CHANGELOG.md

Changelog

0.16.0

Core protocol and architecture routing

• do-it-core: added a single protocol voice for eight execution rules, backed by the closed-set execution-failure-modes.tsv registry and a consistency validator that rejects missing, mutated, duplicated, or restated rule bullets across the core skill, hooks, bridges, and satellite skills.
• do-it-architecture: added load-bearing architecture guidance for authority and ownership, public or persisted contracts, dependency boundaries, migration and cutover, guards, deletion, recovery, and drift.
• Model-adaptive Standard advisory: strong-model hosts receive one compact do-it-core pointer; weak, unknown, or undetected models receive the canonical rule text inline.
• Architecture routing: action-shaped Heavy interface/schema, migration/cutover, and security-boundary changes point to do-it-architecture; explanation questions remain Light, and ordinary Heavy release work stays free of an architecture pointer.
• Prompt-state ordering: Claude, Codex, Cursor, OpenCode, Pi, and Kimi now invoke router then Heavy grill through one serialized prompt-submit hook, preventing concurrent hook scheduling from reading a stale tier; the entrypoint stays fail-open (router guidance is always emitted, grill is gated on state persistence), and state writes use a portable mkdir lock fallback with a bounded wait when flock is unavailable.
• Architecture routing precision: strong Heavy detection now uses bounded, clause-local action/surface pairs for interface, migration, and security work, avoiding substring joins such as token + relationship while still recognizing explicit security-boundary creation.

0.15.0

Code-quality rules move into skills and hooks

• New file-size advisory family in write-quality-lint: whole-file line count against warn (>500) and split (≥800) thresholds, nudging phase-based splits before behavior is added to oversized files. Project override via the data-only .do-it/write-quality.local.tsv (git root of the edited file) or the DO_IT_FILE_SIZE_WARN_LINES / DO_IT_FILE_SIZE_SPLIT_LINES env vars; each level falls back independently, CRLF/no-EOL files tolerated.
• do-it-code-quality: new "Phases, Not Piles" section — compute → apply → audit separation, reuse before repeat, a three-question post-edit re-check, and function-length judgment guidance; scope chain gains a targeted-reads step (no whole-file reads on large files).
• do-it-skill-authoring: new "Where a Rule Lives" — judgment rules in skills, deterministic numbers in hooks, drift contracts in validators; the family reference table is the single prose home for hook-owned numbers.
• Repo root gains AGENTS.md with repo-local contributor facts (source of truth vs build artifacts, closed-set contracts, verification bar).

0.14.2

Release hygiene

• Post-0.14.1 source bump with corrected README/release truth planes for the tagged v0.14.1 publication path, then ship 0.14.2 metadata for the next independent root / OpenCode / Pi npm release.
• Portable Pi smoke npm invocation formatting cleanup on Windows path join.
• Stabilize Pi bridge timeout diagnostic assertion when a setsid grandchild races process-tree cleanup against the grace timer.

0.14.1

Pi host adapter and npm delivery

• Pi package: added independently publishable @tdwhere/do-it-pi with one TypeScript extension, shared skills/hooks, prompt templates, and ten portable package agents discovered under do-it.* names by optional pi-subagents.
• Lifecycle: root sessions receive bootstrap/router/grill, write-quality advice, and a soft next-turn verification reminder; child processes identified by PI_SUBAGENT_CHILD=1 receive only the bounded subagent stance.
• Runtime hardening: ToolResult content remains structured; hook execution is async, abortable, aggregate-output-bounded, and requests Unix process-group or Windows taskkill /T /F tree termination. A bounded cleanup grace closes hook pipes and reports unconfirmed descendant cleanup rather than waiting forever. Bash discovery supports Git for Windows and an explicit DO_IT_BASH override.
• Mapping: Scout handles reconnaissance; code-mapper uses bounded trace/thorough closure with resumable frontiers instead of package-owned fixed turn, tool, or timeout budgets.
• Packages: release automation now verifies and publishes separate npm artifacts for @tdwhere/do-it, @tdwhere/do-it-opencode, and @tdwhere/do-it-pi; Linux and Windows CI exercise Pi build, discovery, hooks, package contents, and optional-dependency degradation.

Kimi Code host adapter

• Root plugin: repository-root kimi.plugin.json installs via /plugins install with no build step — skills, commands, and five advisory hooks; custom subagents are not shipped (Kimi has no custom subagent mechanism).
• Protocol: do_it_emit_context emits plain text under Kimi (no additionalContext); UserPromptSubmit prompt may be a ContentPart array; verification-gate reads session wire.jsonl when transcript_path is absent.
• Validate: npm run validate:kimi-plugin is wired into test, prepack, CI, and release.

Adapter hygiene

• Hook script allowlists and Codex hooks.json commands now share scripts/lib/hook-manifest.mjs; Codex hook paths fall back through PLUGIN_ROOT → CLAUDE_PLUGIN_ROOT → . so an empty expansion cannot become /hooks/....
• manifest.json commonExtras plus per-target overrides; session-dir resolution is aligned across hooks, install/manage.mjs, and the OpenCode bridge (including $KIMI_CODE_HOME/do-it-data/sessions).
• index.json total_skills counts runnable skills only (9); the generated _index.md discovery entry stays in total_discovery_entries.

Skill absorptions (no tenth skill)

• do-it-code-quality: merge-conflict hunk-by-intent rule; name a production evidence surface when shipping runtime features.
• do-it-verify: production-run claim shortcut; deferred-marker sweep on branch closeout.
• do-it-handbook: optional decisions.md promotion when a grill route settles permanently.
• do-it-decide: cite primary sources / capture durable research; doubt applies in-flight, not only after the diff.
• docs/upstream-map.md: lineage principle (convergence ≠ lineage) plus stale-name fixes and inspired-by rows for this round only.

Feedback, routing, and safety profiles

• Direct intent wins: a why/explanation prefix no longer downgrades a concrete fix/edit request to Light; Chinese and English mixed requests have regression coverage.
• Action boundary: explicit no-write / 先不改 state persists until an actual reopen instruction, not a question about whether editing is possible.
• Opt-in retrospective: added default-off, silent, project-local redacted feedback capture plus do-it-retrospective on|off|status|report. Reports propose at most three lessons and never edit durable instructions without confirmation.
• External actions: added a default-off Claude-only named-command PreToolUse profile. ask requests host confirmation; deny is an explicit stop. Codex, Cursor, and OpenCode documentation now points to their real native permission boundaries instead of claiming a universal hook veto.
• Release guard: registered hook sources must be Git-tracked, so a clean checkout cannot silently omit a newly wired hook.

Fixes — Cursor on native Windows

• Install path: install-cursor-local.mjs no longer rewrites %USERPROFILE% into /mnt/c/... on win32 (that landed under e.g. D:\mnt\c\Users\...). WSL→Windows mirroring only runs on Linux when WSL//mnt/c/Users is present.
• Hooks registration: after local/CLI install, merge do-it entries into user-level ~/.cursor/hooks.json (Cursor Hooks service does not load plugin-local hooks/hooks.json today).
• Windows hook entrypoint: all Cursor hook commands go through polyglot hooks/run-hook.cmd <name> (never bare .sh). Cursor on Windows was opening .sh paths in the editor / “Open with” dialog instead of executing them. Runner prefers Git for Windows bash and skips System32\bash.exe.
• Index build: build-index-json normalizes CRLF SKILL.md frontmatter so Windows checkouts do not fail name/description validation.
• Git Bash keyword match: do_it_prompt_has_word uses pure bash on MSYS/MINGW to avoid grep -q + pipefail “Aborted” noise.
• Quality follow-ups: Windows-safe atomic hooks.json replace; no WSL multi-user /mnt/c/Users scan; fail closed on partial multi-home install; do-it hook ownership keyed only on do-it-cursor/hooks/; quote $/backtick paths; Windows CI isolates USERPROFILE for setup/doctor.

0.14.0

Theme: meaning-centric rethink. Skills organize by what they mean for the task instead of a mandatory process chain. Main defense moves to write-time quality; plugin marketplace is the primary install path on all four hosts.

Skills — meaning buckets

Eight installed skills (5 core + 3 extended):

Bucket	Skill
Route	do-it-router
Write defense	do-it-code-quality
Review / repair	do-it-review
Decide	do-it-decide
Verify / close	do-it-verify
Handbook / names	do-it-handbook, do-it-context
Maintain	do-it-skill-authoring

Router stays thin: pick Light / Standard / Heavy, then self-select buckets. No mandatory brainstorm → grill → plan → orch pipeline on Standard.

Migration (old skill → new skill)

Old skill	New skill
do-it-grill	do-it-decide
do-it-brainstorm	do-it-decide
do-it-planning	do-it-decide
do-it-slicing	do-it-decide
do-it-review-loop	do-it-review
do-it-fix-loop	do-it-review
do-it-verification-gate (skill)	do-it-verify
do-it-branch-closeout	do-it-verify
do-it-tdd	do-it-code-quality
do-it-debugging	do-it-code-quality
do-it-architecture-scan	do-it-code-quality
do-it-interface-drill	do-it-code-quality
do-it-worktree-isolation	do-it-code-quality
do-it-comments-discipline	do-it-code-quality
do-it-codebase-design	do-it-code-quality
do-it-subagent-orchestration	(removed — parent delegation contract + subagent-stance hook)

do-it-router, do-it-handbook, do-it-context, and do-it-skill-authoring keep their names. The verification-gate hook remains; only the skill file was folded into do-it-verify.

Hooks

• grill-prompt: injects on Heavy only (or explicit grill request); Standard stays silent.
• grill-pretool removed — no PreToolUse / preToolUse plan-file gate.
• write-quality-lint families overhauled: merged narrative comment families, raised case-list / edit-bloat thresholds, added scope-chain, live-path, secret-leak, type-escape, test-fiction.
• Codex / Claude / Cursor / OpenCode plugin bundles carry hooks; Codex operators install the plugin and trust hooks under /hooks.
• Retained: router, subagent-stance, verification-gate, write-quality-lint.

Agents

Agents reduced to 10 (from ~23). Retained set absorbs merged lenses:

• Decide: product-strategist, architecture-strategist, plan-challenger
• Write: code-mapper, code-quality-cleaner, tdd-red-writer
• Review: reviewer, red-team-reviewer, spec-compliance-reviewer
• Docs: documentation-engineer

Install

• Plugin marketplace is primary for Codex, Claude Code, Cursor, and OpenCode.
• Cursor ships the full 8 skills (ALL_SKILLS), same as the other hosts. Local install must be a real copy under ~/.cursor/plugins/local/ (Cursor rejects external symlinks); use node scripts/install-cursor-local.mjs.
• do-it setup is optional / legacy (doctor, temp-home smoke, migration).
• Removed the v1 guidance to pair Codex plugin install with global setup because plugin_hooks=false.

0.13.1

Fixed

• verification-gate turn slicing: guard against invalid user line numbers so a bad transcript offset cannot widen the evidence window incorrectly.
• write-quality scan hardened for copy-paste git grep and temp-path fallback edge cases.
• Skill references/ link targets corrected so progressive-disclosure sheets resolve from installed skill paths.

Changed

• OpenCode plugin build requires tsc (no silent skip when compile is needed).
• prepack expands Cursor and OpenCode plugin builds alongside the Codex plugin and agent-bundle validation.

0.13.0

Added

• Four-host adapters: Codex, Claude Code, Cursor, and OpenCode share one workflow kernel. Cursor marketplace bundle under plugins/do-it-cursor/; OpenCode TypeScript plugin under plugins/do-it-opencode/.
• Shared skill references/ sheets (integrity, dimensions, write-quality families, per-host install notes).
• Harness adapter matrix at docs/harness-adapter-matrix.md.

Changed

• PostToolUse comments-lint and anti-patterns-lint merge into advisory write-quality-lint (single reminder per file; tier/DIM gated). Legacy wrappers exec into the merged script for one release.
• UserPromptSubmit injection compressed for Standard turns.

0.12.0

Added

• do-it-codebase-design skill — deep-module vocabulary (module, interface, seam, adapter, leverage, locality), wired into router, architecture-scan, and review-loop.
• subagent-stance UserPromptSubmit hook — compact stance reminder for delegated work.

Changed

• do-it-grill and do-it-review-loop tightened with leading words and checkable completion criteria.
• Handbook auto-bootstrap on Standard/Heavy greenfield code turns.
• Handbook simplified to stable project truth (invariants / architecture / glossary) plus worklog template; backlog / code-map / runtime-status templates removed.
• code-mapper returns temporary path maps only (no handbook write claim).

Removed

• code-map-refresh PostToolUse hook.

0.11.0

Fixed

• macOS CI: the anti-patterns-lint hook is now BSD-portable. The no-consumer check uses git grep's own -w word match instead of a GNU \b regex, and file paths are canonicalized with pwd -P before repo-relative stripping so the /var→/private/var symlink no longer breaks self-exclusion. The macOS hook-test job was failing on these for three releases.

Changed

• Handbook slimmed from 12 templates to 7 project-truth files. The five process docs that duplicated skills were removed: the task-card layout now lives in do-it-planning, the review protocol in do-it-review-loop, the dispatch contract in do-it-subagent-orchestration, the execution pipeline in do-it-router, and the handbook maintenance rules in the do-it-handbook skill.
• Three low-utilization skills consolidated (23 → 20): do-it-grill-log folded into do-it-grill (Grill Log Artifact), do-it-domain-language into do-it-context (Domain Glossary Mode), and the optional do-it-visual-planning into do-it-planning (Visual Aids); the visual companion (local browser server + render templates) moves under do-it-planning and its templates were visually redesigned. Existing installs remove the deprecated skills on upgrade.
• do-it-router gains a § Restraint principle: advisory over blocking, reuse over add, no unbounded libraries, check git intent before deleting, and capabilities that auto-fire instead of requiring a /command.
• do-it-context gives CLAUDE.md a formal role in the source-of-truth hierarchy, draws the CLAUDE.md / CONTEXT.md / handbook boundary, and can scaffold a lean CLAUDE.md on demand (additive, never auto-write).

Added

• Advisory nudges (one-shot, never block): after grill on durable-plan work with no plan card, a reminder to land the plan card before implementation drifts ahead; on established projects (.do-it/CONTEXT.md or handbook present) or port/restore prompts, a reminder to read existing structure / grep current code first.
• comments-lint now flags phase N, wave N, and BL-NNN stage markers as task-ref comments.

0.10.0

Added

• Evidence Ledger rows for durable .do-it/plans/<task>.md cards. Heavy, release/install, multi-agent, and explicit durable-plan work can now track each claim by readiness target, truth plane, evidence, result, owner, and residual risk without adding another skill.
• Truth-plane vocabulary for source, task worktree, integration worktree, temp install, live Codex, live Claude, package artifact, host behavior, and external blockers.
• Subagent lane status vocabulary (assigned, running, done_with_evidence, integrated, blocking) so parent closeout can wait for, inspect, and accept delegated work instead of trusting worker DONE.

Changed

• Agent templates are model-agnostic. Source agents/*.toml files no longer pin concrete models or reasoning-effort fields, and Claude generated agents omit model: frontmatter so the host owns model selection.
• do-it-planning, do-it-verification-gate, do-it-review-loop, do-it-fix-loop, do-it-interface-drill, do-it-slicing, do-it-subagent-orchestration, and do-it-branch-closeout now share the evidence-ledger, proof-path, truth-plane, and lane-status contract.
• Release and maintenance docs now require source, package, temp install, live install, and host behavior readiness to be claimed separately.

Fixed

• code-mapper is consistently read-only. It now returns proposed handbook code-map updates to the parent instead of claiming permission to write from a read-only lane.
• Live installs now manage individual do-it hook files instead of claiming the entire hooks/ directory, so non-do-it hook integrations can coexist.
• scripts/validate-agent-bundle.mjs rejects model pins and host-private model/budget fields in portable agent templates and generated Claude agents.
• manifest.json carries a 0.9.x install-state migration range for the 0.10.0 minor upgrade.

0.9.1

Added

• index.json — generated machine-readable inventory for the do-it skill and agent surface. It records package/version metadata, skill group coverage, optional skills, and every manifest-backed skill/agent entry with source and target paths.
• scripts/build-index-json.mjs — builds and validates the root inventory from manifest.json, skill frontmatter, and agent TOML descriptions. Wired into npm run build:generated and package validation.
• Install tests now cover cross-device staged replacement (EXDEV) and the generated index/docs cleanup contract.

Changed

• index.json is included in the npm package file list so external runtimes and marketplaces can discover do-it's installed skill/agent surface without loading every SKILL.md.
• Removed the stale install/migrations/ note directory from the package. Install-state migrations remain active in manifest.json and install/migrate.mjs; the removed file was historical human documentation.

Fixed

• install/manage.mjs now falls back to copy-and-remove when replacing a staged managed target fails with EXDEV, which can happen when temp/staging and the target install root are on different filesystems.

0.9.0

Added

• install/migrate.mjs — version-migration logic (parseMinor, matchesFromRange, needsMigration, applyMigrationAction, applyMatchingMigrations) extracted from install/manage.mjs as pure, importable functions.
• tests/install/manage.test.mjs — 14 tests covering the migration logic plus the end-to-end install flow: fresh install, re-install from an older state version (migration + *.pre-migrate.json backup), and codex/claude target independence. New npm run test-install, wired into npm test.
• scripts/check-skill-links.mjs — fails when a SKILL.md references a `do-it-*` skill or command that does not exist, or a skill directory is missing from manifest.json. New npm run check:skill-links, wired into npm test.
• do-it-router § Integrity — a canonical principle: a failure is a clue to trace to a root cause, not a symptom to make disappear. Names the cover-up patterns (swallowed exception, weakened/deleted assertion, skipped/deleted test, commented-out failing code, hiding fallback, edited evidence). do-it-debugging, do-it-fix-loop, do-it-verification-gate, and the subagent dispatch contract reference it; the reviewer agent treats a cover-up as a Blocking finding.
• do-it-visual-planning/templates/plan-card.html and review-report.html — content-fragment templates that reuse the existing browser-companion frame and CSS; section order matches task-card-template.md and review-loop severity grouping.
• CI test job now also runs on macos-latest, guarding the bash hooks against BSD/GNU tool divergence (find, printf, date).

Changed

• hooks/verification-gate.sh block reasons rewritten as short single-instruction sentences. The verbose "for example a line whose content is the literal phrase…" meta-explanation that caused the model to echo the reason back verbatim is removed.
• hooks/verification-gate.sh scopes edit / evidence / review-loop detection to the current turn — the transcript lines after the last user message. A verification command or review-loop trace from an earlier turn can no longer satisfy a later unverified turn. The tail window is raised 80 → 400 lines (turn scoping removes the staleness cost of a larger window).
• applyMigrationAction now throws on an unknown migration action instead of logging and skipping it — a malformed manifest fails loud rather than leaving a half-migrated state.
• manifest.json migrations gain 0.7.x and 0.8.x entries.
• do-it install and doctor now report the install-state version of every manifest target, surfacing codex/claude cross-host version drift that a single-target run otherwise hides.
• do-it-subagent-orchestration Required Prompt Contract gains an integrity stance field, propagating the integrity principle into every dispatched subagent.
• do-it-router and do-it-planning now point at do-it-handbook and task-card-template.md so the handbook skeleton is discovered and reused instead of ignored.

Fixed

• do_it_emit_block / do_it_emit_context emit valid JSON via a printf fallback when jq is unavailable. Previously a jq-less host silently dropped the Stop-hook block decision and context reminders entirely.
• The 0.8.0 known limitation where a stale review-loop trace in a long session could let a later review-needing turn pass silently — resolved by the current-turn scoping above.
• Stale session directories under the runtime sessions base are pruned after DO_IT_SESSION_TTL_DAYS (default 7) of inactivity (do_it_prune_stale_sessions, run once per session from router.sh), instead of accumulating without bound.

0.8.0

Added

• New PostToolUse hook hooks/anti-patterns-lint.sh — advisory scan after every Edit|Write|MultiEdit for three coarse code-quality anti-patterns: large bash case lists (≥10 consecutive *"..."* branches → suggests data-driven externalisation), newly-exported JS/TS symbols with no other-file consumer (incl. export default function/class), and ≥5-line code blocks duplicated from a same-directory neighbour. Never blocks; emits one system-reminder per file. Implementation uses portable grep+sed (no gawk-specific match(... , arr) three-arg form) so it runs on macOS's bundled BSD awk just like the 0.7.2-era hooks.
• New task pointer protocol: .do-it/runtime/pointer is a single-line file holding the active task slug. do-it-planning and do-it-brainstorm write it when creating a durable artifact; do-it-branch-closeout clears it; do-it-router reads it before routing. Defined once in do-it-router § Task Pointer; other skills short-reference.
• Tests: tests/hooks/anti-patterns-lint.test.sh (12 cases incl. export default coverage) and 4 new cases in tests/hooks/verification-gate.test.sh lock in the new behaviors. scripts/test-hooks.sh gains a dim-aware grill suppression case + counter-case so the new hooks/grill-prompt.sh consumer of dim_touches_code is exercised. npm run test-hooks now exercises the full tests/hooks/*.test.sh suite (router, comments-lint, anti-patterns-lint, verification-gate, common).

Changed

• DIM_ dimensions are now actively consumed* (the 0.7.0 "维度正交化" design finally lands). Each of the five session-state dimensions has at least one explicit consumer:
  • dim_touches_code → hooks/grill-prompt.sh suppresses Standard-tier implicit triggers when the prompt has no code object (discussion turn).
  • dim_crosses_packages → do-it-architecture-scan skill: mandatory trigger when set.
  • dim_breaks_interface → do-it-interface-drill skill (mandatory trigger) and hooks/verification-gate.sh (requires the inline-review marker to name interface / contract / schema / api).
  • dim_needs_tdd → do-it-tdd skill: mandatory trigger when set.
  • dim_needs_review_loop → do-it-review-loop skill (mandatory trigger) and hooks/verification-gate.sh (requires review-loop / review-quick / review-deep / review-adversarial mention in the recent transcript before a "done" claim passes).
  • The canonical one-liner for reading a dimension (and the "missing state degrades to tier-only" fallback) lives in do-it-router § Reading dimensions. docs/routing-matrix.md updated to match: MAY → SHOULD.
• do-it-fix-loop switches default posture from "see-one-fix-one" to "collect all findings → root-cause → batch or pointwise". The Standard tier sequence now requires a written Batch vs Pointwise Decision covering every finding before editing. The Red Flag entry that previously forbade multiple unrelated fixes in one commit is removed; the new red flag is going one-by-one without a written root-cause decision. do-it-review-loop correspondingly requires reviewers to emit findings as a complete batch, not stream them.
• do-it-comments-discipline SKILL.md trimmed from 373 lines to 119 (~68% reduction), keeping the 5 allowed categories (one example + one near-miss each), the 6 forbidden families (one bad line + one fix per family), the review checklist, the hook keyword reference, escape hatch, and output shape. The comments-lint hook strength is unchanged (still advisory) — the bet is that a smaller SKILL gets read by the agent at all.
• 23 subagent TOML files lose their duplicated Common protocol: block (~4-5 lines per agent, ~120 lines total). Replaced with a single line pointing at do-it-subagent-orchestration § Required Prompt Contract, which already defines the canonical dispatch contract. The shared boilerplate now lives in exactly one place.

Fixed

• hooks/verification-gate.sh no longer falls back to a raw tail -n 5 of the JSONL when jq does parse but the last assistant frame is tool-only with no text content. The fallback only fires when jq is missing; with jq present, an empty result is treated as "agent said nothing this turn" and the gate stays silent. Prevents a false-positive block where the new dim_breaks_interface attestation check would see leaked text from earlier turns and refuse to find the interface keyword on its own line.
• Documented Reading dimensions in do-it-router § now distinguishes the hook layer (call do_it_session_state_get from hooks/lib/common.sh, which already honours the 5-level env-var path search) from the agent layer (judge mandatory triggers from prompt content; do not query state at runtime). The earlier shipped one-liner pointed at a path that is incorrect under the plugin-install env layout.
• Each mandatory-trigger sentence inside do-it-tdd, do-it-interface-drill, do-it-architecture-scan, and do-it-review-loop now carries an explicit Light-tier escape clause and points at do-it-router § Mandatory-trigger escape clauses. dim_needs_tdd is no longer treated as a forced ceremony on docs/config edits.
• .do-it/runtime/pointer protocol is reframed as a best-effort hint: read consumers MUST verify the referenced artifact exists, and the document explicitly addresses branch-switch, concurrent-write, and <closed>-sentinel edge cases.
• do-it-comments-discipline gains a Verification (pass criteria) section and a Trigger Event section that the 373→119 slim accidentally dropped; the Anti-Pattern Hook Keywords list is reconciled with hooks/comments-lint.sh (adds 曾经, fix:, bugfix, hotfix, patched, FIXME, XXX; splits history and fix-narrative into separate families to match the hook's _record_family calls).
• do-it-review-loop and do-it-fix-loop now cross-link explicitly: review-loop step 8 names the batch contract fix-loop operates on; fix-loop's Batch vs Pointwise Decision references review-loop step 8 as the source of the complete finding list.
• README.md, README.zh-CN.md, and docs/release.md carry the new 0.8.0 tarball filename and an Upgrading to 0.8.0 section. The READMEs also gain a paragraph describing the new advisory anti-patterns-lint hook.

Known Limitations

• dim_needs_review_loop enforcement in hooks/verification-gate.sh reads TAIL_BUF (last 80 JSONL lines) and accepts any historical review-loop / review-quick / review-deep / review-adversarial mention. In a long session with multiple distinct review-needing turns, a later turn can inherit an older trace and silently pass. Tightening this requires a last_review_seen_at session-state timestamp; tracked as follow-up, not blocking 0.8.0.
• dim_breaks_interface attestation in hooks/verification-gate.sh accepts any inline-review marker that mentions interface / contract / schema / api — best-effort signal, not a semantic check. A marker like inline-review: schema validation looks fine satisfies the gate even on a non-schema change. The signal is intentionally generous; the reviewer / fix-loop is the real gate.

0.7.2

Fixed

• Claude Code plugin hooks now run on macOS's bundled Bash 3.2 by removing Bash 4.3 nameref usage (local -n) from keyword loading and prompt matching.
• Hook scripts avoid a few GNU-only assumptions on macOS (sha1sum, dirname --, stat --, cat --, and template-less mktemp).

Changed

• Tightened selected do-it skills with shorter trigger metadata and explicit stop, red-flag, and verification guidance for closeout, fix-loop, delegation, worktree isolation, grill logging, and skill authoring.

0.7.1

Highlights

• agent-skills method absorption. Rewrote selected addyosmani/agent-skills methods into do-it-native workflow policy without vendoring upstream text or adding installed upstream skill names.
• Skill anatomy hardening. do-it-skill-authoring now defines the minimum installed-skill structure: trigger, tier/process, stop conditions, common rationalizations, red flags, and verification evidence.
• Core workflow density. Planning, slicing, TDD, debugging, interface drill, review-loop, architecture scan, verification-gate, and context now carry stronger anti-skip, boundary, dependency, and proof-quality rules.

Changed

• docs/upstream-map.md records the agent-skills source idea to do-it destination mapping.
• docs/maintenance.md documents external workflow absorption, skill anatomy, and generated-artifact ownership rules.
• README and README.zh-CN acknowledge agent-skills as method inspiration and state the rewrite-not-vendoring boundary.

0.7.0

Highlights

• Codex agent schema repair. Removed the unsupported output_budget top-level field from Codex-installed agents/*.toml; subagent response budgets now live in do-it-subagent-orchestration.
• Agent bundle validation. Added scripts/validate-agent-bundle.mjs and wired it into npm test, build:claude-agents, build:codex-plugin, and prepack so source agents, manifest inventory, Claude output, and the Codex plugin bundle cannot drift silently.
• 0.7.0 release metadata. Package, manifest, plugin metadata, release docs, and README upgrade notes now describe the 23-skill / 23-agent surface.

Added

• Codex plugin marketplace distribution under .agents/plugins/marketplace.json and plugins/do-it/, generated by scripts/build-codex-plugin.mjs from manifest.json.
• Codex global hook install assets: default Codex setup now manages hooks/ and root hooks.json so UserPromptSubmit, PreToolUse, PostToolUse, and Stop can invoke do-it hooks from ${CODEX_HOME:-$HOME/.codex}.

Changed

• Skill frontmatter descriptions now use trigger-first Use when... wording while preserving the existing Problem/Fix body content.
• brainstorm is no longer an escape keyword, so explicit brainstorm prompts continue through router, grill, and verification gates.
• Release and maintenance docs now distinguish Codex global setup, Codex plugin marketplace discovery, and Claude Code plugin hooks.

0.6.1

Highlights

• Codex-compatible agent TOML. Removed the Claude-only claude_model field from Codex-installed agents/*.toml. scripts/build-claude-agents.mjs now owns the Claude model map and still emits model: sonnet for the brainstorm lenses and code-mapper.
• Brainstorm two-core redesign. do-it-brainstorm now starts from product-strategist and architecture-strategist, then dynamically adds task-fit supplemental lenses. Product strategy clarifies product boundary, core goal, requirement shape, and option tradeoffs. Architecture strategy separates core foundation from extension modules and names what should close in the current stage.
• Brainstorm discovers; grill converges. Brainstorm output now centers on Requirement Shape, Product Boundary, Core Goal, Options, Architecture Foundation, Extension Modules, and Must Resolve In Grill. Grill owns convergence; brainstorm does not collapse options into a final answer.

Added

• agents/product-strategist.toml — required product core brainstorm lens.
• agents/architecture-strategist.toml — required architecture core brainstorm lens.

Changed

• skills/do-it/do-it-brainstorm/SKILL.md — replaces the fixed four-persona flow with product + architecture cores, dynamic supplements, discussion-first mode, and option tradeoff output.
• commands/do-it-brainstorm.md — documents default dual cores, explicit lens selection, and discussion-first usage.
• skills/do-it/do-it-grill/SKILL.md — consumes Must Resolve In Grill instead of the legacy per-lens question list.
• README, README.zh-CN, release, and maintenance docs updated for the 0.6.1 compatibility and brainstorm contract.

Migration

• Re-run do-it setup or do-it install to refresh live Codex agent TOML. After upgrade, rg -n "claude_model" ~/.codex/agents should not match do-it-managed Codex agent files.
• No project artifact migration is required. Existing 0.6.0 brainstorm files remain readable by humans, but new artifacts use the 0.6.1 section names.

0.6.0

Highlights

• Brainstorm before grill. Initial do-it-brainstorm support wrote one artifact per task at .do-it/brainstorm/<task>.md. The lens model and section contract were replaced by the 0.6.1 product + architecture core design above.
• Grill convergence mode. do-it-grill reads the brainstorm artifact when one is open, lifts execution-blocking items into candidate premises, and flips brainstorm status: open → converged once each decision is resolved in the grill log. Light tier still uses single-thread grill; Standard / Heavy must consume the brainstorm.
• Project handbook bootstrap. New do-it-handbook skill scaffolds .do-it/handbook/ with twelve generalized templates (invariants, architecture, code-map, glossary, backlog, runtime-status, maintenance, task-card-template, plus three workflow files). Templates are skeletons with placeholders; the bootstrap is additive and never overwrites.
• Persistent code map. code-mapper now writes the "Current Implementation Locations" section of .do-it/handbook/code-map.md when the handbook exists, so future sessions read instead of re-derive. A new code-map-refresh hook marks that section stale on barrel / migration / route / workspace-manifest edits.

Added

• skills/do-it/do-it-brainstorm/SKILL.md — initial divergent brainstorm pass before grill convergence.
• skills/do-it/do-it-handbook/SKILL.md plus twelve templates under templates/ and templates/workflow/.
• agents/ceo-reviewer.toml, ux-designer.toml, end-user-advocate.toml, ops-sre.toml — supplemental read-only brainstorm lenses. Their Claude model handling moved out of Codex TOML in 0.6.1.
• commands/do-it-brainstorm.md, commands/do-it-handbook.md.
• hooks/code-map-refresh.sh — PostToolUse marker that prepends <!-- stale: true; reason: ... --> to .do-it/handbook/code-map.md on structural-file edits. Idempotent (replaces, does not stack).

Changed

• scripts/build-claude-agents.mjs — builds Claude agent Markdown from Codex TOML. Claude-only model selection moved into the generator in 0.6.1.
• agents/code-mapper.toml — adds a Handbook Write Target / Claude Code Adapter section in developer_instructions. Output target shifts from inline-only to .do-it/handbook/code-map.md "Current Implementation Locations" when the handbook exists.
• skills/do-it/do-it-grill/SKILL.md — adds "Convergence after brainstorm" section and Light / Standard / Heavy consumption rules.
• hooks/grill-prompt.sh — when .do-it/brainstorm/*.md has any file with status: open and tier is Standard / Heavy, appends a convergence-mode pointer to the grill reminder.
• hooks/hooks.json — registers PostToolUse for code-map-refresh.sh.
• manifest.json — registers do-it-brainstorm and do-it-handbook skills, plus the initial brainstorm lens agents.

Migration

• No breaking changes for 0.5.x users. Existing .do-it/grill/, .do-it/plans/, and .do-it/CONTEXT.md paths are unchanged.
• The new .do-it/brainstorm/ directory is additive; grill ignores it on Light tier and works as before when no brainstorm artifact exists.
• The new code-map refresh hook only acts when .do-it/handbook/code-map.md already exists; projects that have not run /do-it-handbook are unaffected.

0.5.1

Highlights

• Flow reduction. Standard prompts no longer default to grill -> planning -> review-loop. The router now tells agents to use an inline modification map, add grill only when the decision actually needs it, and select review depth by risk.
• Non-sticky question state. Router now records last_prompt_kind=question|work, so a question turn can bypass grill/gate for that turn without suppressing the next implementation turn.
• Decision-triggered grill. grill-prompt no longer fires just because an intent verb appears. Standard grill triggers only on uncertainty, explicit grill requests, or long input that also has a plan/spec hint. Heavy still auto-grills.
• Lighter plan gate. grill-pretool only requires .do-it/plans/* for Heavy work or explicit durable-plan requests. Standard source edits may use an inline modification map. Existing plan partial edits are no longer blocked just because the edited fragment lacks ## Grill.
• Risk-budgeted review. do-it-review-loop now spells out local Light / docs-only review, at most one focused Standard reviewer, and two default Heavy release/workflow lenses: skill quality and install/release readiness.

Added

• scripts/test-hooks.sh — regression coverage for question-state recovery, reduced grill triggers, Heavy two-signal behavior, long-input length + hint, Standard vs Heavy source plan gates, and existing-plan partial edits.
• npm run lint-hooks and npm run test-hooks; npm test now runs both.

Changed

• hooks/router.sh — separates prompt kind from grill state, records durable plan requirement, and updates Standard/Heavy reminders.
• hooks/grill-prompt.sh — removes intent-only trigger, changes long-input trigger from length OR hint to length AND hint, and changes reminders to fact-first / one-question guidance.
• hooks/grill-pretool.sh — scopes hard plan gates to Heavy or explicit durable-plan work.
• hooks/verification-gate.sh — checks last_prompt_kind=question instead of relying on grilled=skip-question.
• do-it-grill, do-it-grill-log, do-it-planning, and do-it-verification-gate — align the contract around fact verification, one user decision at a time, and kind: fact|decision grill-log items.
• README, routing, release, and migration docs updated for 0.5.1 behavior and tarball examples.

0.5.0

Highlights

• Sharper triggering. hooks/router.sh and hooks/grill-prompt.sh no longer fire on single CJK characters (做, 改, 加, 修, 审, …) and use ASCII word-boundary matching, so fix no longer matches prefix and add no longer matches address. Heavy tier requires ≥2 heavy signals to upgrade.
• Same-session de-dup. Grill emits at most once per session unless the user explicitly asks again (重新 grill / re-grill / 再 pressure-test / 重新审视). Re-implementing in the same conversation no longer eats tokens on a repeat 5-step template.
• Question / discussion mode. Prompts ending in ? / ？ / 吗 / 呢 or matching the question hints get auto-classified Light, suppress grill, and bypass the verification gate. "你觉得 X 怎么样？" is no longer treated as an implementation request.
• Pointer-mode grill output. Standard-tier grill emits ~60-token pointer reminders; full 5-step template is preserved for Heavy tier.
• DO_IT_DEBUG=1 is now first-class. Each hook decision (escape, skip, question, tier, trigger, evidence detection) emits one structured stderr line. Default is silent.
• Session state JSON. Hook invocation counters and tier history are persisted under ${CLAUDE_PLUGIN_DATA}/sessions/<id>/state.json. do-it doctor --session=<id> pretty-prints it.
• Verification-gate is more honest. Reads only the last assistant message for completion language; passes through turns with no Edit/Write/MultiEdit calls; respects question / discussion turns. Evidence pattern expanded to cover pytest, mypy, tsc, eslint, ruff, biome, cargo, go.
• Auto-migrate. do-it install detects 0.4.x install state and silently upgrades it; pre-migrate.json backup is left in place. Use --no-migrate to refuse and exit code 2.

Added

• hooks/data/*.tsv — keyword tables are now data-driven (intent verbs, uncertainty words, heavy/light signals, escape words, long-input hints, question hints). hooks/lib/keywords.sh is a thin loader.
• hooks/data/SCHEMA.md — file format and maintenance checklist for the tsv tables.
• hooks/lib/debug.sh — do_it_debug helper for structured stderr trace when DO_IT_DEBUG=1.
• do_it_prompt_has_word, do_it_prompt_is_question, do_it_session_state_get/set/inc, do_it_session_summary in hooks/lib/common.sh.
• New skills: do-it-context (with CONTEXT-FORMAT.md) and do-it-grill-log (per-task grill artifact under .do-it/grill/<task>.md).
• scripts/lint-hooks.sh — shellcheck on hook entry points; npm run lint.
• .github/workflows/lint.yml — CI shellcheck on push / PR.
• install/manage.mjs — --session=<id> flag for doctor; --no-migrate flag for install; needsMigration / runMigration with manifest-declared migration actions; pre-migrate state backup.
• manifest.json migrations field — declarative migration actions (remove-state-entry, rename-state-key).

Changed

• hooks/router.sh — question short-circuit, ≥2-heavy-signal upgrade rule, compact ≤80-token system-reminder. Full Heavy / Standard / Light recommendations have moved into the corresponding skill SKILL.md files.
• hooks/grill-prompt.sh — same-session de-dup; Heavy tier emits full template, others emit pointer; explicit re-grill phrase detection.
• hooks/verification-gate.sh — last-assistant-only completion scan; no-edit pass-through; question pass-through; expanded evidence patterns.
• hooks/grill-pretool.sh — debug instrumentation only; behavior unchanged.
• skills/do-it/do-it-grill/SKILL.md rewritten to "ask one premise at a time, anchor terms, falsify before debating, sediment to CONTEXT.md / grill log".
• skills/do-it/do-it-planning/SKILL.md — Planning Sequence step 2 now requires reading .do-it/grill/<task>.md and resolving any pending premise before finalizing the plan.
• skills/do-it/do-it-verification-gate/SKILL.md — new "Inputs" section requiring .do-it/grill/<task>.md review at closeout.

Removed

• Single-character CJK intent verbs (做, 改, 加, 写, 修, 审, 搭) from the default intent-verbs table.
• ASCII whitespace hacks (add, add, test, should, could, might, doc) — replaced by word-boundary matching, which handles address / prefix / released correctly without manual padding.

Migration

do-it install detects 0.4.x install state and silently migrates. Behavior is additive only; the auto-migrate is mostly bookkeeping (state version bump + new skills do-it-context, do-it-grill-log).

0.4.0

Highlights

• Claude Code target. do-it now ships as a Claude Code plugin alongside the Codex install. Use /plugin marketplace add tdwhere123/codex-workflow then /plugin install do-it, or do-it install --target=claude from the CLI.
• Hook-driven workflow. On Claude, the three core skills (router, grill, verification-gate) auto-trigger via UserPromptSubmit, PreToolUse, and Stop hooks. No slash commands to memorize. Escape hatch: include yolo, 直接做, skip grill, or /do-it-skip in the prompt.
• Problem/Fix skill descriptions. All 17 SKILL.md descriptions rewritten in the Problem: ...; Fix: ... shape so Claude's implicit-summon path stays reliable when hooks are bypassed.

Added

• .claude-plugin/plugin.json and marketplace.json — Claude plugin metadata with a self-hosted marketplace pointing at the repo root.
• hooks/ — four hook scripts (router.sh, grill-prompt.sh, grill-pretool.sh, verification-gate.sh) plus shared library (lib/common.sh, lib/keywords.sh).
• commands/do-it-skip.md — the only slash command; explicit escape hatch.
• scripts/build-claude-agents.mjs — converts agents/*.toml to dist/claude/agents/*.md. Runs automatically on do-it install --target=claude and via prepack.
• manifest.json — new targets block describing each host's install root, agent file extensions, extras, and pre-install scripts.
• manage.mjs — --target=<name> and --with-optional flags.
• package.json — new install:claude / doctor:claude / build:claude-agents scripts; prepack runs the agent build.

Removed

• do-it-delivery-loop skill — folded into the Heavy path of do-it-router plus do-it-planning. Recorded in manifest.deprecatedTargets; existing 0.3.x installs have it cleaned up on upgrade via legacyHashes.

Marked optional

• do-it-visual-planning — was a default skill; now requires --with-optional to install. Existing installs are not auto-removed but drop out of doctor's managed set; rerun do-it install --with-optional to keep it managed.

Changed

• All remaining 17 SKILL.md descriptions rewritten in Problem: ...; Fix: ... shape.
• manage.mjs rewritten around an installRoot abstraction; codexHome is now installRoot, computed from the active target's rootEnv / rootDefault. CODEX_HOME continues to work for codex; new CLAUDE_PLUGIN_ROOT_OVERRIDE for claude target.
• assertManagedTargetShape now accepts extras (top-level dirs/files added by target config) and the agent target extension is target-driven (.toml for codex, .md for claude).

Compatibility

• Codex install is byte-equal with 0.3.x except for do-it-delivery-loop (deprecated, removed on upgrade) and do-it-visual-planning (now optional; unchanged on disk but no longer in the managed set unless --with-optional).
• State files: codex still uses .do-it-install-state.json; claude uses .do-it-install-state-claude.json to avoid collision when both targets are installed.
• Existing 0.3.x codex users: do-it-delivery-loop will be removed on upgrade (legacyHash matches). do-it-visual-planning will not be auto-removed but will fall out of the managed set; reinstall with --with-optional or remove manually.

Upgrade

npm install -g @tdwhere/do-it@0.4.0
do-it install                       # codex (default)
do-it install --target=claude       # claude (or use marketplace)

To keep do-it-visual-planning as a managed skill on codex:

do-it install --with-optional