydration, renders the plan to a shareable plan.html via mk:visual-plan (opt-in; markdown plan files stay source of truth).

Removals

• Removed mk:preview --html --plan-review — plan-as-HTML rendering is now owned solely by mk:visual-plan (mk:visual-plan <plan-dir>). mk:preview keeps --explain, --diagram, --slides, and --diff.

2.12.2 (2026-06-28) — Visual Plan Rendering

Highlights

A plan directory (plan.md + phase-*.md) now renders into ONE self-contained, shareable plan.html a reviewer can scan in under 30 seconds — phases, architecture, decisions, risks, and files-touched, without reading prose. Consistency comes from a fixed template, one shared theme, and a disciplined block vocabulary, so two unrelated plans look the same. It is prompt-only and self-contained; the only external dependency is the pinned Mermaid CDN, with a visible source fallback when offline.

New Skills

2.12.1 (2026-06-27) — Native Plugin Distribution

Highlights

MeowKit is now installable as a native plugin in both Claude Code and Codex (mk), alongside the existing flat-copy install. The plugin distribution is generated from the same .claude/ source — no second source of truth — so skills (/mk:*), agents (mk:*), and hooks all resolve under the plugin. The flat-copy mewkit init path is unchanged.

New Skills

CLI

• mewkit build-plugin — generate the native plugin distribution (plugin/ payload + Claude and Codex marketplaces) from .claude/.
• mewkit validate --plugin — guard namespace purity, plugin-manifest contract, and manifest-version alignment; runs in CI.
• mewkit doctor now reports the active install mode and warns when flat-copy and the plugin are installed together.

Features

• Native plugin install for Claude Code (claude plugin install mk@meowkit) and Codex (codex plugin add mk@meowkit).
• Hook parity in plugin mode — settings.json hooks are translated into plugin/hooks/hooks.json, so gates and safety hooks fire under the plugin too.

Improvements

• The harness substrate now reports full coverage — a new intervention-recording rule tags the previously-uncovered responsibility, bringing mewkit inventory --substrate to 12/12.

2.12.0 (2026-06-27) — Legible Context + Queryable Substrate

Highlights

The outer harness gains a legible context layer and a queryable substrate across sessions. A new phase × tier read-budget tells any inner harness how much context to load per phase; every harness artifact now carries a vendor-neutral responsibility so coverage is auditable; and an on-demand trace surface recalls repeated friction, drift, and cost aggregates from the existing append logs. Everything is additive, advisory, and inner-harness-agnostic — no gate, verification, or hook behavior changes.

CLI

• mewkit trace score|audit|propose — on-demand trace-quality tier, entropy/drift audit, and advisory grouping of repeated friction over the append logs.
• mewkit trace --friction "<note>" --responsibility <r> — record friction with no inner-harness hook required.
• mewkit inventory --substrate — print a responsibility×coverage matrix generated from the registry; --emit writes the committed view.
• mewkit index and mewkit query — opt-in derived SQLite index over the append logs (rebuild-able and disposable; the logs stay canonical) with read-only aggregate queries.
• mewkit validate --substrate — flag untagged registry artifacts and a stale substrate view, now run in CI.

Features

• New context-budget-rules — a vendor-neutral 7-phase × 3-tier read-set and token-ceiling matrix with retrieval triggers, so an agent loads the smallest right context per phase.
• Every harness artifact now carries a responsibility tag (eleven Runtime-Substrate values plus gate-enforcement), so the dead-weight audit can target by responsibility instead of by file.

Improvements

• The Phase-6 analyst agent now prefers mewkit index + query + trace audit for cross-run cost and pattern rollups instead of hand-parsing the JSONL logs, and falls back to the raw log when the index is absent.

Beta

• mewkit index / mewkit query are opt-in and use the experimental node:sqlite builtin — the derived index is disposable and never the source of truth.

2.11.7 (2026-06-20) — Brainstorming Solution Decompression

Highlights

mk:brainstorming now handles the case where you arrive holding a solution instead of a problem. When the input is itself a solution ("just build X", a preselected feature, a roadmap item), the skill first decompresses it into the underlying problem and three alternative problem framings before generating any ideas — and redirects to mk:office-hours when the problem's value is unvalidated. It stays read-only and never grades evidence inline.

Improvements

• mk:brainstorming decompresses a solution-input to the underlying problem and three or more problem framings before ideating, through a new solution-decompression pre-step technique.
• The pre-decided-user gotcha and the repeated-redirect edge case now recover the problem first, then optionally stress-test the chosen path with reverse.
• Brainstorm reports can be rendered as a self-contained HTML brief through mk:preview --html --explain.

2.11.6 (2026-06-19) — Merge Conflict + Codebase Architecture Skills

Highlights

A new on-demand skill, mk:resolving-merge-conflicts, walks an in-progress git merge or rebase to a clean finish. It reads the conflict state, recovers the original intent behind each side from commit history and PRs, resolves every hunk — preserving both intents where they are compatible — then runs the project's checks and completes the merge or rebase. It always resolves rather than aborting, and never invents new behavior.

A second on-demand skill, mk:improve-codebase-architecture, reviews a codebase for deepening opportunities — shallow modules whose interface nearly matches their implementation — emits them as structured candidates for mk:preview to render, then grills a chosen candidate into a precise, type-safe patch. It owns analysis and patch emission only; all rendering is delegated to mk:preview.

New Skills

2.11.5 (2026-06-18) — Long-Horizon Run Hardening

Highlights

The problem this release fixes: on long, multi-phase plans (and autonomous runs), the context window fills up and gets reset or compacted — and the agent could lose track of which phase it was on, redo finished work, or skip steps.

2.11.5 makes that progress durable on disk instead of trusting in-context memory. After every phase, mk:cook writes a mandatory checkpoint, mk:plan-creator records how much latitude the agent has so it survives a cold start, and plans are now split into runnable end-to-end slices so the project always works between phases. No new features — this hardens how state survives a reset.

Improvements

• Resume on the right phase after a reset. mk:cook now writes a checkpoint after every completed phase — it is mandatory, not best-effort. The checkpoint lands in two durable places (the phase-file checkboxes and the plan.md Agent State) before moving to the next phase, so a brand-new session reads the disk and knows exactly where to continue.
• One source of truth for resume state. Phase-file checkboxes and frontmatter are now the canonical record of progress. .plan-state.json is just a derived cache — regenerated at the final sync-back, never hand-edited mid-run — so the three places that used to drift can no longer disagree.
• Autonomy limits that survive a cold start. For long-horizon plans, mk:plan-creator writes an Autonomy Boundaries block: a simple three-tier guide (Always allowed / Ask first / Never) for how far the agent may act on its own. It is written into the plan so it is not lost on reset. Fast, trivial, and single-phase plans skip it to avoid paying the token cost forever.
• The agent reads those limits according to how you run it. At startup mk:cook applies the block by mode: an interactive run asks you before any significant reversible choice; an autonomous run takes the safe default and logs it instead of stopping. If the block is missing, it degrades gracefully — proceeding on reversible, in-scope changes.
• Each phase delivers a working end-to-end path. mk:plan-creator now splits work into vertical slices by default (one complete feature path per phase) instead of horizontal layers (all DB, then all API, then all UI). The system stays runnable after every phase — which matters most when a run is interrupted partway. Horizontal phases are kept only for foundations like migrations and shared types.
• Smarter phase splitting. Two new signals trigger a split: an "and" in a phase title, or two clearly separable deliverables. This works within the existing rule of at least two and at most seven phases.

Migration Notes

• Run npx mewkit upgrade to pick up the updated mk:cook and mk:plan-creator behavior.
• No CLI package version bump is required for this release because the changed commits do not modify packages/mewkit/src.

2.11.4 (2026-06-18) — Post-Compaction Safety Re-Arm

Highlights

Native compaction no longer leaves a session running without its safety baseline. The PreCompact hook now re-arms the per-session verification marker, so the next turn re-reads the safety rules whose text compaction dropped — instead of trusting a stale "already verified" marker and skipping the re-read.

Bug Fixes

• Safety baseline is no longer skipped after compaction — compaction drops the rule text from context but used to leave the per-session "verified" marker intact, so the PreCompact hook now re-arms that marker to force a fresh re-read on the next turn, and the cached-skip resumes only once a genuine re-read re-verifies.

Improvements

• Long autonomous runs gain an observation-budget rule — summarize tool output to the relevant finding on receipt and drop or compress the oldest observations once new ones exceed the budget, rather than carrying full raw output across turns.

2.11.3 (2026-06-14) — Vue 3 Skill Suite

Highlights

Vue support grows into a full skill suite. mk:vue is restructured from a single patterns file into focused references and now covers the Pinia Colada data layer plus Vue 3.4/3.5 core APIs, <script setup> compiler macros, and advanced reactivity patterns. Two invoke-only advisors join it — mk:vue-best-practices for deep best-practices review and the ordered authoring workflow, and mk:vue-testing-best-practices for Vue test design and test-code review.

New Skills

Improvements

• mk:vue is split from one patterns file into focused references for components, composables, reactivity, routing, and state.
• mk:vue gains Pinia Colada data-layer coverage — core queries/mutations, advanced patterns, plugins, and direct cache access.
• mk:vue adds Vue 3.4/3.5 coverage — new core APIs, <script setup> compiler macros, and advanced reactivity patterns.
• The status line is simplified — heavy rendering and caching logic removed for a leaner, more reliable output.

Migration Notes

• Run npx mewkit upgrade to pick up the new skills.
• No CLI package version bump is required for this release because the changed commits do not modify packages/mewkit/src.

2.11.2 (2026-06-13) — Grill Plan-Interrogation Skill

Highlights

One new pre-planning skill joins the catalog. mk:grill relentlessly interviews you about your own plan or design — one question at a time — resolving each branch of the decision tree until you both share the same understanding. It is the inverse of the other elicitation skills: Claude is the interviewer and you are the interviewee. Every answer is checkpointed to docs/knowledge/<slug>.md so progress survives a context-window reset, anything you cannot answer is flagged as a gap to chase down, and the session ends by handing the front-loaded context to mk:skill-creator, mk:plan-creator / mk:brainstorming, or mk:project-context.

New Skills

Migration Notes

• Run npx mewkit upgrade to pick up the new skill.
• No CLI package version bump is required for this release because the changed commits do not modify packages/mewkit/src.

2.11.1 (2026-06-11) — Ask-Me Project Q&A Skill

Highlights

One new cross-cutting skill joins the catalog. mk:ask-me answers factual and explanatory questions about the current project — "how does X work here", "why is X structured this way", claim-checks — with cited file:line evidence instead of a long grep session. It is read-only by default, enforces hard read budgets per depth tier, separates verified claims from inference with a mandatory confidence level, and redirects ideation, decisions, debugging, and reviews to the owning specialist skill rather than partially answering.

New Skills

Migration Notes

• Run npx mewkit upgrade to pick up the new skill.
• No CLI package version bump is required for this release because the changed commits do not modify packages/mewkit/src.

2.11.0 (2026-06-11) — Context Tiers + Harness Guardrails

Highlights

This release tightens the harness's always-on context and adds guardrails across it. The intent → skill routing table moved out of always-loaded rules, context-budget reporting now splits by load tier, hook-registration drift is caught in both directions, skills can declare their external-service contract, and mk:help now ranks your next steps with rationale.

New Commands

• /mk:help — thin wrapper for the forward-looking navigator that ranks your next steps.

CLI

• mewkit budget context now reports an always-on / conditional / on-demand tier breakdown; --json gains a tiers field alongside the existing top-level totals.
• mewkit validate --rules — new scoped check that warns when a broad intent → skill routing table lives in an always-loaded rule.
• The mewkit package version for this release is 1.10.7.

Features

• mk:help now emits a ranked top-3 of next steps with rationale, reading checkpoint, budget, detected-model, review-verdict, and roadmap state with graceful handling when a source is absent.
• Skill frontmatter gains three advisory tool-contract fields — requires_external_service, default_enabled, and stable_output_contract — surfaced by the frontmatter validator as warnings.
• New benchmark audit mode — git-index-audit.sh records a reproducible git tracked-state fingerprint (tracked-path and index SHA-256, file and directory counts, optional two-checkout comparison) as a JSON artifact.

Improvements

• The intent → skill routing table moved out of always-loaded agent-routing.md into an on-demand mk:agent-detector reference, trimming always-on context.
• mewkit budget context no longer counts conditional rules as always-on.
• Hook-registration validation now detects drift in both directions — a registered hook missing on disk, and a handlers.json entry with no file — not only unregistered scripts.
• Brand-prose lint gains a diff mode that fails on violations in changed files while treating pre-existing violations in untouched files as warnings.
• Team-mode coordination detail moved out of always-loaded rules into a mk:team-config reference, loaded only when team mode activates.

2.10.3 (2026-06-10) — Memory + Plan Completion Cleanup

Highlights

Fix memory capture is now JSON-only, and plan completion no longer waits for the ship phase. Completed plan todos now close the plan lifecycle immediately.

CLI

• POST /api/plan/todo now marks a plan complete and moves it to tasks/plans/archive/ when the final non-abandoned phase todo is checked.
• The mewkit package version for this release is 1.10.5.

Improvements

• mk:fix and mk:cook now direct bug-class learnings to .claude/memory/fixes.json only.
• Plan sync-back now treats completed task checkboxes as the source of truth for plan completion.
• Manual /mk:plan archive remains confirmation-based, while completion-driven archiving runs only after every plan todo is checked.

Bug Fixes

• Fix capture no longer creates an unnecessary duplicate Markdown memory entry.
• Finished plans are no longer left active until /mk:ship runs.

Migration Notes

• Run npx mewkit upgrade to pick up the updated workflow instructions and plan lifecycle behavior.

2.10.2 (2026-06-02) — PR Review + Response Skills

Highlights

Two single-purpose PR-collaboration skills join the reviewer's toolkit. mk:review-pr gives a fast shallow verdict on someone else's PR, and mk:respond-pr triages the comments you receive with verify-before-agree discipline. Both are read-only on code and only write to GitHub when you opt in with --reply.

New Skills

Migration Notes

• Run npx mewkit upgrade to pick up the two new skills.
• No CLI package version bump is required for this release because the changed commits do not modify packages/mewkit/src.

2.10.1 (2026-06-02) — Brainstorming Rigor + Prompt Recipes

Highlights

Two planning-adjacent skills get tighter boundaries. mk:brainstorming now challenges its own idea set before recommending, while mk:prompt-enhancer gains recipe-mode routing for architecture-review and research-style prompts without turning into a reviewer or researcher.

Features

• mk:prompt-enhancer architecture-review recipe — --analyze --deep can rewrite a draft prompt so it asks for architecture findings with severity, evidence, and decisions-needed while the enhancer itself emits no findings.
• mk:prompt-enhancer research recipe — discovery prompts now get grounding, attention-anchored ordering, and explicit acceptance criteria while leaving actual research to mk:scout or the downstream agent.
• mk:prompt-enhancer recipe canaries — architecture-review and research prompt canaries guard against role confusion, alongside updated eval-suite guidance for manual LLM-judged checks.

Improvements

• mk:brainstorming now runs a challenge pass before recommendation, checking duplicate architectures, hard-constraint violations, category diversity, conservative drift, and missing stakeholder failure modes.
• mk:brainstorming now tracks context budget explicitly, keeping scout bridge summaries, rejected alternatives, and handoff packets compact instead of passing full transcripts downstream.
• mk:brainstorming reports now self-check for problem, binding constraint, success criterion, excluded scope, technique choice, challenge result, recommendation evidence, and handoff completeness.
• mk:prompt-enhancer mode routing is now centralized, with default, analyze, score, deep, architecture-review, and research behavior documented in one reference.
• mk:prompt-enhancer --deep clarifies that the local scanner is a bounded hint source and broad repository mapping remains mk:scout.

Migration Notes

• Run npx mewkit upgrade to pick up the updated skill contracts, reference docs, and canaries.
• No CLI package version bump is required for this release because the changed commits do not modify packages/mewkit/src.

2.9.14 (2026-05-30) — Autobuild Rename + mk:loop

Highlights

The autonomous green-field build pipeline is now mk:autobuild — a name that says what it does — and a new mk:loop skill drives one measurable metric (coverage, bundle size, lint count, latency) toward a target through bounded, git-tracked iterations that keep wins and revert regressions.

New Skills

Removals

• mk:harness is renamed to mk:autobuild — the same planner → contract → generator ⇄ evaluator pipeline, clearer name. Invoke it as /mk:autobuild; /mk:harness no longer resolves.
• MEOWKIT_HARNESS_MODE is renamed to MEOWKIT_AUTOBUILD_MODE — values MINIMAL|FULL|LEAN are unchanged.

Migration Notes

• Run npx mewkit upgrade to pick up the renamed skill, command, and rules.
• Replace any /mk:harness invocation with /mk:autobuild.
• Replace MEOWKIT_HARNESS_MODE with MEOWKIT_AUTOBUILD_MODE in your shell config or CI. MEOWKIT_BUDGET_CAP and --budget are unchanged.

2.9.13 (2026-05-30) — Fix Gate Parity & Workflow Evidence Index

Highlights

mk:fix now follows the same no-self-approval discipline as the rest of the workflow, and a new MeowKit-native workflow evidence index gives mk:fix, mk:cook, mk:review, and mk:ship a single traceable record of what ran — without ever approving anything itself. Gate 1 and Gate 2 remain human authority.

Features

• Workflow evidence index — mk:fix and mk:cook populate one workflow-evidence.json per run indexing task, risk flags, diagnosis or contract, verification, verdict path, and approvals; it is traceability only, carries no score, and mirrors validate-gate-1.sh / validate-gate-2.sh rather than replacing them.
• A zero-dependency validator (.claude/scripts/validate-workflow-evidence.cjs) checks evidence completeness before approval prompts; the contract lives in .claude/rules-conditional/workflow-evidence-rules.md and loads on demand, adding no always-on context.
• mk:ship pre-flight reads the evidence index when present — after mk:verify, which keeps abort authority — and surfaces approvals, verdict, and verification as proof; absent evidence ships exactly as before.

Improvements

• mk:fix --auto no longer self-approves from a review score — it auto-fixes blocking issues up to the cycle limit, then stops at ready for user approval, and crossing Gate 2 or a ship boundary always requires explicit human approval.
• mk:fix adds a six-field root-cause proof checkpoint (HARD GATE) before implementation — exact symptom, deterministic reproduction, expected vs actual, root cause with file:line, why now, blast radius — with a compact one-phrase form for --quick.
• mk:cook records evidence across Phase 0-6 from outputs that already exist with no new phase or gate, and --auto wording is clarified — automatic execution between gates, never automatic Gate 2 approval.
• mk:review records the verdict path in the evidence index; the verdict and side-effect behavior are unchanged.

Migration Notes

• If you relied on mk:fix --auto proceeding past review automatically on a high score, you now confirm at the ready for user approval checkpoint — no flag change is required, only the approval step is now explicit.
• Standalone standard/complex mk:fix runs now require the six root-cause fields before a fix is applied; simple and --quick fixes use the compact form and are otherwise unaffected.

2.9.12 (2026-05-30) — JSON-first memory + observability cleanup

Highlights

Curated project memory is now JSON-first — the .json stores are the canonical, schema-validated source and human-readable Markdown is generated from them. The conversation-summary subsystem is retired, so long-session continuity relies on the host runtime's native compaction, and observability consolidates on a single canonical trace stream.

CLI

• mewkit memory validate — schema-validate the curated memory stores; --strict exits non-zero on errors.
• mewkit memory seed-from-md — populate the JSON stores from existing Markdown topic files, run once after upgrading.
• mewkit memory render-views — regenerate the human-readable views/*.md from the canonical JSON.
• mewkit verdict-gate <slug> — validate a review verdict's machine-readable proof bundle; blocks on a blocked, invalid, or missing verdict.
• mewkit doctor --state now reports curated-memory health alongside the runtime-state checks.

Features

• Memory consumers read the .json store first and fall back to Markdown only when it is absent, warning when the two diverge.
• Security findings are now a curated JSON store — security-findings.json.
• mk:review emits a schema-validated tasks/reviews/<slug>-verdict.json proof bundle alongside the human verdict.

Improvements

• trace-log.jsonl is documented as the single canonical event stream — reader-less debug telemetry is gated behind MEOWKIT_HOOK_DEBUG.
• Generated Markdown views carry a "generated — do not edit" banner so the JSON stays the source of truth.

Removals

• Removed the conversation-summary subsystem — the /mk:summary command, the summary cache hook, and the MEOWKIT_SUMMARY_* environment variables are gone, and long-session continuity is now handled automatically by the host runtime's native compaction.
• Removed the inert cost-meter.sh hook — per-call cost tracking is handled by the budget tracker.
• Removed the archived lessons.md and deprecated patterns.json stubs.

Migration Notes

• npx mewkit upgrade to pick up the new memory layout.
• Run mewkit memory seed-from-md once to migrate existing Markdown memory into the JSON stores.
• If you relied on /mk:summary or any MEOWKIT_SUMMARY_* variable, remove them — continuity is now automatic.

2.9.11 (2026-05-24) — Plan-creator mode clarity

Highlights

Plan creation now has clearer mode and flag boundaries. --deep is documented and wired as bounded phase-scoped planning, while --tdd remains an independent opt-in flag for regression-first execution handoff.

Improvements

• mk:plan-creator --deep now documents bounded scope maps and per-phase Deep Phase Maps instead of implying unrestricted repository scanning.
• mk:plan-creator --tdd now documents tdd: true, regression_gate, Tests Before, Protected Change, Tests After, and Regression Gate as the regression-first phase contract.
• TDD handoff now keeps strict execution explicit by printing cook commands with --tdd when the plan was created with TDD enabled.
• Cook now warns when a plan contains TDD markers but execution starts without --tdd or MEOWKIT_TDD=1.
• Scout, brainstorming, and planning boundaries are clearer: scout finds context, brainstorming chooses an approach, and plan-creator turns a concrete approach into executable phases.
• New Core Concepts page "Plan Creator Modes and Flags" explains planning cost, context impact, mode selection, composable flags, lifecycle boundaries, examples, migration notes, and anti-patterns.
• Plan-creator reference docs now describe --deep and --tdd using the same bounded, runtime-neutral terms as the skill implementation.

Migration Notes

• Use npx mewkit upgrade to pick up the updated skill contracts and docs.
• Keep using /mk:plan --deep --tdd only when both risks apply: broad scope and existing behavior that must be preserved.
• If a TDD plan prints a cook command with --tdd, run that exact command so RED-first execution is enforced.

2.9.10 (2026-05-23) — Skill portability + cook context-engineering + provider diagnostics

Highlights

Three workstreams ship together. First, four skills (mk:docs-finder, mk:autobuild, mk:plan-creator, mk:scout) gain a meowkit: frontmatter block declaring portability policy (portability, providers, requires.{surfaces,commands,env}, context_cost) — the first-pass schema that lets the migrate subsystem decide whether each skill can install onto a non-Claude provider. Second, the migrate CLI grows provider-contract diagnostics — a new provider-contract-diagnostics module surfaces support gaps per provider, wired into npx mewkit doctor --providers, npx mewkit doctor --state, and npx mewkit validate --portable; first-pass skill providers expand to Codex, Gemini CLI, Antigravity, and OpenCode in addition to Claude Code. Third, mk:cook gains three new context-engineering contracts — a scout-first contract at Phase 0 (3–6 bullet codebase summary surfaced before any clarifying question), an exact-requirements contract at Phase 1 (5 dimensions must be answered before plan-creator returns), and a no-side-effects regression-recovery pattern at Gate 2 (2–4 typed options presented to the user when a regression is surfaced instead of silently patching). All changes are additive — pre-upgrade plans, verdicts, sessions, and skills without the meowkit: block continue to work unchanged.

Features

• meowkit: frontmatter block on skill SKILL.md — first-pass portability schema with fields portability (generic / provider-adapted / provider-only), providers.{include,exclude}, requires.{surfaces,commands,env}, and context_cost (low / medium / high). Parsed by parsePortabilityPolicy() in skills-discovery.ts and consumed by portability-policy.ts to decide install eligibility per provider. Skills without the block continue to migrate under the existing pre-policy heuristics.
• First-pass skill providers — portability-policy.ts enables skill installation for Codex, Gemini CLI, Antigravity, and OpenCode in addition to Claude Code. Skills with providers.include: [claude-code] only continue to install for Claude Code; other providers receive a PortabilitySkip with reason "needs review before non-Claude install".
• 4 first-pass skills carry the new metadata — mk:docs-finder (provider-only, claude-code, surfaces: [skills], commands: [Bash], context_cost: medium), mk:scout (provider-only, claude-code, commands: [Agent, Grep, Glob], context_cost: medium), mk:plan-creator (provider-only, claude-code, commands: [Agent, AskUserQuestion, Bash], context_cost: high), mk:autobuild (provider-only, claude-code, commands: [Agent, AskUserQuestion, Bash], env: [CLAUDE_PROJECT_DIR], context_cost: high).
• New module migrate/provider-contract-diagnostics.ts — collects provider-support diagnostics with pass / warn / fail severities, exposes collectProviderContractDiagnostics() + summarizeProviderContractDiagnostics().
• New cook contracts — Scout-First Contract (Phase 0 codebase-summary presentation) and Exact-Requirements Contract (Phase 1 plan-creator must answer expected output / acceptance criteria / scope boundary / non-negotiable constraints / touchpoints), with skip-on-plan-path notes. Plan-creator SKILL.md gained a matching Requirements Capture Contract section enumerating the 5 dimensions and the scout-grounded-options rule.
• New cook regression-recovery contract — Side Effects Detected: Yes reviewer signal recognized by validate-gate-2.sh. Blocks Gate 2 until a ## User Decision Addendum block (User selected: + Resumption point:) is appended. Positive-presence-only — absence of the field is never a block signal, so pre-upgrade verdicts validate unchanged.
• New "Regression Recovery Options" subsection in mk:cook/references/review-cycle.md — 5-step procedure with 4 standard options (revert + re-plan / keep + update dependents / compatibility shim / accept the regression).
• New top-level shared rule .claude/rules/anti-rationalization.md — generic implementation-phase rationalizations migrated from mk:cook/SKILL.md, which retains a 1-row TDD-specific extension.

CLI

• npx mewkit doctor --providers — runs provider-contract diagnostics alongside the standard doctor checks. Wraps the new collectProviderContractDiagnostics() output into the doctor report.
• npx mewkit doctor --state — emits state-taxonomy diagnostics.
• npx mewkit validate --portable — adds portable-provider contract checks to the validate pipeline.
• Top-level CLI help (mewkit --help) now lists the three new flags above.

Improvements

• mk:cook/SKILL.md Workflow Modes table demoted from 5 columns to 3 (Mode / Research / TDD) — canonical 6-column mode-behaviors matrix lives in mk:cook/references/intent-detection.md. Gate 2 absolutism statement collapsed to a 1-line callback to .claude/rules/gate-rules.md across SKILL.md, intent-detection.md, workflow-steps.md (review-cycle.md preserved inline for iteration-loop context).
• TDD column labels unified to RED-strict / Plan-level / Skip across cook SKILL.md, intent-detection.md, workflow-steps.md.
• Phase 5 ship subagent renamed in cook SKILL.md, workflow-steps.md, and subagent-patterns.md from git-manager via mk:ship to shipper via mk:ship — shipper orchestrates the full pre-ship pipeline and invokes git-manager internally.
• mk:cook/references/subagent-patterns.md gained a Standard Delegation Skeleton (9-field block per orchestration-rules.md) at the top and a Scope: pass X / do NOT pass Y annotation on each of the 13 template sections (isolation-boundary contract per orchestration-rules.md Isolation Boundaries).
• validate-gate-1.sh runs as an advisory preflight in non-auto cook modes (surfaces structural-check failures to the user before Gate 1 prompt; user retains override) and remains a blocking gate in --auto mode.
• Cook SKILL.md Mermaid diagram edge labels quoted (|"--verify"|, |"--strict"|) for renderer compatibility.
• mk:cook/references/failure-catalog.md gained 4 new failure modes — skipped scout summary, vague clarifying questions, scout-first gate fired on plan-path input, silent patch on review-detected regression.
• migrate/portability-policy.ts widens skill-skip signal detection — .claude paths, CLAUDE.md references, mk / meow slash commands, Claude env vars, Anthropic env vars, and orchestrator-semantics references all trigger skip-with-reason when migrating to non-Claude providers.
• core/smart-update.ts and core/compute-checksums.ts refined for the new portability flow.
• migrate/migrate-orchestrator.ts integrates the new portability policy + diagnostics path.

Bug Fixes

• USD pricing on cook --verify / --strict flags removed from cook SKILL.md, intent-detection.md, workflow-steps.md — replaced with [LIGHT] / [HEAVY] relative-cost labels and a one-line variability note (concrete cost depends on inner harness, model tier, and target surface).
• Inner-harness Mermaid rendering preamble removed from cook SKILL.md (diagram authority is independent of rendering fidelity).
• Phase 4.5 subagent column reframed from agent-browser or curl to browser-automation subagent or HTTP verification tool — concrete subagent name depends on installed skill set.
• ToC anchors in mk:cook/references/workflow-steps.md and subagent-patterns.md updated to match renamed Gate 2 / Phase 4.5 / Ship section headers.

Migration Notes

• npx mewkit upgrade to pick up the new gates and CLI flags.
• Skills WITHOUT a meowkit: frontmatter block continue to migrate under existing pre-policy heuristics — no immediate action required. To opt-in to first-pass portability, add a meowkit: block declaring portability, providers, requires, and context_cost; see the 4 first-pass skills for examples.
• New cook gates (scout-first, exact-requirements) apply only when starting from a fresh task description. Plan-path invocations (/mk:cook tasks/plans/.../plan.md) skip both gates explicitly.
• The new validate-gate-2.sh side-effect signal is positive-presence-only — existing verdicts without Side Effects Detected: Yes continue to pass unchanged.

2.9.9 (2026-05-23) — Plan-creator determinism + handoff

Highlights

mk:plan-creator gains a deterministic post-approval handoff and two consistency-sweep gates so red-team and validation-interview edits cannot leave the rest of the plan inconsistent. After Gate 1 + task hydration, the skill fires an AskUserQuestion whose options are pruned to the gates not already auto-run in the active mode (cook / validate / red-team / harness / end), writes the choice to plan.md frontmatter as a typed enum, and STOPs without auto-invoking. The validator now rejects unknown handoff.next values. Static AskUserQuestion JSON in seven step / reference files becomes a "Recommend When" decision table; runtime-templated payloads stay as JSON because they carry substitution variables a table cannot express.

Features

• New step file step-09-post-plan-handoff.md — runs after task hydration. Mode-pruned options (≤4), live risk re-scan over plan content for risk-checklist.md trigger keywords, defensive fallback advisory when matched_flags is unset, writes handoff: { next, decided_at } to plan.md frontmatter.
• New gates Whole-Plan Consistency Sweep — Gate W1 at the end of step-05-red-team.md, Gate W2 at the end of step-06-validation-interview.md. Stage-then-apply algorithm: read-only Pass 1 stages ### Pending Sweep Edits block; decision check blocks on unresolved contradictions via AskUserQuestion; write Pass 2 applies edits, writes summary, updates consistency_sweeps.{red_team|validation} frontmatter. Recursion bound: 2 resolve attempts per gate. See references/whole-plan-sweep.md.
• New step-04 sub-step Verification Roles — Fact Checker / Flow Tracer / Scope Auditor / Contract Verifier dispatched as READ-ONLY parallel Explore subagents, tier auto-selected by phase count (Light 1–2 / Standard 3–4 / Full 5+). Orchestrator aggregates verdicts and writes ONE ## Verification Log Edit per phase file. Skipped in fast / product-level modes. FAILED claims surface as targeted step-06 interview questions. See references/verification-roles.md.
• validate-plan.py enforces handoff.next enum when the handoff: frontmatter block is present. Plans without handoff: continue to validate.
• .plan-state.json v1.2 — additive schema bump adding optional verification_tier and consistency_sweeps_passed. v1.1 readers ignore unknown keys; v1.2 readers default missing keys.

Improvements

• Seven static AskUserQuestion JSON blocks across step-00-scope-challenge.md, step-07-gate.md, references/gate-1-approval.md, and references/archive-workflow.md converted to "Option / Recommend When / Why" decision tables matching the ck-plan Post-Plan Handoff pattern. Runtime-templated and looped payloads (per-finding loop, file-content embedding, fully generated question text) stay as JSON.
• references/phase-template.md documents ## Verification Log and ## Validation Log as machine-written, optional sections; hand-edit anti-pattern added.
• references/gate-1-approval.md Context Reminder block moved from step-07 to step-09 (fires AFTER the user picks a next-step option, not on Gate 1 approval); contradictory step-07 instruction reconciled.

Migration Notes

• Legacy plans without handoff, consistency_sweeps, ## Validation Log, or ## Verification Log continue to validate as PLAN_COMPLETE. No migration tool required.
• Fixtures live at tests/fixtures/plan-creator/ covering legacy / modernized / partial / interrupted-sweep / handoff-invalid-enum states; runnable via bash tests/fixtures/plan-creator/validate-fixtures.sh.

2.9.8 (2026-05-23) — Memory system deep fix

Highlights

Four verified failures in the memory pipeline removed: last-model-id.txt now persists the real model id, /mk:fix reliably writes its Step 6 entry, the ##pattern: / ##decision: / ##note: capture surface is finally documented as a user-typed keyboard shortcut (the handler never fired on agent output), and analyst.md writes to the v2.4.1 split topic files instead of the deprecated patterns.json / lessons.md stubs.

Bug Fixes

• last-model-id.txt always wrote unknown — post-session.sh was reading env vars (CLAUDE_MODEL, MEOWKIT_MODEL_HINT, ANTHROPIC_MODEL) that Claude Code never exports to Stop-hook subprocesses. New hooks/lib/resolve-model.sh resolves the model id from session-state/detected-model.json (the canonical source) with env-var fallback. cost-log.json entries and trace-log.jsonl session-end events now carry the real model id.
• /mk:fix Step 6 memory write path is now enforced — the command spec previously had no memory write instructions and Phase 6 / Reflect only ran in /mk:cook. New entries land in both fixes.md and fixes.json (with frequency: 1 and today's lastSeen).
• Agent files instructing ##pattern: / ##decision: / ##note: capture were architecturally broken — the handler is bound to UserPromptSubmit and never fires on agent output. 41 files patched (23 agents + 16 skills + 1 command + docs/memory-system.md) to use direct Edit calls instead. Existing user-typed ##prefix: flow is unchanged.
• analyst.md agent wrote to patterns.json (deprecated stub) and lessons.md (archived stub) — Phase 6 entries were functionally lost. Now targets fixes.{json,md}, review-patterns.{json,md}, architecture-decisions.{json,md} per the v2.4.1 schema.

Improvements

• New reference doc .claude/skills/memory/references/capture-architecture.md — the canonical 2-path contract describing user-typed ##prefix: vs agent-authored direct Edit. Referenced from every patched agent and skill.
• docs/memory-system.md §3a clarified that the ##prefix: handler does NOT fire on agent output; new §3d added describing the agent-direct write path; §8 tombstone gains rows for "Agent-output ##prefix: as API" and "patterns.json / lessons.md as active write targets".
• mk:memory SKILL.md gains a "How session-capture is invoked" section clarifying the routine is prose-driven (the agent reads references/session-capture.md and follows its steps via Edit) — there is no mewkit memory session-capture CLI subcommand.
• New tests/resolve-model.test.sh — 7 cases covering happy path, missing file, malformed JSON, literal "unknown" fall-through, env-hint override, file-wins-over-env precedence, and shell-metachar sanitization.
• New tests/spec-drift.test.sh — regression gate ensuring no agent/skill/command file re-introduces ##prefix: as an agent-output instruction or writes to the deprecated patterns.json / lessons.md stubs.

Removals

• patterns.json and lessons.md remain on disk as gravestone stubs; their write-target status was already removed in v2.4.1 and is now also enforced via spec-drift.test.sh. Scheduled for deletion in v2.12.0 — external skills writing to either should migrate to the split topic files before then.

2.9.7 (2026-05-16) — Docs reference contract + validator

Highlights

.claude/ shipped into downstream projects used to leak path references into the kit's own internal docs (docs/dead-weight-audit.md, docs/red-team-overview.md, docs/memory-system.md, etc.) that do not ship via the release zip. This release formalizes the two-tier docs-reference contract, patches every leaked path, and adds a validator that prevents future regressions. Downstream projects no longer carry broken cross-references inside their installed .claude/.

CLI

• npx mewkit validate — new check: "Docs references on Type-1 allowlist". Walks every shipped .claude/ subdir, extracts docs/<path> references, and verifies each against the Type-1 allowlist in .claude/rules/docs-reference-contract.md. Exit code 1 on any error.

Features

• New rule .claude/rules/docs-reference-contract.md — codifies the two-tier resolution model (target-project docs vs kit-internal docs), the Type-1 allowlist (parsed from a fenced <!-- ALLOWLIST-START --> block as the single source of truth), and the skill-author checklist.
• New rule .claude/rules/dead-weight-audit-rules.md — 8 numbered rules covering audit cadence, the benchmark-driven measurement process, decision thresholds (PRUNE / WATCH / KEEP), components that never prune, anti-patterns, and the circular-dependency workaround when the harness itself is under test.
• New rule in skill-authoring-rules.md Rule 5 — Docs References Follow the Two-Tier Contract; enforced mechanically by the validator.
• New CI step Validate docs references on Type-1 allowlist — runs python3 .claude/scripts/check-docs-references.py .claude on every PR.
• New CI step Unit-test docs-references validator — runs the 15-case parser/matcher test suite.

Improvements

• 14 path references inside .claude/ were re-pointed at in-kit rule files or inlined directly into the consuming context — covers benchmark/SKILL.md, harness/references/adaptive-density-matrix.md, harness-rules.md, agents/reviewer.md, commands/mk/summary.md, memory/architecture-decisions.md, hooks/HOOKS_INDEX.md, scripts/lint-brand-prose.sh, benchmarks/README.md, three attribution citations (prompt-enhancer, elicit, jira), and one template placeholder (plan-creator/references/phase-template.md).
• Five false-positive source rewrites cleaned up: slash-separated category list in scale-routing/SKILL.md, CSV example in lazy-agent-loader/SKILL.md, example filename in agent-browser/references/video-recording.md, attribution prose in jira-issue/references/issue-templates.md, heading text in plan-ceo-review/references/post-review.md, plus two scheme-prefixed URL fixes in hooks/HOOKS_INDEX.md and hooks/lib/read-hook-input.sh.
• Phase-template placeholder ../../../docs/relevant-doc.md becomes <path-to-related-doc-in-your-project> so new phase files emit a clean placeholder by default.

Removals

• Bare docs/relevant-doc.md from the Type-1 allowlist — the phase-template now uses an angle-bracket placeholder, so the literal allowlist entry is no longer needed. No migration action required.

2.9.6 (2026-05-16) — Context isolation: SessionStart budgeting, agent-detector sentinel, memory auto-prune

Highlights

SessionStart and per-message context costs cut without changing the safety contract. docs/project-context.md gains a byte cap before injection; the agent-detector skips its 10-file rule re-read on turns 2..N via a session-scoped sentinel; memory .md topic files auto-prune entries older than 90 days on Stop. Combined: ~9,700 tokens saved per turn after turn 1 on a typical session, plus a hard cap at SessionStart on repos with oversized project-context.md.

New Commands

• /mk:delegate — assembles a context-isolated subagent prompt from the orchestration-rules.md template. Wizards through 7 required fields, runs an injection-pattern guard on each user-provided field, and emits a ≤200-token prompt block for review before the Task tool is invoked. --silent mode skips the interactive wizard for inner harnesses that do not support AskUserQuestion. Distinct from mk:spawn (which launches parallel agents in isolated worktrees).

Features

• Env var MEOWKIT_MAX_PROJECT_CONTEXT_BYTES — pre-injection byte cap on docs/project-context.md at SessionStart. Default 12288 (~12 KB ≈ ~3K tokens). Files over the cap are truncated with a visible warning before cat; files over 2× the cap emit a stronger warning. 0 disables the cap entirely.
• Env var MEOWKIT_SKIP_SAFETY_SENTINEL — controls the agent-detector cached-marker optimization. Default on skips the 10-file safety + phase-zero rule Read on turns 2..N of a session. off forces the full check every turn.
• Env var MEOWKIT_MEMORY_PRUNE — auto-prune of stale .md topic-file entries on Stop. Default on. Entries older than 90 days with a parseable YYYY-MM-DD header date are removed; entries with severity: critical or severity: security are NEVER pruned; entries without a parseable date are NEVER pruned (protects manual notes).
• Env var MEOWKIT_MEMORY_PRUNE_AGE_DAYS — overrides the 90-day cutoff for memory auto-prune.
• Memory directory size audit — project-context-loader.sh now reports any .claude/memory/*.md file exceeding 500 lines in the agent-readiness banner.

Improvements

• orchestration-rules.md gains three new sections — Isolation Boundaries (5-row table of what crosses each layer and what does NOT), Inner Harness Compatibility (capability matrix: Task tool surface, context auto-injection root, SubagentStart hook availability), and Rejected Patterns (deliberate non-implementations with reasoning).
• mk:agent-detector step-0 and step-0b first scan the context window for a cached-sentinel marker before invoking the 10-file Read loop. The marker is emitted by a new UserPromptSubmit handler when a session-scoped JSONL log records both flags safety and phase_zero as true. The skill remains allowed-tools: [Read] — all filesystem ops live in hook context.
• post-session.sh model-change detection now writes ## YYYY-MM-DD — dead-weight-audit-needed (auto-flagged) headers so the auto-prune script can parse and eventually rotate them.
• Single state file session-state/session-sentinels.jsonl replaces per-session-ID sentinel files. Truncated on session change.
• session-state/prune-log.md records pruned counts as {file} | {date} | {N} entries pruned — no entry content, breaking the injection-rules.md Rule 11 carrier chain for memory data.

2.9.5 (2026-05-14) — mk:worktree script backing and new commands

Improvements

• mk:worktree is now backed by a Node.js CLI (scripts/worktree.cjs + scripts/lib/worktree-git-helpers.cjs). All 6 commands emit structured JSON (--json) and support --dry-run for safe previewing.
• New status command: health audit of all active worktrees showing ahead/behind divergence and dirty state for each.
• New prune command: removes stale .git/worktrees/ metadata left by manually-deleted worktrees; always dry-run first.
• New --orchestrated flag on create: places the worktree at .worktrees/{agent-name} on a parallel/{agent-name}-{timestamp} branch per parallel-execution-rules.md Rule 3.
• New references/commands.md: full option tables, JSON field docs, and error code reference for all 6 commands.
• SKILL.md updated to document all 6 actions with script invocations, a ## Gotchas section, and compliance with skill-authoring-rules.md.

2.9.4 (2026-05-11) — Agent rules and TOON agent docs

Improvements

• Added a docs reference for MeowKit agent rules covering status handoffs, context loading, agent naming, search-before-building, and file ownership.
• Converted uniform .claude/agents matrices to TOON blocks so agent registries, operations tables, safety tiers, and review checklists stay compact without changing their stated values.
• Updated release documentation to include the 2.9.4 release entry.

2.9.3 (2026-05-11) — Optimize rules

Apply Token-Oriented Object Notation (TOON) to rules. Dropped ~26,288 → ~23,234 tokens (~3,054 saved, 11.6%)

2.9.2 (2026-05-11) — Spec-to-tech-breakdown orchestrator

Highlights

/mk:breakdown ships as a thin skill-composing command that takes any spec source — Confluence URL, Jira KEY, GitHub issue, Linear ticket, file, or paste — and produces refined user stories, optional per-story Jira tickets, and a codebase-aware tech breakdown report. Stops cleanly before plan creation so the user runs /mk:plan later with the report as context. Closes the manual 4-skill chaining gap between spec analysis and planning.

New Commands

• /mk:breakdown <source> [--source <type>] [--project <KEY>] — orchestrates mk:confluence-spec-analyst (or mk:intake for non-Confluence sources) → story decomposition → per-story AskUserQuestion gate → mk:jira-issue create per acceptance → mk:planning-engine plan --tickets → consolidated tasks/reports/breakdown-{YYMMDD}-{slug}.md. Per-story options: Create / Edit-then-create / Skip / Skip-all-remaining. Honors EXT_SYSTEM gate — human confirmation precedes every Jira write.

Improvements

• /mk:meow Step 2 heuristics table routes "break down spec", "decompose spec", "spec to stories", and breakdown <source> phrasing to the new command.
• Footer logic emits one of five state-specific next-action messages — covers Jira-not-configured, --project missing, partial spec (Confluence child-page failures), tech breakdown skipped, and the happy path with cross-link to the planning report.

2.9.1 (2026-05-11) — Brand-prose neutralization for migrate targets

Highlights

mewkit migrate codex|kiro|cursor now strips narrative "MeowKit" / "Claude Code" / non-citation "Anthropic" prose at converter time and substitutes the target name. The merged Codex/Kiro AGENTS.md header no longer leaks "MeowKit" branding. Toolkit identifiers (mk:, MEOW_*, npx mewkit, hook event names) stay verbatim. A markdown-only lint guard prevents source regressions, gated in CI on every PR.

CLI

• md-strip converter accepts a new targetName option and substitutes "Claude Code" → target display name when set (e.g. "Codex", "Kiro IDE", "Cursor"); falls back to "the host runtime" otherwise.
• fm-strip merged-agents header rewritten — > Ported from the toolkit via mewkit migrate replaces the hardcoded "MeowKit agents" string.
• Citation-aware skip preserves <!-- research-citation --> marker, research / thesis keyword windows, and blockquote-style citations.
• All stripClaudeRefs callers (md-to-mdc, md-to-kiro-steering, fm-strip) now pass the provider display name through.

Features

• New lint guard .claude/scripts/lint-brand-prose.sh (orchestrator) + .claude/scripts/check-anthropic-context.py (single source-of-truth check) — runs in under 1 second on the full tree.
• CI step Lint brand prose in .claude markdown blocks PRs that reintroduce banned narrative prose.
• New .claude/.brand-allowlist.txt for toolkit-internal index/log headers, with a 30-entry ceiling.
• New docs/branding-style-guide.md — single-page reference with before/after examples and the documented On Claude Code, … factual-prefix form.

Improvements

• Source cleanup across .claude/ markdown — narrative "MeowKit" / "Claude Code" prose neutralized in rules, skills, commands, hook references, and benchmark canaries. ~47 hits rewritten.
• Code emitters statusline.cjs, scripts/memory-migrator.cjs, and hooks/project-context-loader.sh use neutral strings instead of "MeowKit" in user-visible output.
• One non-citation "per Anthropic's documented behavior" phrase in rules/skill-authoring-rules.md rewritten to "per the runtime's plugin contract".

2.9.0 (2026-05-11) — Pre-ticket story sizing

Highlights

mk:story-sizer ships as MeowKit's first pre-ticket sizing skill — accepts a paste-mode markdown block of stories, produces per-story Fibonacci complexity + rough sizing, and optionally batch-creates tickets via mk:jira-issue and mk:jira-collaborate (--internal audit comments) with mandatory dry-run + single confirmation gate. Closes the spec → ticket-creation gap without violating the "AI provides signals, humans decide" design. v1 ships paste-only; spec/intake adapters are deferred.

New Skills

New Agents

• story-sizer — Heuristic complexity + Fibonacci scoring agent. Read-only at the Jira side — delegates create/comment to mk:jira-issue and mk:jira-collaborate.

Features

• --paste input adapter with strict markdown template and per-story AC validation.
• Optional --scout extracts codebase context from session (matches mk:planning-engine extract-or-prompt pattern; never auto-invokes).
• Advisory split detection for stories sized 13+ points.
• Audit comments use --internal flag on mk:jira-collaborate add-comment (team-only, not customer-facing).
• Audit comment template customizable via MEOWKIT_STORY_SIZER_COMMENT_TEMPLATE env var.
• Workflow page updated — spec-to-pr-walkthrough gains Step 3.5.

Improvements

• The 10-step Spec → PR walkthrough now has a concrete pre-ticket sizing step, not just a feasibility-scan handwave.

2.8.7 (2026-05-11) — Agile/Scrum Rule Layer

Highlights

Adds a thin, conditionally-loaded Agile/Scrum governance layer covering Definition of Ready, Definition of Done with Jira sync, sprint-goal persistence, mid-sprint amendment ceremony, sprint-close hygiene, retro→action loop, and time-boxed spike governance. Three rules in .claude/rules-conditional/agile-*.md load via mk:agent-detector Step 0b ONLY when an Agile context is detected. Non-Agile sessions pay zero context cost — no new agents, no new hooks, no new gates.

New Conditional Rules

Skill Integrations

• mk:plan-creator — DoR advisory at Phase 1 entry; new --spike --timebox <duration> flag with two-phase spike template (assets/spike-plan-template.md); rejects --spike --product-level and --spike + mk:autobuild FULL
• mk:sprint-contract — new sprint-goal subcommand (set / show / align) writing to tasks/contracts/sprint-state-{date}-sprint-{N}.md with flock discipline; new assets/sprint-state-template.md
• mk:retro — post-narrative action-item ceremony parsing ## 3 Things to Improve and ## 3 Habits for Next Week; per-item AskUserQuestion (Create Jira story / Add to plan TODO / Document as no-action / Defer)
• mk:jira-agile — sprint add/remove post-start: flock + amendment append + reason prompt; sprint close: per-ticket disposition prompt + closure summary
• mk:ship — Gate 2 PASS path early-return guard (preserves zero-cost for non-Agile) + 3 opt-in DoD prompts (verdict→Jira comment, status transition, business AC checkbox); covers the mk:cook → mk:ship chain in one insertion
• project-manager agent — Jira-aware status enrichment via single mk:jira-search JQL aggregation when Agile context active

Agile Context Detection (OR-logic)

Any one match triggers load of all 3 rules at mk:agent-detector Step 0b:

• Glob tasks/contracts/sprint-state-*-sprint-*.md returns ≥1 result
• Active plan frontmatter has non-empty jira_tickets:
• MEOW_JIRA_BASE_URL env var set
• Last user message matches [A-Z]{2,10}-\d+

Pre-flight check defends against drift: if .claude/rules-conditional/ directory is absent, the load is skipped silently.

Architecture Notes

• Plan-as-source-of-truth retained; Jira receives projections only (verdict comment, status transition) — never authoritative
• Sprint-state contract path uses leading sprint-state- prefix to prevent glob collision with the existing check-contract-signed.sh glob
• Spike plans are INCOMPATIBLE with mk:autobuild FULL density (harness gate breaks); use mk:cook or mk:plan-creator --fast
• Concurrent-write safety: mk:jira-agile and mk:sprint-contract sprint-goal set MUST acquire flock before any read-modify-write on sprint-state files

ADR: docs/architecture/adr/260511-agile-scrum-conditional-rule-layer.md. Plan: plans/260510-2333-agile-scrum-rule-layer-design/.

2.8.6 (2026-05-10) — Align rules

Improvements

• Generalized generated runtime assets and aligned rule prose across .claude/rules/, .claude/agents/, and .claude/commands/mk/ so wording, role labels, and cross-references match the post-2.8.5 namespace and folder layout.

2.8.5 (2026-05-10) — Rules Folder Reconsolidation + mk:preview

Highlights

Two threads land together. The five Phase-Zero conditional rule files move back into .claude/rules/, simplifying the rule-availability validator and the upgrade orphan-cleanup pass. Alongside that, mk:preview ships as a native MeowKit visual generator — markdown and self-contained HTML for explanations, diagrams, slide decks, git diffs, and plan rendering — with a mandatory theme toggle, anti-slop guarantees, and zero new package dependencies.

New Skills

Features

• 9 modes total — 4 markdown (--explain, --diagram, --slides, --ascii) + 5 HTML (--html --explain, --html --diagram, --html --slides, --html --diff, --html --plan-review).
• Mandatory theme toggle on every HTML artifact, persisted via localStorage with OS-preference fallback on first load.
• Zoom and pan engine for diagram pages — mouse wheel cursor-anchored zoom, drag pan, keyboard + − 0.
• Slide engine for --html --slides — scroll-snap + IntersectionObserver + arrow-key nav + bottom progress dots.
• Plan-aware output paths — tasks/plans/{slug}/visuals/ when an active plan exists, tasks/visuals/ fallback. Reads session-state/active-plan; handles both absolute-path and slug formats; falls back loudly with stderr warn:.
• Style rotation — palette and typography vary per run via ${CLAUDE_PLUGIN_DATA}/preview/style-rotation.json; deterministic-hash fallback when env var is unset.
• 4 clean-room HTML templates in assets/ — architecture, mermaid-flowchart, slide-deck, data-table — each carrying a clean-room provenance comment.
• New embedded reference references/mermaid-essentials.md (≤ 120 lines) — pinned Mermaid v11.4.1, init pattern, .node class trap, layout-direction guidance, color rules. Loaded by 4 other Mermaid-emitting skills via a one-line directive (mk:cook, mk:plan-ceo-review, mk:problem-solving, mk:jira-relationships).
• --html --diff PR mode pipes gh pr diff through .claude/hooks/lib/secret-scrub.sh before HTML interpolation; topic strings are HTML-encoded per element / attribute / CSS / JS context.

Improvements

• The five Phase-Zero conditional rule files (agent-routing.md, model-selection-rules.md, phase-contracts.md, risk-checklist.md, scale-adaptive-rules.md) move from .claude/rules-conditional/ back to .claude/rules/. They keep the same Phase-0 conditional load semantics (read explicitly by mk:agent-detector Step 0b — NOT auto-loaded by Claude Code's directory mechanism), but live alongside the always-on rules so the orphan-cleanup pass and the rule-availability validator see them as a single tree.
• scripts/validate-rule-availability.sh regex simplified — drops the (-conditional)? group now that the legacy path is gone.
• mk:cook, mk:plan-ceo-review, mk:problem-solving, mk:jira-relationships each gain a one-line directive pointing at mk:preview/references/mermaid-essentials.md before any Mermaid emission.
• SKILLS_INDEX.md registers mk:preview under "Cross-Cutting (Any Phase)".

Bug Fixes

• mewkit upgrade orphan-cleanup pass was blind to .claude/rules-conditional/ — DEFAULT_ORPHAN_SCOPES only listed rules, skills, agents, hooks. After this release, the scope list also covers rules-conditional as a one-time migration sunset, so users upgrading from v2.8.3 / v2.8.4 get the legacy directory flagged and removed on the next upgrade. New regression test in tests/find-orphans.test.ts locks the behaviour.

Migration Notes

• Run npx mewkit upgrade to pick up the merged layout. The orphan-cleanup pass will detect leftover files under .claude/rules-conditional/ and prompt to delete (skip with --no-cleanup).

2.8.4 (2026-05-10) — Confluence Ecosystem + Macro-Aware Spec Analysis

Highlights

mk:confluence returns as a 6-skill ecosystem (1 hub + 5 leaves) backed by the confluence-as CLI. The differentiator is mk:confluence-spec-analyst — it fetches pages as ADF (atlas_doc_format) and pipes the body through a new macro-aware walker so panels, decisions, task lists, mentions, expand sections, and Smart Link / Figma embeds survive as explicit markdown labels (> [INFO], > [DECISION], - [ ], <details>, @name, ![alt](attachment:<id>)). Macro labels become first-class signals for downstream gap detection. The 2.8.3 removal is reversed: there is now a first-party Confluence read path, with mk:planning-engine --spec consuming the resulting Spec Research Report.

New Skills

New Agents

• confluence-page — Page CRUD + hierarchy + version + copy/move agent.
• confluence-search — CQL search + saved-filter agent with unconditional sanitization.
• confluence-spec-analyst — Read-only spec analyst with macro-aware ADF fetch path and persisted reports.
• confluence-bulk — Bulk operations agent with mandatory dry-run + typed-token confirmation.
• confluence-collaborate — Per-page collaboration agent (comments, attachments, labels, watchers) with footer-default discipline.

Features

• New macro-aware walker confluence/scripts/adf-to-md.sh + adf_to_md.py — preserves panel / decisionList / decisionItem / taskList / taskItem / expand / mention / media / inlineCard / blockCard nodes with explicit ASCII labels. Two modes: stdin (pipe an ADF body) or --page-id <N> (single-call fetch via the wrapper).
• Single-call fetch pattern in mk:confluence-spec-analyst — extracts metadata + body from one ADF response, halving per-page API request count and staying inside the Cloud rate limit (10 req/s).
• Per-child failure handling — root-page fetch failure aborts the run; child fetch failure appends [INCOMPLETE: child <id>] and continues with the partial corpus.
• Unknown ADF nodes emit [UNHANDLED_NODE: <type>] with keys-only metadata — type name + sorted attribute keys, never raw text or attribute values. Closes the prompt-injection vector through custom macros.
• Mention display names are escaped + scanned against the prompt-injection pattern list before emission; matched names are redacted to [REDACTED_DISPLAY_NAME].
• New env vars MEOW_CONFLUENCE_API_TOKEN, MEOW_CONFLUENCE_EMAIL, MEOW_CONFLUENCE_SITE_URL in .claude/.env. The wrapper scripts/confluence-as.sh translates them to confluence-as's native CONFLUENCE_* names per call and refuses any plaintext-credential fallback in .claude/settings.local.json.
• Cloud-only gate — wrapper exits 3 on non-*.atlassian.net URLs and points at the Atlassian MCP escape hatch.
• mk:planning-engine accepts --spec <report-path> pointing at a mk:confluence-spec-analyst report; the Planning Report gains a ## Spec Context section.
• mk:intake recognizes Confluence URLs (*.atlassian.net/wiki/spaces/...) and raw page IDs as a 5th source.
• Walker fixture meowkit/.claude/skills/confluence/scripts/tests/fixtures/adf-macros.json + adf-macros.expected.md for deterministic regression checks.
• 11 new VitePress reference pages (5 agents + 6 skills) grouped under "Confluence Agents" and "Confluence (Router + Family)" in the sidebar.

Improvements

• Children traversal cap lowered from 10 to 5 in mk:confluence-spec-analyst to stay inside the Confluence Cloud per-user rate limit under the single-call-per-child pattern.
• ADR 260510-mk-confluence-ecosystem-path-b.md extended with an addendum documenting the macro-aware fetch path, the ADF coupling rationale, and why a 7th mk:confluence-convert skill was rejected (wrapping a lossy upstream function inherits the loss).
• confluence/references/cli-idioms.md gains a "Local conversion helpers" section + macro label convention table.
• confluence/references/safety-framework.md adds new failure-mode rows for the walker's exit codes and a sanitization-addendum note.
• confluence-spec-analyst/references/spec-analysis-patterns.md gains 6 macro-label heuristics (decisions without rationale → INFO; task items without owner → GAP; panels with weasel words → AMB; mentions without context → INFO; media without alt → GAP; [UNHANDLED_NODE] → Open Question).

Migration Notes

• Run npx mewkit upgrade to pick up the new defaults — this auto-installs confluence-assistant-skills (CLI distribution) into .claude/skills/.venv via the existing requirements.txt.
• Populate .claude/.env from .claude/.env.example with the three MEOW_CONFLUENCE_* vars.
• Reverse direction (md → adf) is deferred — no consumer in MeowKit currently calls page update --representation atlas_doc_format --content-file <json>. Re-add when a real consumer surfaces.
• The 2.8.3 removal advice ("use Atlassian MCP directly") is superseded — mk:confluence is the first-party path again.

2.8.3 (2026-05-10) — Jira Family + Workflow Discovery

Highlights

mk:jira becomes a pure router. Jira execution moves into 16 thin leaf skills — 13 domain plus 3 intelligence — each forking a dedicated agent in .claude/agents/. The Atlassian-MCP coupling is replaced with the jira-as CLI (auto-installed via npx mewkit setup). Workflow templates are no longer assumed: a new fetch-workflow.sh discovers each instance's actual statuses + transitions and caches them under tasks/jira-workflows/ for mk:jira-lifecycle, mk:jira-bulk, mk:planning-engine, and mk:jira-evaluator to consume.

New Skills

New Agents

• jira-issue — JIRA issue CRUD via the jira-as wrapper.
• jira-search — JIRA search + saved-filter management.
• jira-lifecycle — Workflow lifecycle with cache-first transition discovery.
• jira-collaborate — Comments, attachments, watchers, notifications.
• jira-relationships — Issue link / unlink / blocker / dependency / clone.
• jira-time — Time tracking + worklog management.
• jira-agile — Sprint / epic / backlog / story-point management.
• jira-fields — Custom field discovery + admin configuration.
• jira-bulk — Bulk operations with mandatory dry-run discipline.
• jira-jsm — JIRA Service Management across 8 sub-domains.
• jira-admin — Highest-blast-radius admin agent with 2-step token confirmation on destructive ops.
• jira-dev — Developer artifact generator (branch / PR / commit linking).
• jira-ops — jira-as cache + project-context diagnostics.
• jira-evaluator — Read-only complexity + inconsistency analysis (persists report).
• jira-estimator — Read-only heuristic story-point estimation (persists report).
• jira-analyst — Read-only ticket + media analysis (persists report).

Features

• New script scripts/fetch-workflow.sh — discovers a project's actual workflow via admin workflow for-issue <KEY> (admin path) with non-admin fallback to lifecycle transitions <KEY>. Cache layout: tasks/jira-workflows/_schemes/<PROJECT>.md (project → workflow mapping) + tasks/jira-workflows/<workflow-slug>.md (full statuses + transitions).
• New env vars MEOW_JIRA_API_TOKEN, MEOW_JIRA_EMAIL, MEOW_JIRA_SITE_URL in .claude/.env. The wrapper scripts/jira-as.sh translates them to jira-as's native JIRA_* names per call and sets JIRA_OUTPUT=json as default.
• New SessionStart hook jira-env-loader.sh — validates .claude/.env presence + the 3 required keys; emits [mk:jira] env OK / [mk:jira] <KEY> missing to the status line.
• New env var CLAUDE_CODE_FORK_SUBAGENT=1 set in settings.json env block by default — enables context: fork semantics for the 16 thin leaves.
• New pip dependency manifest meowkit/.claude/skills/jira/scripts/requirements.txt — npx mewkit setup auto-installs jira-as (and the jira-as[keyring] extra) into .claude/skills/.venv.
• 32 new VitePress reference pages — one per leaf skill + one per agent — grouped under "Jira (Router + Family)" and "Jira Agents" in the sidebar.
• 30 high-signal references adopted from upstream JIRA-Assistant-Skills (issue templates, JQL patterns, workflow patterns, comment templates, field-types reference, voodoo-constants, ITIL workflows, branch-naming + smart-commits guides) — all rewritten to use the jira-as wrapper instead of upstream Python scripts.

Improvements

• mk:planning-engine and mk:intake no longer reference Atlassian MCP — both now route through the mk:jira-* family. mk:planning-engine gotchas + mk:intake/references/jira-{awareness,handoff-protocol}.md updated accordingly.
• All 16 jira-* agents conform to MeowKit conventions: Required Context (MeowKit) block (loads docs/project-context.md), Skill Rule of Two self-classification, Memory (MeowKit convention) block using ##pattern: / ##note: / ##decision: prefixes against typed memory files. The 3 intelligence agents persist their reports to tasks/reports/jira-{evaluate,estimate,analyze}-*.md for cross-session continuity.
• The jira-estimator agent verifies the per-instance Story Points field ID at runtime (fields list --search "Story Points") instead of guessing — customfield_10016 is the documented default but the agent always confirms.
• Pattern files at mk:jira-lifecycle/references/patterns/{standard,software-dev,jsm-request,incident}-workflow.md are now labeled "CONCEPT REFERENCE — NOT AUTHORITATIVE" and point at the discovered workflow cache as the source of truth.

Removals

• mk:confluence skill removed — 9 files deleted, 3 cross-references updated. Migration: confluence reads no longer have a first-party path; if needed, invoke Atlassian MCP directly per mk:jira/references/install-and-auth.md.
• Documentation cleanup follow-up: deleted orphaned website/reference/skills/confluence.md reference page and pruned lingering cross-references from docs/project-context.md, docs/meowkit-architecture.md, and website/reference/skills.md index.
• The legacy mk:jira Atlassian-MCP execute path is gone. The router contains zero Agent(...) calls and zero jira-as invocations.
• The 3 intelligence agents previously at skill-scoped meowkit/.claude/skills/jira/agents/ move to project-scoped meowkit/.claude/agents/jira-{evaluator,estimator,analyst}.md. The old directory is deleted.

Migration Notes

• Run npx mewkit upgrade to pick up the new defaults — this auto-installs jira-as into .claude/skills/.venv via the new requirements.txt.
• Populate .claude/.env from .claude/.env.example with the three MEOW_JIRA_* vars.
• Pre-existing scripts that called /mk:jira evaluate KEY should switch to /mk:jira-evaluator KEY (and similarly for estimate / analyze).
• mk:confluence users: there is no drop-in replacement. Use Atlassian MCP directly, or read pages manually before invoking mk:jira-issue for ticket creation.

Breaking Changes

• mk:jira is no longer an executor — it only routes. All previous /mk:jira <verb> invocations route to the appropriate mk:jira-* leaf instead of executing inline.
• mk:confluence removed entirely.
• Required env vars: MEOW_JIRA_API_TOKEN, MEOW_JIRA_EMAIL, MEOW_JIRA_SITE_URL in .claude/.env. Without them, the SessionStart hook surfaces [mk:jira] <KEY> missing and wrapper invocations error out.

2.8.2 (2026-05-09) — Prompt Enhancer Output Modes

Highlights

mk:prompt-enhancer default output is now the rewritten prompt only — no analysis, no preamble, paste-ready into any coding agent. The new --analyze flag opts back into the four-section diagnostic surface (decomposition, detected issues, source-grounded suggestions), and --score adds a deterministic 1–10 quality rating on the original prompt with a transparent rubric breakdown.

Improvements

• mk:prompt-enhancer default mode emits only the rewritten prompt code block — decomposition and weakness detection still run internally to feed the rewrite, but are no longer displayed unless requested.
• New --analyze flag — surfaces the original-prompt analysis (5-component decomposition, FOUND-only issues, one-fix-per-finding suggestions) alongside the rewrite.
• New --score flag — adds a deterministic 1–10 quality score on the original prompt using a fixed rubric (5 components × {0,1,2} − issues × 0.5, clamped to 1–10, banker's rounding). Auto-promotes to --analyze --score when used alone since a bare integer hides the rubric.
• mk:prompt-enhancer frontmatter gains an argument-hint field for slash-command argument completion in supporting clients.
• Skill docs and references cleaned up — removed source-citation § symbols and academic section IDs (§5.2, §4.1, etc.) that added noise without aiding the user. Quoted source-name citations preserved.

2.8.1 (2026-05-09) — The Prompt Enhancer Release

Highlights

mk:prompt-enhancer ships as a model-agnostic prompting framework synthesized from seven prompting and context-engineering source documents. It decomposes a draft user prompt into goal / context / constraints / acceptance / output-format, detects ten weakness patterns, and emits a rewritten prompt using a universal markdown kernel that works across coding agents without model dispatch. An opt-in --deep mode surfaces codebase suggestions from allow-listed sources only.

New Skills

New Commands

• /mk:prompt-enhancer — runs the 5-step framework on a draft prompt and writes the analysis + rewrite to the active plan dir or ${CLAUDE_PLUGIN_DATA}.

Features

• Auto-suggested freedom level (LOW / MEDIUM / HIGH) based on task shape — destructive ops route to LOW, standard feature work to MEDIUM, open-ended exploration to HIGH. Surfaced in the OUTPUT FORMAT section so the user can flip without re-running.
• Auto-suggested verbosity (terse / structured / confirmation) by task type — terse for implementation, structured for review/analysis, confirmation for Q&A.
• Bounded --deep scout — ≤8 files, ≤100 lines/file, ≤30s wall clock, default-deny against .claude/memory/*, .env*, tasks/, secrets. Aborts gracefully with SCOUT_BUDGET_EXCEEDED, NO_GIT_REPO, MINIMAL_DENSITY, or NO_MATCHES. Suggestions live in [FILL-IN: <desc> (suggested: <path>)] brackets — never auto-substituted.
• 10-canary eval suite (six default-mode + four deep-mode) with hard-fail boundary canary that grep-audits saved output and transcript for forbid-list paths.

Improvements

• mk:prompt-enhancer rewrites strip model-coupled framing — XML tags (<context>, <task>), vendor tokens ("think step by step", apply_patch, "Reasoning: high"), and role-as-XML wrapping. Detection item #10 flags the input; the rewrite emits plain markdown headings.
• Added references/context-safeguards.md — six model-agnostic safeguards (right-altitude tone, identifier-based context, long-horizon defenses, tool-result clearing, bloat avoidance, eval discipline) loaded just-in-time on long-horizon signals.

2.8.0 (2026-05-09) — The Cleanup & Audit Release

Highlights

mewkit upgrade now cleans up files the new release no longer ships, so old kit files do not pile up in agent context across versions. A new mk:context-audit skill reports how much of the model context window your .claude/ setup is using and recommends the highest-value trims. Red-team and party deliberation gain cross-persona blind-spot detection — every reviewer is asked to name one dimension outside their specialty that the team may also miss. Four small behavioral rule files were merged into one agent-conduct.md, and conditional rules moved out of the always-loaded folder so the safety bundle stays lean.

New Skills

New Commands

• /mk:context-audit — runs the audit pipeline and prints a markdown report. Read-only; no files are modified.

CLI

• mewkit upgrade gains --no-cleanup (skip orphan deletion) and --yes (auto-confirm). Default behavior is on: orphans are listed and a confirmation prompt is shown before deletion. Non-TTY environments skip deletion automatically (cron-safe).

Features

• Manifest-driven orphan detection in mewkit upgrade. Compares files on disk under .claude/{rules,skills,agents,hooks}/ against release-manifest.json. Anything on disk + not in the manifest + not allowlisted is treated as an orphan.
• Conservative default allowlist exempts user customizations: custom-*, user-*, *.local.md, scope-internal README.md.
• Scope is intentionally limited to kit-owned directories. User-private state — .claude/memory/, .claude/logs/, .claude/.env, .claude/secrets/ — is never inspected.
• New CI step Validate rule availability fails the build if any rule path referenced in .md/.sh/.cjs/.js/.ts/.json/.py/.yaml/.yml/.toml files does not resolve on disk. Catches dead references before they ship.
• New skill-body precheck in mk:agent-detector HARD-GATEs Phase 0 routing on the existence of all 5 always-on safety rules (security-rules.md, injection-rules.md, gate-rules.md, core-behaviors.md, development-rules.md). Aborts with SAFETY BASELINE INCOMPLETE if any are missing.
• New ## Compaction Policy section in CLAUDE.md instructs Claude Code to preserve numbered injection-rules.md rules, the Subagent Status Protocol, and the safety-rule names verbatim across auto-compaction.
• Cross-persona blind-spot detection in mk:plan-creator adversarial review. Each persona names ONE dimension outside their specialty that the team may also miss; aggregated as a separate "Cross-Persona Blind Spots" section in the red-team report.
• mk:party Round 2 gains the same blind-spot probe — each agent calls out a dimension the room may have under-weighted.
• mk:party decision record adds a ## Next Action field — the single concrete next step coming out of the discussion.
• New conditional rule file risk-checklist.md — Phase 0 horizontal-risk flags (auth, authorization, data-model change, audit/security event, external systems, public contracts, cross-platform, existing-behavior touch) that complement the vertical domain match in mk:scale-routing.
• New helper script .claude/scripts/rules-audit.sh for contributors to check rule-file health locally.

Improvements

• smartUpdate now returns orphansDeleted and orphansSkipped arrays in UpdateStats; surfaced in the mewkit upgrade summary output.
• New find-orphans.ts module exported from core/index.ts — reusable for downstream tooling that needs orphan detection independent of the update pipeline.
• Four behavioral rule files (naming-rules.md, output-format-rules.md, context-ordering-rules.md, search-before-building-rules.md) merged into a single agent-conduct.md. Tier A preserves orchestrator-consumed schemas (Subagent Status Protocol, Project Context First, Eureka logging) verbatim; Tier B compresses model-default rules. Net: one file instead of four with no behavior change.
• Conditional rule files moved from .claude/rules/ to .claude/rules-conditional/ so the always-loaded bundle stays small. mk:agent-detector loads them on demand at Phase 0. Affected files: agent-routing.md, model-selection-rules.md, phase-contracts.md, risk-checklist.md, scale-adaptive-rules.md.

Removals

• .claude/rules/naming-rules.md, .claude/rules/output-format-rules.md, .claude/rules/context-ordering-rules.md, .claude/rules/search-before-building-rules.md — content lives in agent-conduct.md. The orphan-cleanup pass in mewkit upgrade removes the old files automatically on next upgrade.

Bug Fixes

• docs/research/claude-memory.md fixed an illustrative ln -s example that referenced a non-existent .claude/rules/security.md path. Replaced with a placeholder name (your-org-security.example.md) that does not collide with the rule-availability validator's regex.

Migration Notes

npx mewkit upgrade

The first mewkit upgrade after this release lists the four removed rule files as orphans and prompts before deleting them. Pass --yes to auto-confirm in scripted environments, --no-cleanup to keep the old files on disk.

2.7.6 (2026-05-09) — Phase 0 risk checklist

Improvements

• New rule file .claude/rules/risk-checklist.md — 9 horizontal-risk flags (auth, authorization, data-model change, audit/security event, external systems, public contracts, cross-platform, existing-behavior touch, performance-critical path) that complement vertical domain matching in mk:scale-routing. Loaded by mk:agent-detector at Phase 0. Together they raise the floor on routing accuracy without adding new tiers or new lanes.
• mk:agent-detector Step 0b now loads the risk checklist alongside phase-contracts.md and agent-routing.md.

Note: The risk-checklist file moved to .claude/rules-conditional/risk-checklist.md in v2.8.0 as part of the rules-folder split.

2.7.5 (2026-05-09) — CLAUDE.md trim + reference cleanup

Highlights

CLAUDE.md is trimmed from 209 to 112 lines to honor Anthropic's memory guidance — files over 200 lines reduce instruction adherence. Phase Composition Contracts, the Agents table, Adaptive Density, the Orchestrator Entry Point Rule, Commands-vs-Skills, advisory skill frontmatter fields, and the Skill Rule of Two now live in .claude/rules/ where they load alongside the other rule files. Cross-references in skills, docs, and scripts that pointed at the old CLAUDE.md sections were updated to the new locations. Broken *_INDEX.md pointers in shipped skill markdown were repaired so they resolve in fresh mewkit init installs.

Improvements

• CLAUDE.md trimmed from 209 → 112 lines; pointers replace inline tables. The slim file fits Anthropic's ≤200 line ceiling with a 150-line internal target.
• New rule file .claude/rules/phase-contracts.md — phase I/O contracts (what each phase expects/produces); loads unconditionally.
• New rule file .claude/rules/agent-routing.md — 17-row agent → role → phase routing table plus the MEOWKIT_PM_AUTO=off opt-out; loads unconditionally.
• .claude/rules/orchestration-rules.md gains "Orchestrator Entry Point Rule" — the mk:cook ↔ mk:workflow-orchestrator mutual-exclusion contract.
• .claude/rules/skill-authoring-rules.md gains "Advisory Frontmatter Fields" (preamble-tier, phase, trust_level, injection_risk) and "Commands vs Skills" (the three valid command patterns).
• .claude/rules/injection-rules.md Rule 11 — Skill Rule of Two as an always-loaded numbered rule.
• .claude/rules/post-phase-delegation.md Rule 4 body compressed to a pointer; eliminates duplicate restatement of the Orchestrator Entry Point Rule.
• CLAUDE.md Planning section now points to tasks/plans/YYMMDD-name/ so Claude knows where mk:plan-creator writes plans.
• validate-docs.py now emits ERROR and exits non-zero when CLAUDE.md exceeds 150 lines — prevents regression of the trim.

Bug Fixes

• mk:agent-detector references — after-detection.md cited .claude/agents/AGENTS_INDEX.md twice; that file is contributor-facing and does not ship via mewkit init. Both pointers now resolve to .claude/rules/agent-routing.md.
• mk:rubric SKILL.md references table — dropped the row pointing at RUBRICS_INDEX.md. The skill's --list flag generates the catalog dynamically, which is the actual interactive entry point.
• mk:rubric calibration checklist — removed the "Add to RUBRICS_INDEX.md" item. Custom user rubrics live in user projects where the index does not ship; --list discovers them at runtime.

2.7.4 (2026-05-02) — Browser skill consolidation

Highlights

mk:browse is retired. All browser automation now goes through mk:agent-browser, eliminating the routing decision between two overlapping skills. A new migration reference ships inside mk:agent-browser with verb-to-verb mapping, copy-paste recipes for capabilities agent-browser doesn't expose natively (responsive screenshots, link/form enumeration, performance timing, eval-based state checks), and a developer-machine handoff runbook for CAPTCHA/MFA flows.

Improvements

• mk:agent-browser SKILL.md gains a data-boundary callout — page content, snapshot text, and eval return values are DATA per injection-rules.md. Recommends AGENT_BROWSER_CONTENT_BOUNDARIES=1 so page-derived strings arrive wrapped in nonce markers.
• Sessions and credentials note added — any caller using --session-name should set AGENT_BROWSER_ENCRYPTION_KEY first, otherwise session state under ~/.agent-browser/sessions/ is plaintext.
• New reference agent-browser/references/migrating-from-browse.md covers verb mapping, recipes (responsive, links, forms, perf, is-state fallback), and the --headed + wait --url runbook for interactive auth.
• mk:qa setup and mk:office-hours preamble now use a command -v agent-browser readiness check instead of looking for the gstack binary, with an actionable NEEDS_SETUP message pointing at npm i -g agent-browser.
• Cross-skill routing tables in mk:agent-detector, mk:evaluate, mk:chom, mk:fix, and mk:qa-manual updated to name mk:agent-browser only — the old mk:browse-or-mk:agent-browser pick-by-complexity guidance is gone.

Removals

• mk:browse skill — directory deleted, manifest entries removed, catalog row dropped from SKILLS_INDEX.md. Historical attribution preserved in SKILLS_ATTRIBUTION.md under ## Removed. Existing bookmarks to /reference/skills/browse redirect to a deprecation stub linking to mk:agent-browser.
• $B shell-alias workflow patterns from mk:qa references and mk:office-hours/phase4-alternatives-and-sketch.md — replaced with agent-browser invocations or with the eval recipes from the migration guide.

Migration Notes

• Install the CLI: npm i -g agent-browser (or brew install agent-browser, or cargo install agent-browser), then run agent-browser install once to download Chrome.
• Set AGENT_BROWSER_ENCRYPTION_KEY in your shell and CI secret store before invoking any session-using recipe — required for cookies/localStorage to persist encrypted at rest.
• Existing gstack auth-state.json files (cookie-import format) are not portable. Re-run agent-browser auth save or capture fresh state with agent-browser state save.
• Personal scripts using $B <verb> should consult agent-browser/references/migrating-from-browse.md for the verb mapping.

2.7.3 (2026-05-01) — npx mewkit resolution fix

Bug Fixes

• npx mewkit CMD failed on some machines with sh: meowkit: command not found (macOS) or 'meowkit' is not recognized as an internal or external command (Windows) — the npm package was named mewkit but exposed a bin called meowkit, so resolvers (older npm, pnpm dlx, yarn dlx, certain Windows shells) spawned the bin by name through PATH instead of the symlinked .bin/ path. Bin renamed to mewkit so package name and bin agree across all package managers.
• pre-completion-check.sh counter never reaching its 3-attempt soft-nudge cap — Claude Code assigns a new session ID for every blocked-Stop cycle, and project-context-loader.sh was resetting session-state/precompletion-attempts.json on every SessionStart, wiping the counter mid-loop. The pre-completion check now owns its own counter lifecycle (clears on verification success or cap-hit); the SessionStart reset is removed.

CLI

• Help banner and mewkit status now print mewkit instead of meowkit to match the published bin name.
• mewkit doctor venv-missing hint updated to suggest mewkit setup --only=venv.

2.7.2 (2026-05-01) — Checkpoint subsystem cleanup

Highlights

The session-state checkpoint subsystem stops accumulating files unboundedly. session-state/checkpoints/ is now a single overwriting checkpoint-latest.json written atomically via POSIX .tmp + rename — replacing the prior pointer-and-numbered-file scheme that had no rotation logic. Resume context, git-drift detection, and budget tracking work the same; the file count just stays at 1.

Improvements

• Checkpoint dir collapsed to a single file — checkpoint-utils.cjs rewritten without nextSequence(), the O_EXCL lock, the .next-seq counter, or the filename validator. Two writers (Stop hook + phase-transition trigger) race to last-writer-wins; sequence is display-only.
• auto-checkpoint.cjs fires on phase-transition writes only — the modulo-20 mid-session trigger and auto-checkpoint-counter.json are removed. The Stop hook still captures end-of-turn state.
• orientation-ritual.cjs reads the checkpoint file directly with one readFileSync and a try/catch — no more pointer indirection, no more two-hop read.
• New safePlanPath guard in orientation-ritual.cjs rejects .. traversal and control-character injection (e.g. embedded \n "ignore previous instructions" payloads) in the plan_path field before stdout emission. Replaces the deleted isValidCheckpointFile filename guard with content-level validation.
• Two superseded shell hook writers removed — .claude/hooks/post-write-build-verify.sh and post-write-loop-detection.sh were marked "Not registered. Do not re-register." Their .cjs ports (build-verify.cjs, loop-detection.cjs) keep working unchanged; env-var bypasses (MEOWKIT_BUILD_VERIFY=off, MEOWKIT_LOOP_DETECT=off) still honored by the active handlers.
• HOOKS_INDEX.md reconciled with disk — stale rows for the v2.4.0-deleted memory-loader.cjs, memory-parser.cjs, memory-injector.cjs, memory-filter.cjs removed, replaced with a single tombstone callout. State-file table updated to reference checkpoint-latest.json instead of auto-checkpoint-counter.json.
• AGENTS_INDEX.md, build-verify-commands.md, and trace-schema.md reconciled — stale "13 agents" phrasing corrected to "17 agents", .sh references updated to .cjs, and the build_verify_result / loop_warning trace events deprecated (their emitters were removed in v2.4.0; the events have not fired since).

Bug Fixes

• Pre-existing silent-failure path closed — under nextSequence() lock contention the function returned null, both writers passed null to writeCheckpoint, producing checkpoint-null.json which isValidCheckpointFile rejected; orientation-ritual then returned empty silently and the user got a cold-start resume with no error visible. The new design has no lock and no null-seq path; resume cannot fail this way.
• Resume injection no longer susceptible to prompt-injection via the plan_path field — safePlanPath filters .. traversal and control-character payloads (newline, escape sequences) before they reach Claude's system context.

Removals

• meowkit/.claude/hooks/handlers/checkpoint-utils.cjs exports nextSequence and isValidCheckpointFile are gone — they were only used inside the checkpoint subsystem and have been replaced by the single-file overwrite pattern.
• meowkit/session-state/checkpoints/.next-seq, .seq.lock, and auto-checkpoint-counter.json are no longer produced. Existing files are stale and safe to delete; orientation-ritual.cjs ignores them.
• Two shell writers removed (see Improvements). Their functionality was already migrated to .cjs handlers in v2.4.0; the .sh files were dead code.

2.7.1 (2026-04-30) — Phase Frontmatter Contract

Highlights

mk:plan-creator now ships a strict YAML frontmatter contract for phase files — frontmatter is the machine-readable source of truth, the Overview block is a human-readable mirror that sync-back regenerates from it. Aligns plan-creator output with the orchviz parser cascade, hardens the validator against drift, and formalizes the cook finalize-step sync-back algorithm.

Improvements

• Phase files (phase-XX-*.md) MUST begin with a YAML frontmatter block — schema: phase, title, status, priority, effort, dependencies. status: pending is the only legal value at creation; failed and abandoned are terminal. Documented in references/phase-template.md and references/output-standards.md.
• Cook's finalize-step sync-back is now a formal algorithm — counts [x] checkboxes in ## Todo List to derive pending / in_progress / completed, regenerates the Overview mirror from frontmatter, never overwrites terminal states, idempotent on re-run. Documented in references/task-management.md.
• Step-08 hydration adds a "Status Read Order" block that mirrors packages/mewkit/src/orchviz/plan/parse-phase-file.ts so plan state stays consistent across the planner, the validator, the harness, and orchviz.
• --two mode approach files (plan-approach-a.md, plan-approach-b.md) get frontmatter at creation with status: pending — draft is no longer used because it is not in the parser union.
• --parallel mode merges ownership and parallel_group into the same frontmatter block instead of a separate one.
• mk:review step-03b whole-plan-sweep gains a drift surface — flags stale dependencies: [N] references when phases are renumbered or removed (informational, non-blocking).

Bug Fixes

• scripts/validate-plan.py now hard-fails on status: completed written at creation time and on unknown / draft / done written from frontmatter — the validator catches the "stamp at creation" anti-pattern before Gate 1.
• scripts/validate-plan.py documents its PyYAML dependency — install via npx mewkit setup (creates .claude/skills/.venv/); run via .claude/skills/.venv/bin/python3 scripts/validate-plan.py <plan.md>.

2.7.0 (2026-04-30) — The Namespace Rename Release

Highlights

Skill folders renamed from .claude/skills/meow:X/ to .claude/skills/X/ and the skill identity prefix moved from meow: to mk:. Cross-platform safe — eliminates the colon that broke Windows NTFS checkouts. Slash commands move from /meow:X to /mk:X. Backward-compat alias keeps in-prompt meow: text working with a stderr warning during the v2.7.x window; hard-cut in v2.8.0.

Features

• New helper packages/mewkit/src/migrate/discovery/skill-id-utils.ts — single source for skill-id parsing, validation, and legacy-prefix resolution. Exports parseSkillId, resolveLegacy, _resetWarnState. Path-traversal guard rejects basenames not matching ^[a-z][a-z0-9-]{0,62}$.
• New CI step Reject residual meow: namespace in .github/workflows/ci.yml — fails the build if any ^name: meow: SKILL.md frontmatter or /meow:[a-z] slash reference exists in .claude/commands/, .claude/rules/, or CLAUDE.md.
• SkillInfo.id is now a required field on the discovery type, derived from frontmatter name: and normalized through resolveLegacy. Forces compile-time handling at every consumer.

Improvements

• discoverSkills() is fully content-based — any directory under .claude/skills/ containing a SKILL.md is a skill, regardless of folder prefix. setup.ts:54 collectSkillDeclaredDeps filter changed from name-based (startsWith("meow:")) to content-based (existsSync(SKILL.md)).
• Schema regex at .claude/schemas/skill-schema.json widened from ^meow:[a-z][a-z0-9-]*$ to ^(mk|meow):[a-z][a-z0-9-]*$ for the deprecation window. Narrows to ^mk: in v2.8.0.
• scripts/validate-skill-frontmatter.py glob fixed from meow:*/SKILL.md to */SKILL.md plus a zero-match guard — empty scans now exit with code 2 instead of phantom-passing.
• scripts/check-skill-cross-refs.sh rewritten — inventory built from bare-name folders, scans for mk:[a-z][a-z0-9-]* references, fails on empty inventory.
• All hook scripts refreshed — gate-enforcement.sh, privacy-block.sh, post-session.sh, project-context-loader.sh, pre-completion-check.sh, memory-topic-file-migrator.cjs. Privacy-block SSRF glob updated to match *web-to-markdown* and *mk:web-to-markdown* so the guard survives the folder rename.
• skill-creator is now self-consistent — init-skill.py and validate-skill.py enforce the mk: prefix; SKILL.md template emits /mk:{slug} invocation syntax and mk:{name} frontmatter.

Removals

• .claude/skills/meow:*/ folder layout removed. Skills now live at .claude/skills/X/. Migration: npx mewkit upgrade runs the legacy-namespace migrator with confirmation prompt.
• .claude/commands/meow/ folder removed. Slash commands now live at .claude/commands/mk/. The 21 existing commands moved as-is; the 56 phantom skill-only aliases were intentionally not generated.

Migration Notes

npx mewkit upgrade

The upgrade pipeline detects existing meow:* folders, prompts the user, and then:

• Renames folders via git mv (preserves history); falls back to fs.rename on non-git repos.
• Rewrites SKILL.md frontmatter name: from meow:X to mk:X.
• Reconciles ~/.mewkit/portable-registry.json — strips meow- prefix from sanitized item keys.
• Aborts if the git tree is dirty, or if any folder basename fails the path-safety regex.

If you have custom scripts or aliases that invoke /meow:cook, retype them as /mk:cook. In-prompt text references like meow:scout continue to work via the resolver during v2.7.x — they emit a one-time stderr deprecation warning per id, and they hard-fail in v2.8.0.

Breaking Changes

• Skill folder paths: .claude/skills/meow:X/ → .claude/skills/X/.
• Skill identity prefix: meow:X → mk:X in SKILL.md frontmatter and all cross-references.
• Slash command paths: .claude/commands/meow/X.md → .claude/commands/mk/X.md. Type /mk:cook instead of /meow:cook.
• SkillInfo.id field is now required (was previously absent). Downstream consumers reading the discovery output now see a guaranteed canonical id.

2.6.2 (2026-04-29) — The Telemetry & Validator Release

Highlights

Hook telemetry, schema-validated skill frontmatter, cross-reference CI, and a re-baselined critical-findings audit. Every probe is data-driven — Phase 3 advanced enhancements stay deferred until the new hook-log.jsonl shows real demand. CI now blocks merges that introduce phantom mk:* references or unregistered hooks.

CLI

• mewkit setup gains a project-context step that warns when docs/project-context.md is absent and tells the user which /mk:project-context action to run.

Features

• New scripts under scripts/ — check-skill-cross-refs.sh, check-hook-registration.sh, validate-skill-frontmatter.py, and telemetry-decisions.py. All four wire into CI; the telemetry tool runs on demand and emits per-item ship/reject verdicts from log data.
• New JSON Schema at .claude/schemas/skill-schema.json — permissive baseline calibrated against all 77 current skills. Validates frontmatter on every PR.
• New hook helper lib/hook-logger.sh — append-only JSONL telemetry at .claude/hooks/.logs/hook-log.jsonl with 50MB rotation (rotated logs stay plain text, no gzip, so telemetry analysis stays cheap).
• Probe hooks for PreCompact and PostToolUseFailure, plus a Stop-event control probe that disambiguates "event unsupported" from "logger broken." Activated immediately; data accumulates in the background.
• New canonical hook regression test at .claude/hooks/__tests__/advisory-boundary.test.cjs runs via node --test and locks @@GATE_BLOCK@@ and @@PRIVACY_BLOCK@@ sentinel output.
• mk:review workflow gains step step-03b-whole-plan-sweep.md between triage and verdict — re-reads plan.md plus phase files, surfaces cross-file drift, never auto-FAILs.
• mk:validate-plan gains a Whole-Plan Consistency Sweep section that emits a sweep_failures block alongside the 8-dimension verdict.

Improvements

• .github/workflows/ci.yml now runs all four validators and the hook regression test on every PR, in the existing validate job. No new workflow file.
• meowkit-architecture.md 10 critical-findings table re-baselined — four of six findings the prior audit listed as OPEN were already CLOSED in the current tree. Component inventory updated to reflect the 8 actual handlers (prior table claimed 12).
• Probe hooks ship behind the existing telemetry path so future advanced-feature decisions consult hook-log.jsonl evidence instead of speculation.

Bug Fixes

• Audit doc claimed mk:cook/SKILL.md:159 had a bare memory/ path — verified-incorrect. Line 170 already used the correct .claude/memory/ prefix. Stale finding closed.
• Audit doc claimed mk:lazy-agent-loader/SKILL.md:13 hardcoded an agent count of 15 — no hardcoded count exists in the current tree. Stale finding closed.
• Audit doc claimed seven mk:* command refs in commands/meow/ were phantom — all seven resolve to a real command file under the post-audit "Commands vs Skills" rule. Stale finding closed.

Migration Notes

• After npx mewkit upgrade, run the new validators locally to catch any drift introduced before the upgrade — bash scripts/check-skill-cross-refs.sh && .claude/skills/.venv/bin/python3 scripts/validate-skill-frontmatter.py.

2.6.1 (2026-04-22) — The project-manager Release

Highlights

New 17th core agent project-manager — a cross-workflow delivery tracker that aggregates plan, test, review, contract, and cost state into an evidence-based status report classified as done / in progress / blocked / not started. Ships the new /mk:status slash command, a new post-phase-delegation.md rule, and a new pm-status-template.md report schema. Five orchestration skills (mk:cook, mk:autobuild, mk:workflow-orchestrator, mk:fix complex path, mk:worktree) now cite the rule to delegate post-phase. Opt-out via MEOWKIT_PM_AUTO=off.

New Agents

• project-manager — cross-workflow delivery tracker (haiku tier). Reads plan + verdicts + contracts + cost-log + git log; writes status reports co-located inside each plan dir at {plan-dir}/status-reports/{YYMMDD}-status.md. Backward-looking ("what's done, what's blocked") — complement, not replacement, for mk:help which remains forward-looking ("what's next").

New Commands

• /mk:status — foreground entry point for project-manager. Resolves the active plan (prompts if multiple), delegates to the agent, prints the report path.

Features

• New rule post-phase-delegation.md — 7-rule charter defining PM fire points, invocation form, skip conditions, orchestrator disambiguation, and the no-hook-dispatch safety property.
• New report template tasks/templates/pm-status-template.md — follows planner's template convention; schema edits happen in one file, not inside the agent body.
• New env var MEOWKIT_PM_AUTO=off — disables all silent PM fires from orchestration skills. /mk:status still works regardless.
• mk:agent-detector gains a Delivery Status intent (keywords: status / progress / what's done / what's blocked / delivery tracking) routing to project-manager, explicitly distinct from mk:help's forward-looking navigation intent.

Improvements

• Status reports are co-located with the plan they describe. Archive moves them together. No central tasks/status-reports/ dir; each plan lazily creates its own subdir on first write.
• Harness iteration cap escalation (step-05) now surfaces current delivery state BEFORE the AskUserQuestion escalation — user sees what's done and what's blocked before deciding ship / abort.
• Parallel-execution Rule 5 appends: after the integration test passes, PM emits a merge report summarizing what each parallel branch contributed.
• Agent count: 16 → 17. Architecture 2 counts, 6 roster, AGENTS_INDEX, CLAUDE.md, and agent-detector routing data all synchronized.

Bug Fixes

• docs/meowkit-architecture.md mermaid subgraph caption "Key Agents (16)" updated to (17); slash-commands count bumped 20 → 21.

2.6.0 (2026-04-22) — The Skills Compliance Release

Highlights

A 7-agent audit of all 77 mk:* skills against Anthropic's Skill-authoring best practices, Lessons-from-building-Claude-Code, and MeowKit's own skill-authoring-rules.md. Ships ~220 edits across description fields, frontmatter, reference integrity, scripts, and grounding. Zero new skills, zero breaking changes, measurably cleaner routing.

Improvements

• All skill descriptions normalized to Anthropic third-person format — mk:cook, mk:fix, mk:agent-detector, mk:session-continuation rewritten out of imperative/greedy mood that cannibalized sibling routing.
• Five overlap clusters disambiguated with explicit NOT for X (see mk:Y) clauses — bug-fix (fix / investigate / build-fix), pipeline (cook / workflow-orchestrator / harness), browser (browse / playwright-cli / agent-browser / qa-manual), code-quality (review / clean-code / simplify / evaluate), planning (plan-creator / planning-engine). Twenty additional skills got focused exclusion clauses where latent overlap existed.
• mk:agent-browser 8 inter-ref cross-links flattened and authentication.md ↔ session-management.md circular reference broken — all 9 reference files remain directly linked one level deep from SKILL.md.
• New step-file-rules.md Rule 6 — step-file skills (mk:review, mk:trace-analyze, mk:plan-creator, mk:autobuild, mk:evaluate) formally allowed to chain SKILL.md → step-NN.md → references/X.md provided each step file opens with a Contents TOC.
• Seven skills that reference .claude/skills/... paths directly gained a > Path convention: note declaring $CLAUDE_PROJECT_DIR as the assumed cwd — mk:rubric, mk:multimodal, mk:skill-creator, mk:intake, mk:llms, mk:investigate, mk:jira.
• MCP prerequisite hardening for mk:jira, mk:confluence, mk:planning-engine — server-key assumption (.mcp.json key atlassian) documented via Gotchas without hardcoding prefixes, preserving install portability.
• Twenty-two skills renamed their trigger section to canonical ## When to Use — previously fragmented across "When to Invoke", "When to Activate", "Trigger Conditions".
• 115 reference files over 100 lines auto-gained a ## Contents Table of Contents via idempotent generator, safe to re-run on future drift. Original audit counted 25; actual was 116 (4.6× under-count, largely from mk:angular).
• Frontmatter normalization — preamble-tier: 4 → 3 on mk:qa / mk:review / mk:ship (valid range is 1-3). Dead autoInvoke and priority fields deleted from four skills after grep-confirming zero readers in hooks, scripts, or CLI. sources → source singular unified across seven skills. mk:chom injection_risk bumped low → medium since it replicates external code into the project.
• mk:skill-creator compliance — now has its own ## Gotchas section, 500-line cap matches authoritative Rule 3 (was 150), emitted template includes mandatory ## Gotchas header so every future scaffold inherits Rule 1 compliance. scripts/validate-skill.py gained check 8/8 for Gotchas header presence and auto-passes step-file skills.

Bug Fixes

• mk:lint-and-validate was undiscoverable — malformed YAML frontmatter had an unquoted colon in Triggers onKeywords: which corrupted parsing. Description now wraps cleanly in double quotes.
• Empty Python venv — .claude/skills/.venv/ existed but pip was half-installed with no RECORD file and no bin/pip symlink, so all seven Python-backed skills crashed on first import. Bootstrapped via get-pip.py --ignore-installed --no-deps, then npx mewkit setup --only=deps installed the nine required packages. npx mewkit doctor now reports 13 PASS / 1 WARN (optional Playwright only).
• mk:rubric/scripts/validate-rubric.sh --help no longer crashes on macOS — added -h/--help case before arg dispatcher, which previously fell through to basename "--help" triggering a BSD illegal option error.
• Replaced non-existent debugger agent reference in mk:cook's Phase-3 dispatch table with developer via mk:investigate — .claude/agents/ never had a debugger.md.
• HARD-GATE decorative tags in mk:cook and mk:fix bodies replaced with **HARD GATE** bold markdown — grep against hooks/, scripts/, and the CLI confirmed zero consumers of the tag format.
• Phantom research-01 citations removed from mk:evaluate (4 locations) — no such research file exists; constraints now cite Anthropic harness research honestly or declare themselves as heuristics.
• Phantom research-02 citations removed from mk:rubric/references/calibration-guide.md (2), mk:sprint-contract/SKILL.md, mk:sprint-contract/references/bdd-to-ac-mapping.md (2), and .claude/rubrics/schema.md — same non-existence.
• mk:multimodal "332 system voices" unsourced count corrected to "300+ (see provider catalog)" with the MiniMax URL attached.
• mk:react-patterns unverifiable "45+ rules across 8 priority categories from Vercel Engineering" attribution rephrased to "curated rules drawn from framework docs and production practice".
• mk:ui-design-system CSV counts corrected — colors.csv 161 → 160, ux-guidelines.csv 99 → 98 (matches actual row counts).
• mk:vulnerability-scanner dropped time-anchored "2025 threat landscape" tagline.
• mk:multimodal/scripts/minimax_api_client.py magic numbers poll_interval=10 and max_wait=600 now carry derivation comments. media_optimizer.py tokens_per_sec = 100 / 263 constants sourced to ai.google.dev/pricing.
• mk:multimodal/SKILL.md — check_setup.py gained the bash invocation block every other script in the file already had.
• mk:scout 6-agent cap now documented as an intentional exception to parallel-execution-rules.md Rule 2 (read-only Explore agents produce no merge-conflict risk).
• mk:browse gained a Gotcha for the session-scoped $B alias silent-fail mode.
• mk:problem-solving unclosed backtick at line 148 fixed.

Migration Notes

• npx mewkit upgrade picks up all description changes. No user action required.
• If any custom hook was reading autoInvoke / priority frontmatter fields: they were grep-confirmed unread before deletion. If a hook outside the repo tree was silently depending on them, restore them explicitly in the affected skills.
• preamble-tier: 4 → 3 on mk:qa / mk:review / mk:ship may cause tiny context-ordering shifts. Functionally identical outputs.

2.5.1 (2026-04-20) — mk:henshin

Highlights

New cross-cutting skill mk:henshin — planning front door for transforming existing code into agent-consumable surfaces (CLI + MCP server + companion skill). Adapted from claudekit-engineer/agentize (tier 2). Produces a Transformation Spec and hands off to /mk:plan-creator → /mk:cook for the build. Not a builder on its own.

New Skills

• mk:henshin — 6-phase planning workflow: Discover → Inventory → Capability Map → HARD GATE → Spec Write → Handoff. Non-bypassable human approval for package name, license, and ownership (business decisions); --auto and --lean never skip these. Writes an architectural decision record to .claude/memory/architecture-decisions.md with ##decision: prefix. Ships 5 progressive-disclosure references under 200 lines each (agent-centric-design, auth-resolution-chain, mcp-transports, monorepo-layout, challenge-framework).
• Boundary vs mk:chom — henshin is outbound (local code → agent surfaces); chom is inbound (external repo → local project). No semantic overlap. Paste a GitHub URL into henshin and the Error Recovery section redirects to chom.
• Triggers on agentize, henshin, expose as MCP, wrap as CLI, publish to npm, make LLM-accessible, turn into agent tool.

Bug Fixes

• mk:henshin/references/mcp-transports.md — SSE code snippet rewrote to use Express + a per-session transport map. Previous snippet mixed Hono-style c.res with a Node.js ServerResponse constructor, producing a copy-paste type error in Hono environments, plus a scope leak where a single t served multiple clients.
• mk:henshin/references/monorepo-layout.md — commander dependency bumped from ^12.0.0 to ^14.0.0 (current major).

Documentation

• New per-skill reference page: /reference/skills/henshin.
• /guide/agent-skill-architecture — henshin added to the Cross-Cutting Skills and Quick-Start tables.
• /mk:henshin — new skill for agentizing code into CLI, MCP, and companion skill surfaces. v2.5.1 entry with workflow summary and boundary rationale.

2.5.0 (2026-04-19) — The Native Fit Release

Highlights

Skills stop branding themselves and start reading like the project's own workflow. Roughly 70 user-facing phrases were rewritten across 50+ files so installed kits refer to "the project" / "this workflow" / "the kit" instead of "MeowKit's X", while real infrastructure names (binaries, env vars, CLI commands, frontmatter source: fields) stay intact. Ships the full 64-skill audit: mechanical cleanup, reference fixes, boundary disambiguation for seven collision clusters, and architecture wiring for memory persistence and dual-orchestrator arbitration.

Features

• mk:evaluate, mk:benchmark, and mk:party now persist their outputs across sessions — verdicts to .claude/memory/review-patterns.md, baselines to cost-log.json, architectural decisions to decisions.md. Each writer runs mkdir -p .claude/memory first so the append never fails silently.
• CLAUDE.md gains an Orchestrator Entry Point Rule — explicit /mk:cook wins; mk:workflow-orchestrator defers for the rest of the session. Ends duplicate Gate 1 enforcement.
• CLAUDE.md gains a Skill Frontmatter Schema section defining preamble-tier (1 | 2 | 3), user-invocable (true | false), phase (0-6 | on-demand), trust_level (kit-authored | third-party), and injection_risk (low | medium | high). Twenty-one skills were already using preamble-tier: 3 without a documented schema.
• mk:careful emits an audit-trail log entry to .claude/memory/security-log.md on every warn/override event (timestamp, pattern, severity, command). Log file is auto-initialized with a markdown-table header.

Improvements

• Installed skills read as the project's own workflow — MeowKit follows the Boil the Lake principle → This workflow follows the Boil the Lake principle; Help MeowKit get better! → Help improve this workflow!; table column MeowKit → AI-assisted; dozens of similar edits.
• Seven collision clusters disambiguated — each pair now declares use when + use OTHER instead when at the top of both skills; no more buried routing rules.
• NOT this skill if: differentiators added to the reasoning cluster (mk:elicit, mk:brainstorming, mk:problem-solving).
• mk:workflow-orchestrator bare "implement" trigger replaced with "implement feature" — no longer fires autoInvoke:true on trivial requests.
• mk:ship/references/preamble.md declares a memory-read of .claude/memory/architecture-decisions.md at task start. Previously the memory reference existed in frontmatter without a body instruction.
• mk:review and mk:ship look up plans from .claude/plans/ and tasks/plans/ first; ~/.claude/plans/ is retained only as a legacy fallback with an explanatory comment.
• mk:cook intent-detection table carries a green-field escalation callout pointing at mk:autobuild.
• mk:help documents the /mk:plan alias — confirmed via the slash-command router at .claude/commands/meow/plan.md.
• mk:docs-finder documents Node.js 18+ as a prerequisite for its .js scripts and flags unbounded .claude/memory/docs-cache/ growth in Gotchas.
• mk:investigate Process step 4 clarifies the freeze hook is a no-op without an explicit mk:freeze TARGET-DIR invocation.
• mk:benchmark/SKILL.md accurately describes run-canary.sh as a step-1-of-2 manifest emitter (not a suite runner); compare-runs.sh distinguishes null scores (PENDING, excluded from averages) from valid 0 scores.
• mk:jira description narrowed from CRUD + evaluation to CRUD only — ticket analysis now routes to mk:intake.
• mk:scale-routing marked user-invocable: false — it's a Phase 0 sub-skill, not a user-facing entry point.
• mk:party removes retro from its triggers with a pointer to mk:retro for dedicated retrospectives.
• Six skills gained phase: on-demand frontmatter (chom, clean-code, docs-finder, docs-init, lint-and-validate, multimodal); mk:jira corrected from phase: 3 → phase: on-demand.
• Antigravity-kit-sourced skills (clean-code, lint-and-validate) gained trust_level: third-party + injection_risk: low provenance markers.
• 34 hardcoded /Users/sangnguyen/Desktop/compare-kit/... lines stripped from 10 mk:angular/references/*.md files — source-import metadata with zero runtime value.

Removals

• Dead Codex Review | /codex review table row removed from 6 shared-protocol tables (mk:browse, mk:document-release, mk:investigate, mk:office-hours, mk:plan-ceo-review, mk:retro) — /codex review referenced the OpenAI Codex CLI, not a meowkit skill.
• claudekit-engineer provenance prose genericized in body HTML comments, script file headers, and reference attributions. Frontmatter source: claudekit-engineer fields are policy-preserved.

Bug Fixes

• mk:careful FLUSHDB severity aligned — destructive-patterns.md said HIGH, check-careful.sh enforced CRITICAL. Both now say CRITICAL.
• mk:ui-design-system — 57 font pairings count fixed to 73 (CSV has 73 data rows; documented count had drifted 28%).
• mk:brainstorming — invented BMAD's ~10% pivot frequency statistic removed (no source document existed).
• mk:vulnerability-scanner — fabricated OWASP Top 10:2025 version label relabeled to OWASP Top 10 (2021 ranking, 2025 threat context) with a disclaimer. No official 2025 release exists as of April 2026.
• mk:multimodal — gemini-3.1-* and veo-3.1-* preview model IDs now carry a verify-before-deploy warning pointing at the Gemini model docs.
• mk:scout and mk:docs-finder — P1-P8 / P1-P14 citations removed from Anthropic context-engineering-research attributions (neither research doc uses those labels).

Migration Notes

• mk:workflow-orchestrator no longer fires on bare "implement" — use compound triggers ("implement feature", "build feature", "create feature", "complex task") or invoke /mk:cook explicitly.
• mk:party memory write is now mandatory, not optional. Decision records land in .claude/memory/decisions.md. If a shell blocks writes to .claude/memory/, whitelist the directory before invoking party.
• Upgrade: npx mewkit upgrade. Fresh install: npx mewkit init. Existing .claude/memory/ topic files are preserved; new writers mkdir -p safely if a parent directory is missing.

2.4.6 (2026-04-19) — mk:ship Cleanup + Design Review Checklist

Highlights

mk:ship drops its unused Codex (OpenAI CLI) integration, fixes broken bash in the preamble, and removes phantom slash-command references. mk:review gains a lite design-review checklist — source-level pattern detection for frontend diffs, adapted from gstack with additions from claudekit-engineer (Strategic Omissions, Fix Priority Order) and everything-claude-code (anti-template patterns).

Improvements

• mk:review/design-checklist.md — new lite design-review reference with six categories (AI Slop Detection, Typography, Spacing & Layout, Interaction States, DESIGN.md Violations, Strategic Omissions) and a three-tier confidence system ([HIGH] / [MEDIUM] / [LOW]) for grep-actionable pattern detection.
• mk:ship large-diff adversarial review is now Claude-only — two passes (Claude structured + Claude adversarial subagent) replace the prior four-pass cross-model scheme that depended on an uninstalled external CLI.
• mk:ship and mk:review Gotchas sections document the scope-gated frontend trigger on the design check and the source-pattern-only limitation (no visual rendering).

Removals

• Codex (OpenAI CLI) integration removed from mk:ship — codex exec, codex review, and which codex guards dropped across five reference files. Claude adversarial subagent was already the runtime fallback.
• Phantom slash-command references removed from mk:ship — /qa-only, /plan-design-review, and /design-review (none exist as meowkit skills). The plan-verification step now emits a post-ship reminder to run /mk:qa against the deployed build instead of attempting to invoke a missing skill.
• design-review-lite as a fake skill-name in the review-log schema — renamed to "source":"ship-design-check" so the readiness dashboard no longer pretends a non-existent skill produced the entry.

Bug Fixes

• mk:ship/references/preamble.md — two bash syntax errors fixed (unterminated $(# comment) upgrade-check and orphan for...done with a comment-eaten done). bash -n now passes on every code block; prior runs silently failed before any shipping work started.
• mk:ship/references/plan-completion-audit.md — the dead /qa-only invocation that silently failed every run is replaced with a one-line reminder to run /mk:qa against the deployed build.

2.4.5 (2026-04-19) — The Thinking Skills Release

Highlights

A new strategic-unsticking skill and three diagnostic frameworks for evidence-based debugging. mk:problem-solving ships seven non-default techniques for approach-level stuck-moments — distinct from mk:sequential-thinking, which gains five-whys-plus, scientific-method, and kepner-tregoe references for methodology-aware root-cause work.

New Skills

Improvements

• mk:sequential-thinking gains three diagnostic-framework references — five-whys-plus.md (bias guards + stopping criteria), scientific-method.md (falsifiable prediction discipline), kepner-tregoe.md (IS/IS-NOT matrix for multi-system bugs). Core workflow unchanged; load when a specific methodology fits.
• SKILLS_INDEX.md and website skill indexes register the new skill and clarify cross-cutting placement for both thinking skills.

2.4.4 (2026-04-19) — Deprecated Skill Cleanup + Brand Refresh

Highlights

Three deprecated skills removed permanently. Brand assets (animated logo, SVG favicon, OG image) wired into VitePress. Vercel routing fix for direct URL access.

Removals

• mk:debug — deprecated since v2.0.0, superseded by mk:investigate. Directory removed from .claude/skills/.
• mk:documentation — deprecated since v2.0.0, superseded by mk:document-release. Directory removed.
• mk:shipping — deprecated since v2.0.0, superseded by mk:ship. Directory removed.
• Dead references in SKILLS_INDEX.md, website/reference/skills/index.md, website/reference/skills-index.md, website/reference/skills/docs-init.md, docs/project-context.md, and docs/meowkit-architecture.md cleaned up.

Improvements

• Brand assets wired into VitePress — animated SVG logo (meow-logo-animated.svg) replaces raster logo.png/logo.webp in navbar and homepage hero.
• Favicon refreshed — SVG primary (favicon.svg) + regenerated multi-resolution .ico + 16/32 PNGs + apple-touch-icon from the new brand source via ImageMagick.
• OG image swapped to the /meow| brand social card (1200×630).
• Site title hidden in navbar (siteTitle: false) — the logo carries the brand; no duplicate text.
• Color token system synced with brand source-of-truth (assets/css/colors_and_type.css): added --mk-neutral-200/800/950, --mk-accent-soft, full radius scale (xs/xl), full shadow set (card/lift/glow-strong/inset-glow), and motion easings (--mk-ease-out / --mk-ease-in-out + --mk-dur-fast/normal/slow).
• Changelog reorganized to a strict section schema (Highlights / New Skills / Improvements / Removals / Bug Fixes / CLI / etc.) — removed internal IDs, test counts, audit metadata. Added npx mewkit upgrade preamble.
• What's-new index compressed to 1-line headlines + ≤3 bullets per release.
• RELEASING.md step 2 rewritten with the new changelog schema, style rules (DRY), and a patch-vs-minor-vs-major required-section matrix.

Bug Fixes

• Direct URL access on Vercel (/changelog, /guide/whats-new) returned 404 — added cleanUrls: true and trailingSlash: false to vercel.json. SPA-navigation from / worked; direct requests did not because Vercel served only /changelog.html without the rewrite.
• Homepage hero image was hardcoded to /logo.png in index.md frontmatter, separate from themeConfig.logo — the navbar swap didn't propagate. Now /logo.svg.
• README hero image pointed at assets/branding/meowkit-logo-black.png which never existed — swapped to assets/branding/meow-logo-dark.svg.
• Stale favicon issue — browsers auto-request /favicon.ico from root regardless of <link> order, picking the old raster. Regenerated all raster variants from the new SVG.

2.4.3 (2026-04-18) — Brainstorming v2: Discovery, Scope, Anti-Bias Pivot

Highlights

mk:brainstorming rewritten with patterns extracted from BMAD-METHOD, ClaudeKit, and the everything-claude-code structural template. Adds discovery protocol, scope assessment, single mid-session anti-bias pivot, and 3 new techniques.

Features

• Discovery protocol — AskUserQuestion capped at 3 questions per batch, targeting binding constraint, success criteria, and ruled-out options.
• Scope assessment — 3+ independently-shippable concerns (heuristic: would each be its own GitHub issue?) → user decomposes before brainstorming proceeds.
• Anti-bias pivot — one mandatory orthogonal-category pivot at idea #4 (midpoint).
• Idea Format Template — every idea carries a mandatory Novelty line; idea is dropped as a duplicate if you can't write one.
• Technique selection tiebreaker — explicit order when multiple techniques match.
• Output templates upgraded with audit-trail fields (Discovery Trace, Scope Decision, Technique Selection rationale, pivot record, Category Distribution, scoring risk callouts).

New References

• references/techniques/analogical-thinking.md — cross-domain transfer (forces non-software analogues).
• references/techniques/scamper.md — 7-lens checklist for iterating an existing thing.
• references/techniques/perspective-shift.md — Six Hats narrowed to dev contexts (on-call SRE, security, future-you, end user).
• references/anti-rationalization.md — 4 categories of skip-the-process excuses with counter-arguments.
• references/edge-cases.md — 8 documented cases where the obvious brainstorming approach is wrong.

Improvements

• Process steps reframed from prescriptive script to outcome-oriented list.
• "Hard gate" wording → "Behavioral hard rule" with explicit note that it is not hook-enforced (see gate-rules.md for actual gates).
• references/gotchas.md expanded from 6 → 12 entries (scope-explosion, question-fatigue, technique-mismatch, semantic-clustering, user-pre-decided, empty-intersection).

2.4.2 (2026-04-18) — Memory Fix

Highlights

Closes the second red-team round on the memory subsystem and corrects three published-doc inaccuracies introduced during the v2.4.1 rewrite.

Bug Fixes

• acquireLock now uses exponential backoff + jitter (10ms → 400ms cap, 8 retries). Eliminates the 20–40% concurrent-write drop rate.
• secret-scrub.sh DB-URL expression split per-scheme. BSD sed on macOS was rejecting the alternation, aborting the pipeline and discarding any content containing a DB URL.
• Stripe sk_live_ / sk_test_ / rk_ / pk_ patterns added to secret-scrub.sh to match the JS path. Shell paths previously leaked Stripe keys while the capture path redacted them.
• findMemoryDir sentinel check now fires at depth=0. Previously the walk could continue into a parent project, causing data loss on mewkit memory --clear from a nested dir.
• appendToQuickNotes now uses the same per-file lock as the other capture paths. Previously concurrent ##note: captures could interleave.

Improvements

• Auto-memory documentation corrected: memory/ directory (not MEMORY.md file), 200-line / 25 KB cap stated explicitly, subagent memory isolation documented, /memory (Claude Code) distinguished from ##prefix: (MeowKit), .claude/memory/ clarified as machine-local (gitignored — not team-shared).
• Removed contradictory "commit them via git" instruction in memory-system guide.

2.4.1 (2026-04-18) — Memory Simplification + Red-Team Hardening

Highlights

Deletes the auto-inject memory pipeline (memory-loader + parser + filter + injector) and replaces it with on-demand topic-file reads per consumer skill. Closes all 15 red-team findings (3 critical, 3 high, 9 medium) from the memory audit; most close by deletion.

Removals

• Auto-inject memory pipeline (memory-loader.cjs, memory-parser.cjs, memory-filter.cjs, memory-injector.cjs). UserPromptSubmit no longer runs a global memory-injection step.
• Lexical keyword-to-domain match retired with the filter; topic-file retrieval is now an explicit Read.
• Commit-message privilege escalation in post-session.sh (auto-tagged NEEDS_CAPTURE CRITICAL from commit message keywords, bypassing filter budgets).

Improvements

• Topic-file layout: fixes.md/json, review-patterns.md/json, architecture-decisions.md/json, security-notes.md replace the lessons.md + patterns.json monolith. Each file has a single consumer skill.
• On-demand retrieval: mk:fix, mk:cook, mk:plan-creator, mk:review read the relevant topic file via their SKILL.md Read step.
• Atomic capture writes: immediate-capture-handler.cjs uses temp-rename for all JSON writes; crash mid-write no longer corrupts split files. Dual-lock race on architecture-decisions.json eliminated.
• Memory is machine-local by default — .claude/memory/* is gitignored; mewkit setup scaffolds a blank directory. Downstream installs no longer inherit the MeowKit dev team's learnings.

Features

• New lib/secret-scrub.cjs with 16 regex patterns (Anthropic, OpenAI, Stripe, AWS, GitHub, JWT, DB URL, Bearer tokens, etc.). Wired into the capture path so leaked secrets no longer re-enter future session context.
• Fresh-install guard: handler auto-creates MEMORY_DIR on first run. Captures on blank projects no longer silently fail.

CLI

• packages/mewkit/src/commands/memory.ts — findMemoryDir exported with project-root sentinel and 5-level depth cap; clearMemory writes valid v2.0.0 skeletons; showStats / showSummary read the three split files.

Bug Fixes

• post-session.sh lock-failure fallthrough — the acquire_lock || branch could still execute the heredoc write. Block removed.
• Stale-lock false-eviction — dual stat failure fell back to always-stale. Safe mtime retry pattern replaces it.
• clearMemory wrote bare "[]" destroying the schema. Now writes proper skeletons per file scope.
• cost-log.json writer/spec schema drift fixed; session_id, model, cache_write_tokens, cache_read_tokens fields added.
• cost-log.json write now uses temp-file + os.replace rename for atomicity.
• SESSION_ID piped unvalidated into sed in conversation-summary-cache.sh — format validated against ^[a-f0-9-]{8,36}$ before use.

2.4.0 (2026-04-18) — The Agent Constitution Release

Highlights

docs/project-context.md becomes the single source of truth for every agent — a 286-line, 11-section "agent constitution" loaded at SessionStart by every agent. Resolves the agent-context-drift issue open since the 260411 audit. Full skill audit cycle: 64 findings, 61 resolved, 0 regressions.

Features

• docs/project-context.md agent constitution (tech stack, conventions, anti-patterns, testing, deployment, memory layout, hook chain). All 16 agents wired identically via ## Required Context.
• New SessionStart hook ensure-skills-venv.sh — idempotent bootstrap that creates .claude/skills/.venv if absent. Composes with npx mewkit setup.
• CLAUDE.md "Commands vs Skills" section documents the 3 valid command patterns (skill-composing, agent-invoking, standalone), preventing false-positive phantom flagging.

New Commands

• mk:project-context init — writes a TODO-filled docs/project-context.md skeleton for users starting from scratch. Refuses to overwrite an existing file.

Improvements

• 12 SKILL.md files got real domain-specific Gotchas sections (5–6 entries each, no generic filler): vue, typescript, database, build-fix, lint-and-validate, frontend-design, project-organization, jira, intake, figma, docs-finder, elicit.
• 7 gate-owning skills gained gate-rules.md references: mk:plan-creator, mk:workflow-orchestrator, mk:sprint-contract, mk:cook, mk:ship, mk:review. mk:cso gained security-rules.md.
• README + CLAUDE.md + project-context.md surface npx mewkit setup as the required post-install step.

Bug Fixes

• Deprecated skills (mk:debug, mk:documentation, mk:shipping) gained deprecated: true + superseded_by: YAML keys (previously only described in prose, invisible to parsers).
• Phantom skill refs in dispatcher fixed: mk:plan → mk:plan-creator, mk:test → mk:testing.
• Silent python3-absent skip in post-session.sh:27 upgraded to a warning.
• Silent-fail on missing docs/project-context.md in mk:plan-creator/step-02 upgraded to explicit warning with graceful fallback.

2.3.12 (2026-04-17) — External Codebase Packing + chom v2 Rigor

New Skills

Features

• mk:pack --compress — Tree-sitter signature extraction for API-surface queries (bodies omitted).
• self-pack-guard.sh blocks packing the current git root unless --self is passed.
• mk:chom v2 — 4 user-explicit modes (--compare / --copy / --improve / --port) replacing v1's 2-mode scheme.
• mk:chom speed flags: --lean (skip Phase 1 researcher for freeform inputs) and --auto (auto-approve non-HARD-GATE steps).
• mk:chom intent detection — keyword hints map to suggested mode flags.
• mk:chom explicit Boundary Rules — emits handoff text only, does NOT invoke plan-creator / brainstorming / cook / sequential-thinking mid-flow.

Improvements

• mk:chom Phase 4 HARD GATE language hardened — explicitly non-bypassable in all modes including --lean / --auto.
• chom handoff text enriched with challenge-reds summary + risk score. plan-creator owns adaptation-depth decisions downstream.

Bug Fixes

• Removed fabricated "Skills cannot call skills" claim from chom SKILL.md (contradicted by lessons-build-skill.md Composing Skills).
• Removed fabricated "40–70% context burn" claim from pack SKILL.md (replaced with honest context-isolation framing).
• Added Error Recovery for empty / unreachable / invalid sources (chom).

2.3.11 (2026-04-14) — Env Var Handling Hardening

Highlights

.claude/settings.json adopts Claude Code's native env field for team-shared defaults. Three-layer precedence: shell export > .claude/.env > settings.json env.

Features

• Native env field in .claude/settings.json for 9 control flags (MEOWKIT_TDD, MEOWKIT_BUILD_VERIFY, MEOWKIT_LOOP_DETECT, etc.). load-dotenv.sh / dispatch.cjs parsers now fallback-only for secrets and per-project overrides.
• project-context-loader.sh emits ## MeowKit Config block at SessionStart so the agent sees active control vars. Gated on new sessions only (not resume/clear/compact).
• MEOWKIT_HOOK_PROFILE alias introduced (legacy MEOW_HOOK_PROFILE still accepted).

Bug Fixes

• Quoted values with # preserved literally — MEOWKIT_API_KEY="abc#123" no longer truncated.
• Inline comments stripped from unquoted values only (VAR=on # comment → on).
• Indented keys trimmed ( MEOWKIT_TDD=1 now loads correctly).
• Dangerous keys blocked (PATH, LD_PRELOAD, LD_LIBRARY_PATH, DYLD_INSERT_LIBRARIES, IFS, BASH_ENV, ENV) — prevents env injection via rogue .env.
• Key validation against POSIX var name pattern.
• pre-implement.sh now loads .env via script-relative path fallback when CLAUDE_PROJECT_DIR is unset.
• Symlink-safe guard prevents walking into install source when .claude/ is symlinked.
• CWD mismatch guard added — warns if MeowKit hooks not detected at project root.

2.3.10 (2026-04-13) — Jira Ticket Intelligence + Confluence & Sprint Planning

New Skills

Features

• mk:jira evaluate — qualitative complexity assessment (Simple/Medium/Complex with Fibonacci range).
• mk:jira estimate — heuristic story point estimation with escalation triggers.
• mk:jira analyze — full ticket context analysis with structured RCA output.
• Inconsistency detection — missing AC, vague language, unlinked dependencies, contradictions.
• Injection defense — ticket content wrapped in DATA boundary markers.
• Goal-oriented decision tree replaces operations-centric SKILL.md.
• Add comment / add attachment as inline Tier 2 operations.

Improvements

• mk:jira SKILL.md restructured as thin routing layer (~150 lines).
• jql-patterns.md pruned from 50+ to 15 core patterns.
• sprint-operations.md and workflow-transitions.md pruned (REST details removed).
• Tier 2 batch creates (3+) now require preview + confirmation.
• Partial failure behavior defined for sequential operations.

2.3.9 (2026-04-12) — Memory System Hardening

Highlights

Memory loader split into 3 focused modules. 4 critical security/correctness fixes (tag escape, budget split, YAML validation, per-entry caps). Adds ##prefix: immediate capture, opt-in anchored summarization, and project preferences.

Features

• ##decision:, ##pattern:, ##note: message prefixes auto-route to typed memory files. Content validated against injection patterns before writing.
• quick-notes.md staging area for ##note: captures.
• MEOWKIT_SUMMARY_MODE=merge enables merge-based summarization (preserves earlier context across compressions). Default remains full-regen.
• .claude/memory/preferences.md — team-shared preferences loaded at SessionStart.
• Agent readiness banner — 5-point score (CLAUDE.md, project-context, test, lint, typecheck) shown at session start. Detects Node.js, Python, Rust, Makefile projects.

CLI

• mk:memory --prune — archives old standard-severity entries to lessons-archive.md, recovering injection budget. Critical entries exempt.

Bug Fixes

• Tag escape — MEMORY-DATA wrapper tags in content are escaped before injection, preventing DATA boundary escape.
• Budget split — 60% for critical entries, 40% for domain-filtered. One oversized entry no longer starves all others.
• YAML validation — malformed frontmatter entries rejected with visible [parse-errors:] marker instead of silent fallback.
• Per-entry caps — 3000 chars for critical (security findings preserved), 800 for standard.
• mkdir-based atomic locking for all memory file writes (POSIX portable; flock doesn't exist on macOS).
• O_EXCL checkpoint sequence lock — prevents TOCTOU race between concurrent checkpoint writers.
• Staleness filter — standard entries >6 months skipped (configurable). Critical entries never expire.
• Pattern expiration — patterns older than 12 months from lastSeen auto-expire. Critical/security patterns exempt.
• Cost-log append in post-session.sh with 1000-entry cap.
• Shell-to-Python injection eliminated across 6 hook files (single-quoted heredocs replace shell interpolation).
• Domain keyword extraction — 30+ domain keywords (api, auth, db, sql, etc.) bypass stop-word filter.

Environment Variables

2.3.8 (2026-04-12) — Multimodal Resilience, MiniMax & Provider Fallback

Highlights

Major overhaul of mk:multimodal — multi-provider generation with intelligent Gemini → MiniMax → OpenRouter fallback, MiniMax integration (image, video, TTS, music), document conversion, and MEOWKIT_ env namespace.

Features

• MiniMax image generation (image-01).
• MiniMax video generation (Hailuo 2.3) with async polling.
• MiniMax text-to-speech (speech-2.8-hd, 332 voices, 24 languages).
• MiniMax music generation (music-2.6).
• Intelligent provider router auto-selects Gemini/MiniMax/OpenRouter by available API keys; --provider flag forces one.
• Document → Markdown converter with batch mode (document_converter.py).
• Env-driven provider chains (MEOWKIT_IMAGE_PROVIDER_CHAIN etc.).
• OpenRouter fallback for image gen (opt-in via MEOWKIT_OPENROUTER_FALLBACK_ENABLED=true).
• API key rotation (MEOWKIT_GEMINI_API_KEY_2/3/4) for free-tier throughput (4x).
• Default image model: Nano Banana 2 (gemini-3.1-flash-image-preview).
• --resolution low-res flag for video analysis (62% token savings).
• Media pre-optimization via ffmpeg (optional, 10–20% savings).
• Cost estimation for video analysis in --verbose mode.

Improvements

• MEOWKIT_ env var prefix with backward-compat fallback to legacy names.
• DEFAULT_PROMPTS request structured JSON output (~50% more token-efficient).
• Output truncation enforced at 6000 chars (CJK-safe).
• media_resolution defaults per modality (image=high, pdf=medium, video=low).

2.3.7 (2026-04-12) — The Copy-Cat Release

New Skills

Features

• Smart input routing — auto-detects git URLs (clone + scout), web URLs (web-to-markdown), local paths, freeform text (researcher), screenshots (multimodal).
• 7 challenge questions — Necessity, Stack Fit, Data Model, Dependency Cost, Effort vs Value, Blast Radius, Maintenance Burden.
• Risk scoring — 0–2 proceed, 3–4 resolve first, 5+ reject.
• Two modes — --analyze (full workflow → Replication Spec) and --compare (analysis only → Comparison Report).

2.3.6 (2026-04-11)

Removals

• Unused files removed.

2.3.5 (2026-04-11) — CEO Review Layered Verification

Highlights

Redesigns mk:plan-ceo-review from single-pass deep review to layered verification pipeline. Strengthens decision quality without changing the 4-mode system.

Features

• Pre-screen gate (Layer 0–1) — mode-aware placeholder scan, structural completeness check, requirements coverage mapping. Returns for amendment, never rejects.
• Two-lens evaluation (Layer 3) — Intent Alignment + Execution Credibility. Each grades PASS/WARN/FAIL. Any FAIL → NEEDS REVISION.
• Severity tiers — all findings classified BLOCKER / HIGH-LEVERAGE / POLISH.
• Adversarial necessity — each section must surface ≥1 finding or document why clean. Prevents rubber-stamping.
• Append-only output — ## CEO Review block appended to plan.md (never overwrites).
• Merged Failure Analysis — Error & Rescue Map + Failure Modes Registry combined into single table with severity column.

Improvements

• plan-creator step-08 auto-suggests CEO review after plan creation (gated by planning_mode).
• harness step-01 suggests CEO review after product spec.

2.3.4 (2026-04-11) — Centralized Dotenv Loading

Highlights

Project-level .claude/.env support so all hooks and handlers can read MEOWKIT_* env vars without polluting shell profiles.

Features

• Shared dotenv loader lib/load-dotenv.sh sourced by all 11 shell hooks; no eval, uses printenv for safe key checking.
• Inline parser in dispatch.cjs loads .claude/.env for all 8 .cjs handlers (zero external dependencies).
• .env.example template with all 19 documented env vars across 5 categories (Core, Harness, Summary, Memory, Hook Controls).
• Precedence: shell export always wins over .env (no-override semantics).

Removals

• CLI package builds removed from release.sh — harness releases no longer trigger CLI package builds. Build step replaced with JSON config validation (settings.json, handlers.json, metadata.json).

2.3.3 (2026-04-11) — The Wiring Integrity Release

Highlights

5-agent parallel red-team audit of the full MeowKit harness. 7 critical breakpoints, 12 high-severity issues, and 30 medium/low cleanup items fixed across 25+ files.

Bug Fixes

• Gate 2 NON-NEGOTIABLE violation — fast.md and cost-saver.md modes auto-approved Gate 2 without human confirmation; now require explicit human approval.
• TDD sentinel persistence — --tdd flag wrote sentinel to .claude/session-state/ but session reset cleared session-state/ at project root (different dirs); sentinel now cleared on new session.
• Memory system dead by default — post-session.sh exited on standard profile, disabling memory capture, cost tracking, and trace records; now runs by default.
• Phantom agent dispatch — mk:cook referenced 4 nonexistent agents; remapped to real agents.
• Memory path wrong system-wide — CLAUDE.md + 19 skills referenced memory/ instead of .claude/memory/; all paths corrected.
• Model detector silent failure — model-detector.cjs guard on ctx.hook_event_name silently killed detection when field absent.
• Config file missing — meowkit.config.json referenced by 4+ consumers but never existed; created with version + features object.
• Budget thresholds — code defaults $10/$25, docs said $30/$100; aligned to $30/$100 with MEOWKIT_BUDGET_CAP override.
• 8 orphaned skills (api-design, build-fix, database, decision-framework, figma, intake, jira, verify) added to skill registry.
• 4 phantom skill refs in commands (/arch, /audit, /canary, /ship) fixed to actual skill names.
• ADR path conflict — /arch command wrote to docs/adrs/ but architect agent wrote to docs/architecture/adr/; unified.
• TURN_GAP default — harness-rules.md said :-5, code was :-30; docs aligned to code.

New Commands

• /harness — created missing slash command for primary green-field build entry point.

2.3.2 (2026-04-11) — The Agent-Skills Integration Release

Highlights

Integrates correctness patterns from Anthropic's agent-skills system: 6 core operating behaviors, per-skill failure catalogs, phase composition contracts, and lifecycle-aware skill routing.

Features

• core-behaviors.md — 6 mandatory operating behaviors (Surface Assumptions, Manage Confusion, Push Back, Enforce Simplicity, Scope Discipline, Verify Don't Assume) + 10 failure modes. Loaded via CLAUDE.md preamble.
• Per-skill failure catalogs — Common Rationalizations + Red Flags for mk:cook, mk:plan-creator, mk:review. Merged entries into mk:fix gotchas.
• Phase composition contracts — embedded in CLAUDE.md; documents expects/produces/breaks-if-missing per phase.
• Lifecycle routing table — task signal → phase → skill mapping in mk:agent-detector. Surfaced via mk:help.

2.3.1 (2026-04-11) — The Plan Creator Intelligence Release

Highlights

Plan-creator's biggest upgrade since v1.3.2. 4-persona red team, --deep mode, --tdd composable flag, standalone subcommands, and enhanced validation framework.

Features

• 4-persona red team — Security Adversary + Failure Mode Analyst added to existing 2 personas. Phase-count scaling: 1–3 phases = 2 personas, 4–5 = 3, 6+ = 4.
• Separate red-team-findings.md file with full 7-field detail, linked from plan.md summary.
• --deep mode — hard pipeline + per-phase scouting with file inventory and dependency maps per phase.
• --tdd composable flag — combines with any mode; injects Tests Before/Refactor/Tests After/Regression Gate into phase files.
• Memory capture at Gate 1 — planning decisions persisted after approval.
• Solution design checklist — 5-dimension trade-off analysis reference for Architecture/Risk sections.

New Commands

• /mk:plan red-team {path} — runs adversarial review on existing plans.
• /mk:plan validate {path} — runs critical question interview on existing plans.
• /mk:plan archive — scans completed plans, optionally captures learnings, archives or deletes.

Bug Fixes

• step-03 duplicate section label — 3i appeared twice (Parallel + Two-Approach); renamed to 3i/3j.
• phase-template wrong step reference — hydration reference said step-05 instead of step-08.
• step-08 incomplete schema — .plan-state.json missing deep and product-level as valid planning_mode values.

2.3.0 (2026-04-11) — The Hook Dispatch Release

Highlights

Node.js hook dispatch system with 8 handler modules, cook verification flags, review skeptic anchoring, structured memory filtering, and tool output limits. TDD enforcement is now opt-in.

Features

• Central dispatch.cjs with handlers.json registry — parses stdin once, routes to 8 handlers across 4 lifecycle events.
• model-detector handler — auto-detects model tier + density from SessionStart stdin model field; replaces MEOWKIT_MODEL_HINT as primary source.
• orientation-ritual handler — resumes from checkpoint on session resume.
• build-verify handler — compile/lint after file edits, cached by file hash (ported from shell to Node.js).
• loop-detection handler — warns at 4 edits, escalates at 8.
• budget-tracker handler — token cost estimation with $10 warn / $25 block session-level thresholds.
• auto-checkpoint handler — crash-recovery every 20 tool calls + phase transition detection.
• checkpoint-writer handler — sequenced checkpoint with git state + budget snapshot on Stop.
• cook --verify — advisory browser check after review (~$1).
• cook --strict — full mk:evaluate after review; FAIL blocks ship (~$2–5).
• cook --no-strict — suppress auto-strict trigger.
• Auto-strict — scale-routing level=high auto-enables --strict in cook.
• Review skeptic anchoring — re-anchor prompt injected per adversarial persona dispatch.
• Structured memory — lessons.md YAML frontmatter with two-phase domain-filtered loading.
• Tool output limits — Glob head_limit=50, Grep head_limit=20, Read offset+limit for >500 lines.

Improvements

• TDD now opt-in via --tdd flag or MEOWKIT_TDD=1; default mode skips RED-phase gate.

2.2.2 (2026-04-10) — Homoglyph Detection Refinement

Improvements

• mk:web-to-markdown injection_detect.py homoglyph detection now flags only mixed-script tokens (Latin + Cyrillic within a single word) instead of consecutive foreign characters. Reduces false positives on legitimate multilingual content while still catching homoglyph spoofing.

2.2.1 (2026-04-10) — Bug Fix

Bug Fixes

• mk:web-to-markdown robots_cache.py _fetch_robots_txt() raised UnboundLocalError. A function-local import urllib.request shadowed the module-level urllib.robotparser binding, breaking rp = urllib.robotparser.RobotFileParser(). Hoisted import urllib.request to module-level imports.

2.2.0 (2026-04-08) — Generator/Evaluator Harness

Highlights

Largest architectural addition since 1.0.0. Autonomous multi-hour build pipeline, adaptive scaffolding density per model tier, middleware layer, trace-driven meta-loop, and a conversation summary cache — without loosening any hard gates.

New Skills

New Agents

• evaluator — skeptic persona distinct from reviewer. Drives running build, re-anchors persona per criterion, propagates FAIL verdicts hard. Self-evaluation forbidden — always runs in fresh context.

New Commands

• /mk:summary — conversation summary cache inspector. --status health check, --force re-summarize, --clear reset.

Features

• 4 new middleware hooks: post-write-build-verify.sh, post-write-loop-detection.sh, pre-completion-check.sh, conversation-summary-cache.sh (Haiku-powered, secret-scrubbed, throttled by size/turns/growth).
• 2 new rules files: harness-rules.md (11 rules for generator/evaluator discipline), rubric-rules.md (10 rules for calibration).
• Adaptive density auto-selected per model tier (TRIVIAL=MINIMAL, STANDARD=FULL, COMPLEX/Opus 4.5=FULL, COMPLEX/Opus 4.6+=LEAN). Override via MEOWKIT_AUTOBUILD_MODE.

CLI

• Schema-driven system-deps registry (packages/mewkit/src/lib/system-deps-registry.ts) — replaces hardcoded ffmpeg/imagemagick install paths with a typed registry. Skills declare optional_system_deps; CLI parses and validates against the registry.
• mewkit doctor — generic registry loop replaces hardcoded checks. Playwright entry has dedicated two-probe doctorCheck.
• mewkit init + setup --system-deps — flat list prompt; FFmpeg / ImageMagick / Playwright + Chromium iterated from registry insertion order.

Bug Fixes

• privacy-block.sh exit code corrected — was using exit 1 (non-blocking per Claude Code hooks docs). Changed to exit 2 so the hook actually blocks. Block messages moved stdout → stderr.

Migration Notes

• export MEOWKIT_MODEL_HINT=opus-4-6 in your shell profile if on Opus 4.6 — enables LEAN density auto-detection. Without it, Opus 4.6 users silently get FULL density.
• Try /mk:autobuild "build me a THING" for your next green-field build.
• Run /mk:summary --status after your first long session to verify the conversation cache is healthy.

Breaking Changes

• Hooks now read JSON on stdin via lib/read-hook-input.sh instead of positional $1. Legacy fallback preserved for existing hooks; custom hooks should migrate.
• No CLI, agent, or skill syntax changes. No gates loosened.

2.1.0 (2026-04-04)

Highlights

Custom statusline, dependency management, SEO, and mewkit CLI improvements.

Features

• .claude/statusline.cjs — 5-line ANSI status bar for Claude Code: model+tier, context usage bar with /clear warning at 60%/80%, active plan+phase tracking, 5h/weekly rate limits with reset countdown, token usage breakdown.
• Settings merge preserves new top-level keys (like statusLine) during updates.
• Sitemap generation, robots.txt, OG/Twitter meta tags, canonical URLs.

CLI

• Install prompt during init — asks "Install Python skill dependencies?" after project description (default: no). Installs into .claude/skills/.venv only.
• Per-skill requirements.txt — walks skills/*/scripts/requirements.txt, merges and deduplicates with input validation.
• mewkit setup --only=deps — manual re-run with smart skip (verifies already-installed).
• mewkit doctor pip check — verifies installed pip packages against expected skill dependencies.
• Version picker — shows top 4 versions + "Enter version manually..." option.
• Cross-platform Python detection — where on Windows, py launcher support.
• Security — package name validation, path traversal prevention, 120s pip timeout, execFileSync array args.

2.0.0 (2026-04-04) — The Leverage Release

Highlights

Extracted high-leverage patterns from ECC's 38-agent ecosystem. 5 new skills, 17 reference merges, hook profiling, naming cleanup.

New Skills

Features

• 17 reference merges across 10 skills (RCA selection, plan-creator ops/cold-start/mutation, QA browser checklist, agent-detector token budget, office-hours product lens, typescript review checklist, cook loop safety, review iterative protocol, frontend anti-slop, testing E2E best practices).
• Hook runtime profiling — MEOW_HOOK_PROFILE env var: strict (all), standard (default), fast (gate + privacy only). Safety-critical hooks never skip.
• Mandatory simplification — mk:cook now requires mk:simplify between Phase 3 (Build) and Phase 4 (Review).
• Proactive learning — learning-observer.sh PostToolUse hook detects churn patterns, feeds into retroactive capture.

Removals / Renames

• mk:shipping → mk:ship.
• mk:documentation → mk:document-release.
• mk:debug → mk:investigate.
• (Redirects in place for 2 releases.)

Improvements

• MICRO-TASK TDD exemption — non-production code <30 lines exempt from TDD if classified MICRO-TASK by orchestrator.
• Staged parallel mode — overlapping files handled sequentially, non-overlapping in parallel.
• Memory capture enhancement — budget 2 min → 5 min, markers 3 → 5, CRITICAL/SECURITY markers always processed, --capture-all flag.
• mk:scale-routing — 4-layer detection (CSV + task content + context + confidence scoring), 8 task type classifications, optional product-areas.yaml.

CLI

• npx mewkit init now prompts for optional system deps (FFmpeg, ImageMagick).
• npx mewkit setup --system-deps for deferred install.
• npx mewkit doctor reports status.

1.4.0 (2026-04-03) — The Plan Intelligence Release

Highlights

Dedicated plan red-team with CK-style adjudication, plan-specific personas, and new workflow modes.

Features

• Plan red-team extraction — monolithic step-04 split into steps 04–07; dedicated step-05-red-team.md with 7-field findings (Severity / Location / Flaw / Failure Scenario / Evidence / Suggested Fix / Category).
• Agent adjudication (Accept/Reject + rationale), 3-option user review gate (Apply all / Review each / Reject all), deduplication, severity sorting, 15-finding cap.
• 2 new plan-specific personas: plan-assumption-destroyer (unvalidated scale, dependency, team, infrastructure, timeline, integration assumptions) and plan-scope-complexity-critic (YAGNI violations, over-phasing, scope creep, premature abstraction).
• Dynamic persona scaling — phase-count thresholds: 1–3 phases = 2 personas, 4–5 = 3, 6+ = 4.
• Red Team Review section — auditable finding table written to plan.md.
• workflow-fast.md — compact path (skips research, scout, red-team, interview).
• --parallel mode — file ownership matrix in plan.md ## Execution Strategy, parallel group task hydration, max 3 groups.
• --two mode — 2 competing approach files + trade-off matrix; user selects approach at step-04.

Improvements

• mk:plan-creator workflow expanded from 6 steps (00–05) to 9 steps (00–08).
• .plan-state.json schema bumped to v1.1 with optional parallel_groups and selected_approach fields.
• 6 new gotchas and 2 new reference files (parallel-mode.md, two-approach-mode.md).

1.3.4 (2026-04-02) — Hook path resolution fix

Bug Fixes

• All hooks use $CLAUDE_PROJECT_DIR for absolute paths in settings.json and CWD guard. Fixes "No such file or directory" when CWD differs from project root.

1.3.3 (2026-04-02) — The Hook Safety Release

Bug Fixes

• cost-meter.sh — always exited 1 because settings.json passes no arguments; now exits 0 for missing args.
• post-write.sh — exited 1 on empty/missing file path; now exits 0 (matches PreToolUse safety fallback).
• pre-task-check.sh — used exit 2 for WARN findings; Claude Code treats non-zero as error; now exits 0.

1.3.2 (2026-04-01) — The Plan Quality Release

Highlights

Complete redesign of mk:plan-creator to match/exceed ck-plan across 15 dimensions.

Features

• Step-file architecture — SKILL.md (thin entry) + workflow.md + 6 step files. JIT loading.
• Multi-file phase output — plan.md overview (≤80 lines) + phase-XX files (12-section template each).
• Scope challenge — Trivial → exit, simple → fast, complex → hard. User chooses EXPANSION / HOLD / REDUCTION.
• Plan red team — 2 adversarial personas (Assumption Destroyer + Scope Critic) review plans before validation (hard mode).
• Research integration — bounded (2 researchers, 5 calls each), findings cited in phase Key Insights.
• Sync-back — .plan-state.json checkpoint enables cross-session resume.
• Critical-step tasks — [CRITICAL] / [HIGH] todo items get dedicated Claude Tasks.
• Richer frontmatter — description, tags, issue, blockedBy / blocks fields.

1.3.1 (2026-03-31) — The Red Team Depth Release

Highlights

Hybrid adversarial persona system for mk:review.

Features

• Scope gate — step-01 classifies diffs as minimal (≤3 files, ≤50 lines, no security, domain ≠ high) or full. Minimal runs Blind Hunter only.
• Hybrid persona system — Phase B: 4 adversarial persona subagents (Security Adversary, Failure Mode Analyst, Assumption Destroyer, Scope Complexity Critic) run after base reviewers, informed by Phase A findings. 2-at-a-time batching.
• Forced-finding protocol — zero findings triggers 1 re-analysis with "look harder" prompt. Prevents rubber-stamp approvals.
• 4-level artifact verification — Exists, Substantive, Wired, Data Flowing checks in verdict step.

1.3.0 (2026-03-31) — The Integration Integrity Release

Highlights

Full red-team audit: 98 components, 11 batches, 43 criticals found, 42 fixed.

Bug Fixes

• Hooks enforcement restored — gate-enforcement.sh and privacy-block.sh were completely non-functional (argument mismatch). Fixed argument passing; all 9 hooks now registered and working.
• Agent naming standardized — 5 phantom subagent_type values in Task() calls mapped to real agents.
• 7-phase model everywhere — workflow-orchestrator migrated from 5-phase to 7-phase. Gate 2 no longer bypassable.
• Path consistency — plan files, memory, ADRs, scripts all use canonical full paths.
• Verdict taxonomy unified — PASS/WARN/FAIL everywhere. Review dimensions aligned (Correctness/Maintainability/Performance/Security/Coverage).
• Python venv enforced — all scripts use .claude/skills/.venv/bin/python3. SessionStart warns if missing.
• pre-ship.sh guarded — no longer runs test suite on every Bash call, only on git commit/push.
• Security BLOCK → FAIL — security agent BLOCK verdict now automatically fails Gate 2 Security dimension.

Improvements

• Created missing templates: party prompts (agent-selector, synthesis), team-config ownership map.
• Created mk:fix/references/gotchas.md (7 anti-patterns).
• Fixed mk:development/references/skill-loader.md — all 13+ broken paths corrected.
• Fixed mock guidance contradiction in tester agent (unit tests may mock, integration tests must not).
• Honest documentation: mk:careful now states 8/30 patterns are hook-enforced (was claiming all 30).

1.2.1 (2026-03-31)

Bug Fixes

• mk:cook Phase 6 (Reflect) now spawns a dedicated subagent for mk:memory session-capture. Previously memory write was an inline bullet point that could be skipped if session was interrupted. Now enforced as MUST-spawn.

1.2.0 (2026-03-31) — The Memory Activation Release

Highlights

Fixed the dormant memory system and enriched it with cross-framework insights from 6 agent frameworks.

Features

• Memory capture pipeline — post-session.sh now writes structured NEEDS_CAPTURE markers instead of invisible HTML comment placeholders.
• Retroactive capture — Phase 0 processes pending markers from previous sessions (max 3 markers, 2-min budget).
• Live capture — Phase 5 captures non-obvious decisions, corrections, and rejected approaches before shipping.
• 3-category extraction — patterns, decisions, or failures.
• New patterns.json fields — category, severity, applicable_when (all optional, backward compatible).
• Stronger promotion criteria — patterns promoted to CLAUDE.md only when frequency ≥ 3, severity = critical OR frequency ≥ 5, generalizable, saves ≥ 30 min. Human approval still required.
• Consolidation rubric — 4-branch classification (clear match auto-merge / ambiguous ask / no match create new / no signal skip).

1.1.0 (2026-03-30) — The Reasoning Depth Release

Highlights

Deeper review reasoning, resumable builds, and systematic coverage mapping. Inspired by comparative analysis of BMAD-METHOD, ClaudeKit-Engineer, Khuym Skills, and Get-Shit-Done.

New Skills

Features

• mk:review now recommends running mk:scout before review for complex changes (3+ files).
• After review verdict, users can run mk:elicit for deeper analysis through a named reasoning method.
• Beads pattern — COMPLEX tasks (5+ files) decompose into atomic, resumable work units. Each bead has acceptance criteria, file ownership, and ~150 lines size. Progress tracked in session-state/build-progress.json. Interrupted builds resume from last completed bead.
• Subagent Status Protocol — all subagents report structured status (DONE, DONE_WITH_CONCERNS, BLOCKED, NEEDS_CONTEXT) with controller handling rules.

Improvements

• Sub-agent type classification — support agents now have subagent_type in frontmatter: advisory (brainstormer, researcher, ui-ux-designer), utility (git-manager), escalation (journal-writer).
• Pre-delegation checklist added to orchestration-rules.md: work context, plan reference, file ownership, acceptance criteria, constraints.

1.0.0 (2026-03-30) — The Disciplined Velocity Release

Highlights

The biggest MeowKit update yet. 13 new capabilities inspired by deep analysis of BMAD-METHOD and ClaudeKit-Engineer. Theme: scale throughput while maintaining absolute discipline.

New Skills

Features

• Planning Depth Per Mode — 7 modes declare researcher count: strict / architect run 2 parallel researchers; default / audit run 1; fast / cost-saver / document skip research.
• Multi-Layer Adversarial Review — mk:review now runs 3 parallel reviewers (Blind Hunter, Edge Case Hunter, Criteria Auditor) with post-review triage. Catches 2–3x more bugs than single-pass review.
• Anti-Rationalization Hardening — agents cannot downgrade complexity, minimize tests, skip security, or dismiss WARN verdicts without 3-part justification.
• Project Context System — docs/project-context.md is the agent constitution. All agents load it at session start.
• Parallel Execution & Teams — COMPLEX tasks with independent subtasks run up to 3 parallel agents with git worktree isolation. Integration test required after merge.
• Step-File Architecture — complex skills decompose into JIT-loaded step files. First skill: mk:review (4 steps).
• Hook-Based Enforcement — 3 shell hooks upgrade behavioral rules: privacy-block.sh (blocks sensitive reads), gate-enforcement.sh (blocks writes before Gate 1), project-context-loader.sh (auto-loads context).

Breaking Changes

• None. All additions are backward-compatible.

0.1.2 (2026-03-29)

Features

• Interactive version selection when running npm create meowkit@latest.
• git-manager agent for commit/push workflows.
• Confirmation step before Gemini API key input.

0.1.1 (2026-03-29)

Removals

• Excluded runtime dirs (session-state, memory, logs) from release zip and git tracking.

0.1.0 (2026-03-29)

Highlights

Initial pre-release of MeowKit agent toolkit.

Features

• Core skill set (cook, fix, ship, review, memory, testing).
• Sequential thinking and fix diagnosis references.