@moretea-labs/forge 1.5.0-rc.1 → 1.5.0

package/docs/CHANGELOG.md

@@ -1,978 +0,0 @@
-# Changelog
-
-## 1.4.0 — Unified Forge Runtime
-
-- Made Direct Edit a multi-revision transaction with savepoints and partial rollback.
-- Removed ordinary local risk approval gates and the new-work approval queue.
-- Made Task Agent assignment optional and selected executors at Run time.
-- Reorganized the Local Controller into Overview, Work, Activity, and Settings.
-- Added V7 state migration and V8 compatibility tests.
-
-All notable changes to this skill are documented here.
-
-## [Unreleased]
-
-### Fixed
-
-- Isolated Controller Runtime Source Identity from execution repositories. MCP `rh_status`, CLI controller status, and Local Bridge access state now compare startup runtime source against the package/source authority instead of the selected business repository, eliminating false `RUNTIME_SOURCE_SNAPSHOT_STALE` when switching multi-repo sessions. Missing runtime snapshots remain fail-closed with structured reasons.
-
-### Performance
-
-- Removed the independent Controller Daemon auto-start API; compatibility status reads now project the single Canonical Runtime owner/readiness without spawning a process.
-- Cached full-table process scans (~2.5s), git snapshots (~3s), and runtime source identity collection (~2.5s) so status/readiness storms do not re-run the same host/git work every request.
-
-## [1.3.0] - 2026-06-22
-
-### Added
-
-- Added a shared Task-local execution policy for launch, verification, approval, automatic continuation, and completion.
-- Added read-only and destructive risk classes, ephemeral Quick Agent Issues, startup-deadlock detection, reported command evidence, compact snapshots, and explicit Connector surface drift metadata.
-- Added regression coverage for Task-local readiness, no-check launch, dynamic risk gates, dependency migration, ephemeral cleanup, startup timeout, concurrent scope conflicts, repository glob semantics, and dynamic progress.
-
-### Changed
-
-- Issue readiness is now an aggregate view; unrelated blocked Tasks, multiple active Issues, and current focus no longer block independent execution.
-- Named checks are completion evidence instead of a universal launch prerequisite. Validation and human approval are dynamically reduced for read-only and low-risk work and strengthened only for high-risk/destructive work.
-- Successful Runs continue automatically through applicable checks and completion, while real check failures remain authoritative.
-- Quick Agent sessions are ephemeral by default and no longer pollute the durable Issue board.
-- Job/Run terminal status, `finishedAt`, dependency state, repository search/read policy, progress calculation, log bounds, and `project_snapshot` size are reconciled.
-- Versioned the Controller surface as `controller-execution-first-v7`, schema `9`, surface version `7`, and package line as `1.3.0`.
-
-## [1.2.0] - 2026-06-21
-
-### Added
-
-- Added request-mode assessment that routes work to `direct_edit`, `quick_agent`, or `issue_task` before creating durable Issues.
-- Added first-class direct-edit session listing, persisted unified patches, real named-check verification, reviewer evidence, and finalization/rollback lifecycle.
-- Added a primary File Changes view and local APIs for reading edit details and patches, running verification, finalizing, and rolling back.
-- Added CLI and MCP surfaces for work assessment, edit-session history, patch inspection, verification, finalization, and rollback.
-- Added full-file SHA-256 to `read_repository_file`, including ranged reads, so bounded replacements can enforce stale-write preconditions.
-
-### Changed
-
-- Small known code, configuration, and documentation changes now default to direct edits instead of Issue creation.
-- `create_issue` is documented and instructed as the complex-work path for investigation, dependencies, broad scope, parallelism, long checks, or high risk.
-- Direct edits now require persisted patch evidence and verification before finalization; failed verification can be rerun or rolled back without losing history.
-- Unified worklog and Controller snapshots now include edit-session events and summaries alongside Issue/Task/Run evidence.
-- Versioned the Controller fingerprint as `controller-direct-change-v6`, the default Connector identity as `forge-controller-v6`, the Controller schema as `8`, and the package/template line as `1.2.0`.
-
-## [1.1.0] - 2026-06-21
-
-### Added
-
-- Added a single current execution Issue and project-state policy (`open`, `focus_only`, or `paused`) so project-wide execution no longer mixes unrelated active Issues.
-- Added Issue-creation guardrails for duplicate titles, active-focus proliferation, explicit policy overrides, and automatic first-Issue focus.
-- Added project governance diagnostics and safe reconciliation for dead dependencies, stale superseded dependencies, retryable failed attempts, review/acceptance backlog, closeout drift, and archiving drift.
-- Added evidence-gate progress based on implementation, integration, named checks, acceptance criteria, and explicit human acceptance.
-- Added direct Controller actions to launch ready work, retry failed attempts, run declared named checks with explicit criterion confirmation, accept verified Tasks, request changes, cancel Tasks, repair dependencies, and archive/restore terminal Issues.
-- Added current-work versus archived-history separation while retaining the full Task, Run, Verification, GitHub, and worklog evidence chain.
-
-### Changed
-
-- Failed, unknown, or cancelled Runs now remain attempt history while the Task returns to a retryable state unless a real dependency or human decision blocks it.
-- Project-wide ready-task dispatch now requires an explicit or current Issue and no longer traverses every active Issue by default.
-- Readiness scoring can no longer report 100 when no Task is dispatchable.
-- `projectBoard.readyTasks` is scoped to the current execution Issue; archived Task counts are separated.
-- Versioned the Controller fingerprint as `controller-execution-closure-v5`, the default Connector identity as `forge-controller-v5`, the Controller schema as `7`, and the package/template line as `1.1.0`.
-
-## [1.0.0] - 2026-06-21
-
-### Added
-
-- Added a project progress ledger that aggregates Issue, Task, Run, approval, verification, and GitHub synchronization state into one effective progress model.
-- Added durable controller worklog events under `.ai/harness/controller/worklog.jsonl`, unified Task timelines, evidence inspection, and Markdown/JSON export under `tasks/reports/`.
-- Added a V4 localhost control center with progress, Task detail, Run monitoring, worklog, approval, checks, and optional GitHub plugin views.
-- Added SSE dashboard refresh with polling fallback.
-- Added optional GitHub Issue/Project plugin configuration. Local controller files remain authoritative and synchronization is explicit.
-- Added MCP and CLI surfaces for project progress, Task detail, timeline inspection, worklog export, and GitHub plugin configuration.
-
-### Changed
-
-- Versioned the Controller fingerprint as `controller-progress-ledger-v4`, the default Connector identity as `forge-controller-v4`, the controller schema as `6`, and the package line as `1.0.0`.
-- Task progress now uses effective status derived from durable Task state plus the latest associated Run, preventing stale Task labels from hiding active or failed execution.
-- GitHub integration is disabled by default and exposed as a plugin so local execution and evidence remain usable without GitHub.
-
-
-## [0.9.0] - 2026-06-20
-
-### Added
-
-- Added a redesigned localhost task workstation with overview, Task board, live Run monitoring, approvals, checks, activity timeline, console, diff, heartbeat, timing, and integration state.
-- Added structured Codex/Claude activity parsing so raw output is translated into inspecting, editing, testing, finalizing, waiting, completed, and failed phases.
-- Added automatic execution placement: the first local Task runs directly in the current workspace and concurrent Tasks receive isolated worktrees.
-- Added automatic post-success integration and cleanup for isolated worktree Runs, while preserving archived diff evidence and retaining conflicted worktrees for manual review.
-
-### Changed
-
-- Made omitted `isolate` mean automatic placement instead of forced worktree isolation. Explicit `false` now refuses to start when another local Run is active.
-- Kept Verification Gate mandatory after direct changes or automatic integration; worktree usage is now an execution-safety detail rather than an extra normal user step.
-- Versioned the Controller fingerprint as `controller-live-workspace-v3`, the default Connector identity as `forge-controller-v3`, and the package line as `0.9.0`.
-
-## [0.8.0] - 2026-06-20
-
-### Added
-
-- Added GitHub Issue/Project publication for controller Issues, including optional Task sub-issues and explicit remote links stored in local durable state.
-- Added GitHub Copilot coding-agent cloud Sessions as a visible Task execution provider alongside local Codex and Claude Runs.
-- Added Issue Launcher readiness previews, dynamic Task append/split/supersede/dependency operations, structured Run events/logs, and a mandatory Verification Gate.
-- Added `forge controller board|runs|watch|github-status` for local visibility into task execution and GitHub integration readiness.
-- Added live local Codex/Claude log streaming and `controller_capabilities` tool-surface diagnostics for stale ChatGPT Connector detection.
-- Added a localhost-only visual controller for Issue/Task boards, Local Job approvals, live Run logs/events, named checks, cancellation, and retry.
-- Added explicit Run timeout metadata, deadline progress, 30-second heartbeat events, process-group termination, and persisted retry history.
-
-### Changed
-
-- Made `controller` the documented default ChatGPT MCP workflow and retained planner/orchestrator as compatibility profiles.
-- Raised the default local Agent timeout from 120 seconds to 60 minutes, added a 12-hour configurable ceiling, and reject invalid values instead of silently downgrading them.
-- Versioned the default ChatGPT Connector identity as `forge-controller-v2` and added runtime fingerprints to detect stale tool snapshots or wrong processes on the MCP port.
-- Task acceptance now requires recorded named-check and criterion-level evidence; isolated local Runs must also be integrated first.
-
-## [0.7.2] - 2026-06-19
-
-### Added
-
-- Added ChatGPT native browser product-session binding for user-selected Chrome
-  profiles, with printed bind-page URLs, Chrome profile-directory support,
-  doctor session validation, and fail-closed native consults when no ChatGPT
-  profile is bound.
-- Added Oracle browser-provider heartbeat support. GPT Pro / ChatGPT browser
-  consults now pass `--heartbeat`, defaulting to 59 seconds, so long-running GPT
-  analysis stays observable through Cloudflare and local process boundaries.
-
-### Fixed
-
-- Corrected gbrain remediation guidance to install the official GitHub GBrain
-  CLI instead of the unrelated npm registry `gbrain` package.
-- Fail closed when native ChatGPT browser setup targets the default Chrome data
-  directory, because Chrome 136+ blocks remote debugging there; doctor now
-  reports `blocked_default_profile` instead of opening Chrome and timing out.
-- Removed the static `file://` ChatGPT bind page path. Product-session binding
-  now goes through `forge chatgpt browser-bind`, whose **Bind ChatGPT**
-  action calls a local authorization endpoint instead of linking directly to
-  ChatGPT.
-- Added the ChatGPT bridge provider for existing signed-in Chrome profiles. It
-  generates a product-scoped unpacked extension for ChatGPT domains plus
-  localhost, validates composer readiness through extension heartbeat, and
-  supports `browser-consult --provider bridge` without copying cookies or
-  browser storage.
-- Hardened ChatGPT bridge authorization diagnostics for unpacked extensions
-  recorded in Chrome `Secure Preferences`, and isolated bridge-provider tests
-  from user-installed extensions polling the default local port.
-- Excluded `gbrain` from setup readiness dependencies: missing or stale gbrain
-  stays visible as advisory tooling state without creating Agent actions.
-- Kept ChatGPT MCP stable endpoints out of tracked generated guides, storing
-  real Connector URLs only in ignored local config while public docs remain
-  placeholder-only.
-- Preserved existing ChatGPT MCP endpoint, auth, transport, profile, and dev-mode
-  settings when rerunning `forge mcp setup chatgpt --server-name`.
-- Made `forge mcp doctor --profile chatgpt` report whether the ChatGPT
-  server name is actually configured instead of masking missing local state with
-  the default `forge` name.
-- Made Oracle doctor repair actions source-aware, so explicit `--oracle-bin`,
-  `REPO_HARNESS_ORACLE_BIN`, and repo-local Oracle installs get actionable repair
-  guidance instead of a generic global install suggestion.
-- Limited SessionStart tooling update advisories to one render per cached
-  report, so weekly Waza/CodeGraph update checks do not re-inject stale
-  update instructions on every new agent turn.
-
-### Release Notes
-
-- Prepared the `forge@0.7.2` patch line for npm publish, registry
-  readback, clean-room install smoke, Git tag, and GitHub release creation.
-
-## [0.7.1] - 2026-06-18
-
-### Added
-
-- Added `forge-gptpro-setup` / `forge:gptpro_setup` to guide
-  local `gptpro_browser` browser/session setup and `gptpro_mcp` ChatGPT
-  Connector MCP setup without treating ChatGPT Pro as API quota.
-- Added `forge-gptpro` / `forge:gptpro` as the local consult
-  skill for GPT Pro browser-session assistance, using `gptpro
-consult/read/continue/open` language over the existing ChatGPT Web browser
-  session engine.
-
-### Fixed
-
-- Preserved `.ai/harness/checks/latest.json` for authoritative
-  `forge-run-trace.v1` evidence by writing PostBash advisory metadata to
-  `.ai/harness/checks/post-bash-latest.json`.
-
-### Release Notes
-
-- Prepared the `forge@0.7.1` patch line for npm publish, registry
-  readback, clean-room `npx` smoke, Git tag, and GitHub release creation.
-
-## [0.7.0] - 2026-06-18
-
-### Added
-
-- Added `forge chatgpt browser-*`, a policy-checked ChatGPT Web browser
-  session engine with dry-run prompt assembly, repo-local session records,
-  linked follow-ups, output copying, session read/list/open, and dry-run-first
-  cleanup.
-- Added optional MCP ChatGPT browser tools behind
-  `forge mcp serve --enable-chatgpt-browser`, keeping browser consults
-  out of the default MCP surface.
-- Added an Oracle provider wrapper for `oracle --engine browser` and a native
-  installed-Google-Chrome CDP provider spike for logged-in ChatGPT Web sessions.
-- Added a bundled `forge-chatgpt-browser` skill and user guide at
-  `docs/forge-chatgpt-browser-engine.md`.
-- Added a hosted GitHub CI gate for pull requests and pushes to `main` /
-  `codex/**`.
-- Added the Harness Engineering Optimization sprint closeout surface, including
-  task profiles, Human Review Card enforcement, trace/eval evidence grading,
-  handoff UX, delegation contract roles, and compressed spec/onboarding paths.
-- Added `forge uninstall` to remove forge managed Codex/Claude
-  hook adapters while preserving sibling user hooks and Codex trust-state
-  residue.
-- Added fs-transaction manifests for `forge adopt --experimental-ts-apply`
-  and `forge adopt rollback --transaction <manifest>` so the safe
-  TypeScript applicator has an executable recovery path.
-- Added `scripts/check-release-published.sh` for post-publish npm registry,
-  dist-tag, tarball integrity, Git tag, and local version readback.
-- Added `scripts/check-tarball-install-smoke.sh` and wired it into the CI gate
-  so release checks install the packed tarball in a temporary project and start
-  the packaged `forge` and `forge-hook` bins.
-
-### Changed
-
-- Made `forge install` the primary first-run global runtime bootstrap
-  command, while preserving `forge init` as a compatibility alias and
-  keeping `forge install --target <host> --location <scope>` as the
-  adapter-only path.
-
-### Fixed
-
-- Made generated ChatGPT Connector `/goal` prompts and `forge-goal`
-  reporting guidance language-neutral, so installed skills and MCP goal handoff
-  output follow the user's language or repo-local instructions instead of
-  hard-coding Chinese prompt lines.
-- Made non-standard `forge adopt --mode minimal|self-host` fail closed
-  unless it routes to ordinary TypeScript `--dry-run` or
-  `--experimental-ts-apply`,
-  avoiding a mismatch where dry-run showed a TypeScript mode plan but apply
-  still used the standard shell migrator.
-- Made minimal TypeScript adoption install
-  `.ai/harness/workflow-contract.json`, preserving the hook opt-in marker for
-  all adoption modes.
-- Normalized protocol v1 JSON error output for adoption target validation
-  failures.
-- Added safe-applicator preflight inside `applyAdoptionPlan()` so exported
-  callers cannot partially write a plan before encountering an unsupported
-  operation.
-- Preserved CRLF `.gitignore` managed blocks without duplicate insertion and
-  added planned/skipped/failed summary counts to adoption plan output.
-- Added a bounded CLI process runner for init/adopt/global-runtime, CodeGraph
-  setup, and `forge run` helper dispatch, with default timeout, output
-  cap, and common secret redaction.
-
-### Release Notes
-
-- Prepared the `forge@0.7.0` minor release line for npm publish,
-  registry readback, clean-room `npx` smoke, Git tag, and GitHub release
-  creation.
-
-## [0.6.0] - 2026-06-16
-
-### Added
-
-- Added the Transactional Adoption Planner foundation for
-  `forge adopt --dry-run --json`, including protocol v1 operation plans,
-  redacted JSON rendering, fixture-backed planner tests, and a safe applicator
-  subset for `mkdir`, `writeFile ifMissing`, and managed `.gitignore` blocks.
-- Added workflow-contract manifest planning to the TypeScript adoption plan so
-  `standard` and `self-host` dry-run JSON now report installation of
-  `.ai/harness/workflow-contract.json` from the canonical tracked asset.
-- Added workflow-contract-backed adoption templates for `docs/spec.md` and
-  `tasks/current.md`, keeping their `writeFile ifMissing` planner behavior while
-  moving the file body and reason out of `plan.ts`.
-- Added standard-mode helper wrapper planning to the TypeScript adoption plan so
-  downstream dry-run JSON reports generated `scripts/<helper>` compatibility
-  wrappers from the workflow contract helper manifest.
-- Added an atomic safe-applicator writer with target locks, temp-file fsync,
-  parent-directory fsync, and backup metadata for existing targets.
-- Routed human-readable `forge adopt --dry-run` output through the
-  TypeScript adoption planner text renderer, matching the JSON dry-run source of
-  truth without writing repo files.
-- Added `forge adopt --experimental-ts-apply` as an opt-in TypeScript
-  safe-applicator path, with preflight rejection for plans containing
-  unsupported operations.
-- Added rollback metadata to adoption operation plans so dry-run JSON and
-  experimental apply reports expose the planned recovery strategy per operation.
-- Added workflow-contract install support to the experimental TypeScript
-  adoption applicator, including atomic replacement backups for stale manifests.
-- Added workflow-contract-backed adoption templates for `tasks/todos.md` and
-  `tasks/lessons.md`, completing manifest ownership of the initial bootstrap
-  ledger file bodies.
-
-### Release Notes
-
-- Prepared the `forge@0.6.0` minor release line for npm publish,
-  registry readback, clean-room `npx` smoke, Git tag, and GitHub release
-  creation.
-
-## [0.5.3] - 2026-06-15
-
-### Fixed
-
-- Fixed `forge update --version <version>` so the update subcommand can
-  install an explicit `forge@<version>` package. The top-level
-  `forge --version` / `forge -V` shortcut still prints the CLI
-  version, but it no longer intercepts the update command's package-version
-  option.
-
-### Release Notes
-
-- Prepared the `forge@0.5.3` patch line for npm publish, registry
-  readback, clean-room `npx` smoke, Git tag, and GitHub release creation.
-
-## [0.5.2] - 2026-06-15
-
-### Added
-
-- Added a weekly timestamp cache for SessionStart tooling-update advisories so
-  Waza and CodeGraph update checks run at most once per week by default unless
-  `REPO_HARNESS_TOOLING_ADVISORY_TTL_SECONDS` overrides the TTL.
-- Added reviewed security exceptions to `forge security scan` so
-  user-level warning-only hook findings can be accepted through exact
-  `filePath` + `ruleId` + `command` matches while high/fail findings stay
-  active.
-- Added transcript recovery to the bundled `claude-review` cross-review skill
-  for Claude Code print-mode runs that write the final assistant message to the
-  session transcript but produce empty stdout.
-
-### Changed
-
-- Treated `gbrain doctor --json --fast` connection warnings that only report
-  skipped DB checks as an accepted fast-mode readiness state, while preserving
-  real gbrain warnings such as freshness drift.
-- Pinned the repo-local CodeGraph dev dependency to `1.0.1` so local-first
-  readiness checks do not float across patch releases.
-
-### Fixed
-
-- `forge setup check --target codex --check-updates --json` now reports
-  a fully green setup on this machine after Waza and CodeGraph are up to date,
-  reviewed user-level safety hooks are recorded, and gbrain fast-mode DB checks
-  are treated as intentionally skipped.
-
-### Release Notes
-
-- Prepared the `forge@0.5.2` package line for npm publish, registry
-  readback, clean-room `npx` smoke, Git tag, and GitHub release creation.
-
-## [0.5.1] - 2026-06-14
-
-### Fixed
-
-- Fixed CodeGraph readiness detection so repo-local checks prefer the installed
-  platform bundle before the npm shim. This prevents a broken `.bin/codegraph`
-  shim from turning a ready local index into `project_index=unavailable`.
-- Synced the same CodeGraph resolver into the generated-project helper template.
-
-### Release Notes
-
-- Prepared the `forge@0.5.1` package line for publish; npm publish,
-  registry readback, and GitHub release creation remain explicit release
-  actions.
-
-## [0.5.0] - 2026-06-14
-
-### Added
-
-- Added README release art from `docs/images/image.png` and documented the
-  install/refresh split for first-run bootstrap, user-level runtime updates,
-  read-only setup audit, and repo-local adoption.
-- Documented the eight managed hook routes installed by the Claude/Codex
-  adapters: `SessionStart.default`, `PreToolUse.edit`,
-  `PreToolUse.subagent`, `PostToolUse.edit`, `PostToolUse.bash`,
-  `PostToolUse.always`, `UserPromptSubmit.default`, and `Stop.default`.
-
-### Changed
-
-- Breaking: `forge update` now owns CLI/user-level runtime refresh only;
-  use `forge adopt` for repo-local workflow install, refresh, and
-  migration.
-- `forge update --check` / `--no-runtime-refresh` now route to the
-  read-only setup checklist, and third-party skill/CodeGraph refreshes require
-  explicit opt-in.
-- Added `forge setup check` as the productized read-only readiness
-  command while keeping `forge init-hook` as a compatibility alias.
-
-### Fixed
-
-- Refused `$HOME` as a repo adoption target before any mutation and hardened
-  legacy context discovery so vendored/cache trees such as `go/pkg/mod`,
-  nested `node_modules`, and `vendor` are not mirrored into.
-
-### Release Notes
-
-- Prepared the `forge@0.5.0` package line for publish; npm publish,
-  registry readback, and GitHub release creation remain explicit release
-  actions.
-
-## [0.4.3] - 2026-06-13
-
-### Added
-
-- Added `forge docs list|path|show` so bundled runtime/reference docs
-  resolve from the user-level/package install instead of copied repo prose.
-- Added `forge init-hook --json` bootstrap audit guidance for working
-  rules, adapter drift, stale CLI installs, and tooling readiness.
-- Added first-principles edit guard coverage to the managed hook route set as
-  advisory anti-overengineering guidance.
-
-### Changed
-
-- Generated and migrated repos now write deterministic
-  `docs/reference-configs/*.md` pointer stubs while keeping `.ai/harness/*` and
-  `.ai/context/*` as repo-local runtime artifacts.
-- Retired `AGENTS.md` and `CLAUDE.md` from the reference-doc asset surface and
-  stopped publishing duplicate `docs/reference-configs/` runtime docs in the
-  npm package file list.
-
-### Fixed
-
-- Clarified plan-completeness capture wording so release/readiness checks do
-  not mistake a Draft scan plan for executable implementation state.
-- Made `scripts/check-npm-release.sh` refresh the current handoff before the
-  Codex resume packet so the release gate works from a clean checkout without
-  ignored runtime handoff files.
-
-### Release Notes
-
-- Prepared the `forge@0.4.3` package line for publish; npm publish,
-  registry readback, and GitHub release creation remain explicit release
-  actions.
-
-## [0.4.2] - 2026-06-13
-
-### Added
-
-- Added the `forge-prd` command facade, PRD template, and PRD eval
-  fixtures for generating upper-layer PRDs under `plans/prds/`.
-- Added a subagent return-channel guard to the managed Claude/Codex hook routes
-  so delegated runs are nudged back through the parent session instead of
-  leaking completion claims through the wrong channel.
-
-### Changed
-
-- Split upper-layer PRDs from Sprint backlogs: PRDs stay in `plans/prds/`,
-  Sprint backlogs move to `plans/sprints/*.sprint.md`, and `forge-sprint`
-  now supports PRD-to-Sprint planning without re-deciding product intent.
-- Isolated generated-project helper implementations under `.ai/harness/scripts/`
-  while keeping `scripts/*` as compatibility command wrappers.
-- Updated skill eval 24 and added evals for PRD generation and PRD-to-Sprint
-  backlog creation so the public command surface covers the new hierarchy.
-- Aligned the `forge-prd` command guidance with the installed
-  `.ai/harness/scripts/check-task-workflow.sh` runtime path.
-
-### Release Notes
-
-- Prepared the `forge@0.4.2` package line for publish; npm publish,
-  registry readback, and GitHub release creation remain explicit release
-  actions.
-
-## [0.4.1] - 2026-06-12
-
-### Fixed
-
-- Scoped the CodeGraph route nudge to the real hook stdin `session_id` by
-  exporting `HOOK_SESSION_ID` from the shared hook input parser and preferring
-  it in `session_state_resolve_key`. CodeGraph one-shot state is now per
-  Claude/Codex session instead of being pinned to a stale `.claude/.session-id`
-  fallback.
-- Prevented stale repo-local hook scripts from racing user-level host adapters:
-  generated and migrated repos now prune top-level `.ai/hooks/*.sh` unless
-  `.ai/harness/policy.json` explicitly pins `"hook_source": "repo"`.
-- Treated a missing `post-tool-observer.sh` on `PostToolUse.always` as a
-  soft-missing advisory route with an update hint, instead of hard-failing the
-  hook runtime when a copied repo-local hook set is stale.
-
-### Changed
-
-- Kept non-pinned downstream repos on the user-level hook runtime by retaining
-  only `.ai/hooks/lib/` helper fallbacks plus a README tombstone; this self-host
-  repo can still pin live hook development to `.ai/hooks`.
-- Normalized active workflow documentation to `tasks/todos.md` for the deferred
-  ledger and topic-scoped `docs/researches/*.md` for durable research reports.
-- Added `tasks/.current.md.tmp.*` and `.claude/.plan-state/` to the managed
-  runtime ignore block, and gave plain `bun test` the same 60s per-test timeout
-  used by the release gate.
-
-### Release Notes
-
-- Prepared the `forge@0.4.1` package line for publish; npm publish,
-  registry readback, and GitHub release creation remain explicit release
-  actions.
-
-## [0.4.0] - 2026-06-12
-
-### Added
-
-- Added a loop-engine evidence surface: `forge-hook state-snapshot
---json`, an NL decision-table reference, route NL-vs-TS benchmark fixtures,
-  and a cutover gate that keeps TypeScript routing authoritative unless
-  measured evidence passes.
-- Added `scripts/architecture-queue.sh` plus
-  `scripts/check-architecture-sync.sh` for derived architecture request indexes,
-  stale-index detection, and strict/advisory finish gates.
-- Added contract delegation metadata (`budget`, `permission_scope`, and
-  `roles`) to contract templates and generated workflow contracts.
-- Added `scripts/contract-run.ts`, a repo-local pilot runner that executes
-  explicit worker/verifier child commands and validates the verifier output
-  against contract exit criteria.
-- Added `scripts/heartbeat-triage.sh` and the `.ai/harness/triage/` surface for
-  scheduled workflow, sprint-next, and architecture-request triage.
-
-### Changed
-
-- Productized architecture queue assets into both self-host scripts and
-  generated-repo templates, including workflow contract inventory, reference
-  docs, migration handling, and scaffold parity tests.
-- Updated session-start/current-status projection and task workflow checks to
-  account for archived sprint plans, archived notes, contract delegation fields,
-  and the completed loop-engine sprint.
-- Retired the separate generated workflow compatibility `5.x` line; package,
-  skill, and template stamps now share the `forge@0.4.0` release line.
-- Added `check:architecture-sync` to the npm scripts and release verification
-  surface.
-
-### Removed
-
-- Retired `scripts/architecture-drift.sh` and its helper-template copy in favor
-  of `architecture-queue.sh`; migrations remove legacy copies from downstream
-  repos.
-
-## [0.3.0] - 2026-06-11
-
-### Added
-
-- Added the sprint program layer: `plans/sprints/`, sprint templates,
-  `scripts/sprint-backlog.sh`, the `forge-sprint` command facade, active
-  sprint markers, current-status projection, session-start projection, workflow
-  validation, and generated-repo parity copies.
-- Added `src/cli/hook/prompt-intents.ts`: every prompt-text intent classifier
-  now lives in TypeScript with real Unicode semantics, fixing
-  locale-dependent Chinese misclassification (UTF-8 continuation bytes
-  matching `[[:punct:]]` under `LC_ALL=C` grep, e.g. "实现会在这个 worktree
-  里完成。" misread as a done declaration on GNU grep).
-- Added an edit-layer plan gate to `pre-edit-guard.sh`: implementation edits
-  (paths outside plans/tasks/docs/deploy/harness/markdown surfaces) block
-  unless the active plan is Approved/Executing and `docs/spec.md` exists.
-  Modes `enforce` (default) | `advice` | `off` via policy
-  `.guards.edit_plan_gate` or `REPO_HARNESS_EDIT_PLAN_GATE`.
-- Added the `prompt-guard-decide` prompt protocol: the shell hook pipes
-  `{"prompt": ...}` on stdin and receives one verdict JSON line (action,
-  intent facts, derived strings). Legacy copied hooks that send env facts
-  still receive the bare action enum.
-
-### Changed
-
-- Hook runtime resolution is now central-first: user-level adapters dispatch
-  into `forge-hook`, central packaged hooks are the default runtime, and
-  repo policy can pin self-host development back to the repo copy without
-  changing downstream user adapters.
-- Prompt-layer plan/spec/contract gates became advisory routing; hard
-  enforcement moved to the PreToolUse edit layer where it keys off path +
-  plan state instead of natural-language guessing. Done-claim gates keep
-  blocking because they verify file-backed completion evidence.
-- Merged the PostToolUse always-route observers (`trace-event.sh` +
-  `context-pressure-hook.sh`) into one `post-tool-observer.sh`: one dispatch,
-  one stdin parse, and one library load per tool call. `.claude/.trace.jsonl`
-  is now the single tool-trace record (handoff "Commands Run" reads it
-  directly), and context-budget bun probes are sampled every 5th call. The
-  route tuple (PostToolUse, always) is unchanged; a workflow-contract
-  upgrade entry prunes the retired split hooks from migrated repos.
-- Unified `run-hook.sh`'s two Codex stdout-filter branches into one
-  parameterized path.
-
-### Removed
-
-- Removed the duplicated shell fallback decision table from
-  `prompt-guard.sh` (the 0.2.4 copied-hook fallback). Without a reachable
-  TypeScript engine the prompt layer now degrades to a one-shot advisory and
-  defers enforcement to the edit layer.
-- Removed the orphan `scripts/check-versions.ts` and its test, and the
-  hidden `prompt-guard-decision` CLI alias (use `prompt-guard-decide`).
-- Retired the `project-initializer` legacy name, `PROJECT_INITIALIZER_*`
-  environment fallbacks, and the `forge-skill` compatibility alias;
-  installed-copy sync now deletes both retired skill directories.
-
-## [0.2.4] - 2026-06-07
-
-### Added
-
-- Added shell fallback routing for copied prompt hooks when the TypeScript
-  decision engine is unavailable, preserving PlanCaptureGate guidance instead
-  of failing closed on installed hook copies.
-- Added workflow readiness checks for stale handoff/resume plan references and
-  action-command skill quality gates.
-- Added benchmark quality metrics so release evidence distinguishes
-  authoritative non-dry-run skill evals from dry-run smoke output.
-
-### Fixed
-
-- Treated plan/workflow consultation prompts as advisory text instead of
-  sending them into `PlanStatusGuard`, so questions that mention `new plan`,
-  `方案`, hooks, or workflow routing no longer create plan files or block with
-  "No active plan found" unless they explicitly start execution.
-
-### Changed
-
-- Updated the self-host CodeGraph development dependency to `0.9.9` and made
-  gbrain readiness probe `doctor --json --fast` before falling back to the full
-  doctor command.
-- Updated maintainer release docs and README verification guidance to require
-  non-dry-run skill eval evidence when claiming skill effectiveness.
-- Refreshed the ignored Codex resume packet inside the npm release gate before
-  strict workflow validation, so release tests that update handoff runtime state
-  cannot leave the gate failing its own stale-resume invariant.
-
-### Removed
-
-- Retired the self-host-only `autoresearch-advisory.sh` hook from `.ai/hooks`,
-  the generated hook adapter installer, and hook parity exceptions. Autoresearch
-  now stays an explicit agent-run workflow instead of a user-level background
-  hook route.
-
-## [0.2.3] - 2026-06-05
-
-### Changed
-
-- Replaced the public `forge init` path with a typed global bootstrap
-  that installs the current package as the global CLI, refreshes forge
-  skill aliases, installs user-level hook adapters, configures Waza
-  `think`/`hunt`/`check`/`health`, persists the brain root, and configures
-  CodeGraph MCP without applying repo-local workflow files to the current
-  directory.
-
-### Removed
-
-- Removed the Superpowers Claude marketplace installer path entirely from the
-  active `forge init` flow and from `scripts/setup-plugins.sh`.
-
-## [0.2.2] - 2026-06-04
-
-### Fixed
-
-- Streamed `forge init` setup output directly to the terminal so the
-  first-run `npx -y forge init` path no longer looks hung while
-  `setup-plugins.sh` clones skills or runs Claude plugin setup.
-- Made the Superpowers Claude marketplace plugin opt-in via
-  `forge init --with-superpowers` instead of installing it by default.
-
-## [0.2.1] - 2026-06-02
-
-### Added
-
-- Added `forge init` as a thin npm CLI wrapper around
-  `scripts/setup-plugins.sh`, so users can run
-  `npx -y forge init` for first-run global Claude plugin and hook-profile
-  bootstrap without cloning the source repository.
-- Added a prompt-guard CodeGraph self-heal path: before emitting the first
-  structural code-navigation hint in a session, a missing `.codegraph` index is
-  initialized with the local or PATH-visible CodeGraph binary without running the
-  heavier readiness probe.
-
-### Changed
-
-- Moved the existing repo-local harness install/refresh CLI surface to
-  `forge update`, keeping `forge init` focused on global runtime
-  initialization.
-- Updated the English, Chinese, Japanese, French, and Spanish READMEs for the
-  `0.2.1` npm release line and the split `init` / `update` lifecycle.
-
-### Fixed
-
-- Kept automatic hook-side CodeGraph initialization non-blocking and cleaned up
-  the Cursor rule file if current CodeGraph created it only as a side effect of
-  this automatic init.
-
-## [0.2.0] - 2026-06-02
-
-### Added
-
-- Added a read-only config security scan (`forge security scan [--json]`) that checks high-value hook and editor-task config (`~/.claude/settings.json`, `~/.codex/hooks.json`, repo-local `.vscode/tasks.json`, and legacy project-level `.claude`/`.codex` adapters) for suspicious command patterns — remote-shell pipes, base64-decode-to-exec, `osascript`, `launchctl`/`crontab` persistence, netcat, and inline interpreter execution — plus unmanaged hook commands and auto-run `folderOpen` tasks. It reports findings only and never mutates config.
-- Added a low-frequency `SessionStart` sentinel (`.ai/hooks/security-sentinel.sh`, wired into the `SessionStart.default` route) that fingerprints the config set and re-scans only when a fingerprint changes, surfacing a one-line `[SecurityConfig]` reminder when findings appear.
-- Added a `security-config` check to `forge doctor` backed by the same read-only scan.
-
-### Changed
-
-- Bumped the npm package release line from `0.1.5` to `0.2.0`; generated workflow compatibility stays on the `5.2.3` model line, and `forge --version` / `forge status` now report `0.2.0`.
-- Added `Why forge` and `What's New in 0.2.0` sections to the English, Chinese, Japanese, French, and Spanish READMEs, promoting file-backed cross-session coordination, CodeGraph-plus-progressive-context token savings, the `scripts/setup-plugins.sh` installer, the config security sentinel, and the Claude/Codex draft-plan lifecycle.
-- Added the README hero image to the npm package allowlist so package consumers get the same visual surface as the source checkout.
-- Fixed the Chinese README, which still referenced `0.1.4`, to track the current release version.
-
-## [0.1.5] - 2026-06-01
-
-### Changed
-
-- Added `REPO_HARNESS_*` environment variable aliases for scaffold, migration, context-block selection, external-tooling checks, and contract-worktree controls while preserving `PROJECT_INITIALIZER_*` as legacy fallbacks.
-- Switched new runtime `.gitignore` and Codex resume generated markers to `forge` while keeping dual-read compatibility for legacy `project-initializer` markers.
-- Added a dirty merged linked-worktree closeout guard to `ship-worktrees.sh --cleanup-merged`, requiring useful deltas to be committed, picked, or applied before cleanup and allowing only explicit scaffold-only discard.
-- Made `prepare-codex-handoff.sh` prefer Node for global handoff file updates, with Python retained as a fallback, so release verification does not depend on the local `python3 -` execution path.
-
-## [0.1.4] - 2026-05-31
-
-### Changed
-
-- Switched generated plan task artifacts from slug-only names to the active plan stem (`YYYYMMDD-HHMM-<slug>`) for `tasks/contracts/`, `tasks/reviews/`, and `tasks/notes/`.
-
-### Fixed
-
-- Kept workflow-state, handoff, archive, and contract-worktree helpers compatible with existing slug-only task artifacts while preferring the new plan-stem paths.
-
-## [0.1.3] - 2026-05-31
-
-### Added
-
-- Added AI-native scaffold profiles as overlays on the existing A-K plan catalog, including runtime-console, product-copilot, and sidecar-kernel project structures without introducing new public plan codes.
-- Added AI-native template variables so selected profiles can project focused project structures, runtime-console defaults, and tech-stack guidance while ordinary A-K scaffolds stay unchanged.
-- Added a typed prompt-guard decision engine behind `forge-hook prompt-guard-decide`, keeping host adapters stable while making `intent x plan state` routing table-driven and testable.
-- Added CLI and route-level regression coverage for the internal prompt-guard decision command, the lightweight hook entrypoint, and the public `UserPromptSubmit --route default` path through real hook assets.
-- Added an optional deploy SQL invariant coverage check: when `tests/sql/control_plane_invariants.sql` exists, `check-deploy-sql-order.sh` now verifies every `deploy/sql/*.sql` migration is referenced by full path or basename.
-- Added a dated release filing under `deploy/release-checklists/260531-forge-0.1.3.md` and documented the `YYMMDD-<package>-<version>.md` filing rule.
-
-### Changed
-
-- Split prompt-guard responsibilities so shell continues to parse hook JSON, read workflow files, perform capture side effects, and render host-safe output while TypeScript owns the explicit decision table.
-- Documented the 0.1.x release surface as `forge@0.1.3`, still separate from the generated workflow compatibility line (`5.2.3`).
-- Expanded the English and Chinese README plus the hook operations reference to show the current host adapter -> CLI route registry -> shell hook -> TypeScript decision table architecture.
-
-### Fixed
-
-- Routed active Draft plan prompts such as `implement this plan` and `执行这个方案` to the non-blocking PlanCaptureGate instead of hard-blocking under PlanStatusGuard.
-- Routed no-active-plan and Approved-plan execution projection prompts through the appropriate capture/projection advice instead of collapsing them into generic PlanStatusGuard or ContractGuard failures.
-- Treated copied worktree status, retrospective completion reports, and next-slice planning summaries as passive context so they do not start implementation gates merely because they quote implementation vocabulary.
-- Ensured linked contract worktrees include `.ai/harness/planning/` before pending orchestration cleanup, preserving strict workflow verification in generated worktrees.
-- Filtered `tasks/.current.md.tmp.*` refresh scratch files out of generated `tasks/current.md` snapshots, including generated repo helper parity.
-- Aligned `forge --version` and `forge status` with the `package.json` release version for `0.1.3`.
-
-## [0.1.2] - 2026-05-30
-
-### Added
-
-- Added `forge init` as a one-shot existing-repo bootstrap that defaults `--repo` to the current working directory, refreshes host adapters, applies the harness, installs Waza runtime skills, syncs `diagram-design`, and verifies the repo-local workflow.
-- Added `forge init --no-codegraph` and `--configure-codegraph` so existing-repo bootstrap can either skip CodeGraph readiness or explicitly register CodeGraph MCP after building the index.
-- Added `check:release` / `prepublishOnly` npm release gates that check the official npm registry and reject already-published package versions before running tests, workflow checks, migration dry-run, and pack dry-run.
-- Added a GitHub-facing bilingual README path with `README.zh-CN.md` and a Mermaid task workflow from plan to contract worktree checkout, guarded implementation, verification, review, external acceptance, finish, merge, and cleanup.
-
-### Changed
-
-- Retired `project-initializer` as a Codex/Claude installed skill path and upstream resolver fallback; installed-copy sync now removes those directories instead of maintaining them.
-- Switched generated footer stamps to `forge@...` while keeping `.claude/.skill-version` semantic version fields stable.
-- Prepared npm publishing under the unscoped `forge` package name, made `forge` the primary installed command, and kept `forge-skill` as a compatibility alias.
-- Split the npm/CLI package release line (`0.1.x`) from the generated workflow compatibility line (`5.2.3`).
-- Updated GitHub repository metadata and source checkout docs for the `Ancienttwo/forge` rename.
-- Forced copy-based installed-skill sync when `forge init` runs from an npm `_npx` cache source, avoiding symlinks to temporary npx cache directories.
-- Clarified the product boundary, three-layer operating model, and task lifecycle on the README landing page.
-
-### Fixed
-
-- Rebuilt Claude skill aliases during installed-copy sync so `~/.claude/skills/project-initializer` cannot remain on a stale legacy repo while Codex runtime aliases are current.
-- Reduced full-suite release flakiness by giving `doctor` environment-probe tests a wider timeout budget.
-
-## [5.2.3] - 2026-05-27
-
-### Fixed
-
-- Expanded anchored approval intent variants such as `go ahead with it`, `please proceed`, and `可以干了` so post-plan approvals reach `PlanCaptureGate` / `PlanExecutionGate` without treating broad bug-fix wording as approval capture.
-
-## [5.2.2] - 2026-05-27
-
-### Fixed
-
-- Started a Draft `plans/` artifact as soon as explicit Codex Plan mode or Waza `/think` planning begins, so plan lifecycle state exists before approval and execution gates run.
-- Let terse approval prompts such as `GO` and `可以干` reach the approved-plan capture/projection path instead of being blocked before the agent can run `capture-plan.sh` or `plan-to-todo.sh`.
-
-## [5.2.1] - 2026-05-27
-
-### Fixed
-
-- Fixed terse `GO` approval prompts after Codex Plan mode or Waza `/think` so they trigger `PlanStatusGuard` and route execution through captured `plans/` artifacts instead of bypassing the workflow gate.
-
-## [5.2.0] - 2026-05-27
-
-### Changed
-
-- Added passive plan capture so Codex Plan mode, Waza `/think`, and `forge-plan` outputs can become file-backed `plans/plan-*.md` artifacts through `scripts/capture-plan.sh`, with approved captures able to project directly through `plan-to-todo.sh`.
-- Added opt-in default-brain document mirroring through `scripts/sync-brain-docs.sh`, manifest `sync.direction=repo-to-brain` entries, and PostEdit hook integration for registered valuable docs.
-- Promoted CodeGraph from advisory setup guidance to required Codex agent readiness for code navigation, with read-only detector support, strict readiness checks, generated repo `.codegraph/` ignores, and non-vendored host install guidance.
-
-## [5.1.2] - 2026-05-27
-
-### Added
-
-- Added generated Codex hook adapter support through `.codex/hooks.json` while keeping `.ai/hooks/` as the shared hook implementation layer.
-- Updated init, scaffold, migration, workflow contract, docs, and tests so generated repos install both Claude and Codex hook adapters.
-
-## [5.1.1] - 2026-05-26
-
-### Fixed
-
-- Refreshed stale `references/` docs for the current `forge` hook, migration, eval, plugin, and minimal-documentation contracts.
-- Updated public-surface spec and architecture docs to reflect the full 13-command `agentic-dev-*` facade inventory.
-- Removed empty optional doc placeholders so generated/self-hosted docs match the `minimal-agentic` profile.
-
-## [5.1.0] - 2026-05-26
-
-### Added
-
-- Added filesystem-owned Evidence Contract fields and guards so approved plan execution must name state/progress path, verification evidence, evaluator rubric, stop condition, and rollback surface before implementation or completion.
-
-### Changed
-
-- Made broad research delegation a main-agent spawn decision based on context impact and callable runners, with bounded main-thread fallback when spawning is not useful or available.
-- Hardened Waza external-tooling checks to compare whole skill directories and shared `rules/` files instead of only `SKILL.md`, catching broken `references/`, `scripts/`, `agents/`, and cross-skill rule links.
-
-## [5.0.2] - 2026-05-25
-
-### Fixed
-
-- Excluded ignored repo-local runtime state from Codex installed-copy sync outputs.
-
-## [5.0.1] - 2026-05-25
-
-### Added
-
-- Added the forge plugin architecture map, domain/module docs, and capability-indexed local context contracts for Claude and Codex.
-
-### Fixed
-
-- Fixed Codex installed-copy sync for symlinked legacy `project-initializer` fallback paths.
-- Removed tracked Claude trace state from the release surface and ignored repo-local Codex/runtime logs.
-
-## [5.0.0] - 2026-05-25
-
-### Fixed
-
-- Made repeated `migrate-project-template.sh --apply` idempotent after a clean migration commit by normalizing first-write JSON output and preserving unchanged version stamps.
-- Removed stale `3.1 guidance` wording from migration dry-run output.
-
-### Changed
-
-- Added `deploy/sql/` as the tracked deployment SQL surface and wired a filename-order check for `0001_name.sql` style files.
-- Split deployable operations assets into tracked `deploy/` while keeping `_ops/` fully ignored for local private operations state and secrets.
-- Externalized long-form optional reference configs into the default brain file vault while keeping repo-local runtime contracts, hooks, scripts, and required minimal docs authoritative.
-- Added a repo-local brain manifest and workflow check for default brain pointers without making hooks depend on gbrain or iCloud.
-- Renamed the skill/package/repo display surface to `forge` while keeping `forge-skill` and `project-initializer` as legacy aliases, install paths, and generated stamp compatibility surfaces.
-- Added action-style `agentic-dev-*` command skill facades for plan, review, autoplan, init, scaffold, migrate, upgrade, repair, and check while keeping hooks/docs initialization internal.
-- Added advisory prompt-hook route hints for reusable-workflow packaging, with `forge-autoplan` handling evidence-first plans only after user authorization.
-- Added a Codex installed-copy sync helper that keeps command facades only in the canonical `forge` copy while legacy directories remain runtime fallback bundles.
-
-## [4.0.2] - 2026-05-20
-
-### Fixed
-
-- Installed `inspect-project-state.ts`, `migrate-workflow-docs.ts`, `workflow-contract.ts`, `check-skill-version.ts`, and a delegating `migrate-project-template.sh` wrapper into generated repos so the router verification path is not left stale.
-- Made generated capability discovery ignore `.worktrees/` and `_ref/` caches, preventing local worktree contracts from polluting `.ai/context/capabilities.json`.
-
-## [4.0.1] - 2026-05-20
-
-### Added
-
-- Added a versioned upgrade strategy to the workflow contract, inspector output, harness policy, and migration cleanup path so legacy reconfiguration, archives, preserves, and removals are auditable.
-- Added `docs/reference-configs/global-working-rules.md` as the user-level Claude/Codex rule template with enforceable P1/P2/P3 due diligence.
-
-## [4.0.0] - 2026-05-20
-
-### Changed
-
-- Removed `docs/PROGRESS.md` from default generated and required workflow surfaces; legacy progress files are now archived during migration instead of normalized in place.
-- Replaced default root `specs/` scaffolding with `docs/spec.md`, `interfaces/`, and tests as the stable product/runtime truth surfaces.
-- Promoted `_ops/` as the trackable operations workspace for runbooks, submission materials, release checklists, and helper scripts, while keeping `_ops/secrets/` and `_ops/env/.env*` ignored.
-- Made `_ref/` an ignored external comparison cache and added hook guards that block product edits under `_ref/` and sensitive `_ops` env/secret paths.
-- Updated workflow contracts, generated templates, reference docs, architecture index, and tests to use `tasks/workstreams/` for durable progress and `docs/CHANGELOG.md` for release history.
-
-## [3.6.0] - 2026-05-19
-
-### Added
-
-- Added `minimal-agentic` documentation generation so default scaffolds keep only required docs plus a small reference-config set, with `PROJECT_INITIALIZER_DOCUMENTATION_PROFILE=full` preserving the previous full docs surface.
-- Added `docs/reference-configs/document-generation.md` to document required docs, on-demand docs, and the Agent-owned decision boundary.
-- Added `lsp_profiles` metadata to policy and context maps so selected functional blocks can carry lightweight tooling hints without expanding root prompt context.
-- Added `worktree_strategy` policy for conflict-triggered `codex/<task-slug>` worktrees, Waza `/check`-style validation, and merge-back to `main` without absorbing unrelated dirty changes.
-- Added implementation notes as a task-local workflow artifact under `tasks/notes/`, with plan, contract, review, handoff, and archive integration.
-- Added raw verification run snapshots under `.ai/harness/runs/` so `checks/latest.json` remains a pointer while durable evidence stays inspectable.
-
-### Changed
-
-- Updated scaffold, migration, init, ensure, workflow contract, and tests to install reference configs through the documentation profile instead of copying every reference doc by default.
-- Changed init/migration external-tooling reports to skip update checks by default; set `PROJECT_INITIALIZER_CHECK_TOOLING_UPDATES=1` when an advisory run should also check upstream versions.
-- Updated harness policy and reference docs to distinguish notes, evidence, promoted assets, and advisory memory instead of collapsing task-local decisions into long-term memory.
-
-## [3.5.0] - 2026-05-11
-
-### Added
-
-- Added machine-readable `agentic_development` routing so product discovery uses gstack `office-hours`, complex engineering plans use gstack `plan-eng-review`, design plans use gstack `plan-design-review`, and daily small/medium work uses Waza `/think`, `/hunt`, and `/check`.
-- Added `docs/reference-configs/agentic-development-flow.md` to keep detailed gstack/Waza routing and P1/P2/P3 due-diligence triggers out of root prompts.
-- Added plan and review template sections for selected route, routing reason, and P1/P2/P3 evidence.
-- Added `scripts/select-agent-context-blocks.sh` as the functional-block selector hook for paired `CLAUDE.md` and `AGENTS.md` generation, so Claude Code and Codex receive the same local module contract without inferring boundaries from broad layout globs.
-
-### Changed
-
-- Stopped generating repo-local `.claude/hooks/` shim scripts by default; `.ai/hooks/` is now the shared hook implementation layer and `.claude/settings.json` is the Claude adapter.
-- Updated scaffold, migration, workflow contract, policy defaults, reference configs, and tests to keep self-host and generated repos aligned.
-- Hardened workflow verification and legacy task migration around runtime contract parsing and partially migrated `tasks/todo.md` files.
-
-## [3.4.0] - 2026-05-06
-
-### Added
-
-- Added Codex-first Waza policy metadata to the harness contract and generated repo policy defaults.
-- Added host-aware Waza detection for real Claude/Codex skill paths, per-skill versions, symlink targets, staging drift, and upstream stale status.
-- Added tests covering Claude staging symlinks, Codex independent runtime copies, read-only update checks, and Codex stale drift reporting.
-
-### Changed
-
-- Changed Waza `--check-updates` handling to compare upstream `tw93/Waza` raw `SKILL.md` hashes without running mutating `npx skills check`.
-- Documented the Waza stage -> copy into Codex -> `cmp` verification workflow for generated and self-hosted harnesses.
-
-## [3.3.0] - 2026-04-19
-
-### Changed
-
-- Removed repo-local Skill Factory and Claude auto-memory surfaces from the shared harness, migration path, and self-hosted repo.
-- Added `scripts/check-agent-tooling.sh` plus generated `docs/reference-configs/external-tooling.md` so init and migrate flows can report gstack, Waza, and gbrain advisory status safely.
-- Merged guidance-only `external_tooling` defaults into `.ai/harness/policy.json` during scaffold and migration without overwriting explicit repo overrides.
-
-## [3.2.1] - 2026-04-19
-
-### Fixed
-
-- Added progressive context and harness policy surfaces alongside the workflow contract manifest so generated repos keep root context stable while exposing deeper context on demand.
-- Wrote directory-level `AGENTS.md` files to discoverable module paths like `apps/*/AGENTS.md` instead of the container roots.
-- Stopped custom plan `K` from creating `apps/`, `packages/`, and `services/` unless the target repo already has real module directories there.
-- Corrected `scripts/inspect-project-state.ts` routing so initialized repos with bundled Skill Factory assets still classify as `audit` instead of collapsing to `skill-factory`.
-- Tightened `scripts/check-task-workflow.sh` so strict workflow verification now fails explicitly when no `node`, `bun`, or `python3` runtime is available to read the workflow contract.
-- Extended `scripts/migrate-workflow-docs.ts` to normalize legacy `tasks/todo.md` content in partially migrated repos and preserve the prior checklist in `tasks/archive/legacy-tasks-todo.md`.
-
-## [3.2.0] - 2026-04-08
-
-### Changed
-
-- Added `assets/workflow-contract.v1.json` as the single machine-readable workflow contract and installed `.ai/harness/workflow-contract.json` in generated and self-hosted repos.
-- Introduced `scripts/inspect-project-state.ts` so routing starts from structured repo inspection instead of prompt-only branching.
-- Added `scripts/migrate-workflow-docs.ts` to preserve and migrate legacy `docs/plan.md`, `docs/TODO.md`, and execution-log style `docs/PROGRESS.md`.
-- Updated migration, scaffold, and workflow verification paths to consume the shared contract manifest and verify it after migration.
-
-## [3.1.0] - 2026-03-29
-
-### Changed
-
-- Added `run_id` to trace events, verification reports, and task-state snapshots for tighter report correlation.
-- Expanded harness defaults to five dimensions by adding recovery and state profiles to the initializer question pack and plan map.
-- Added structured `failure_class` logging plus `scripts/summarize-failures.sh` for guard failure aggregation.
-
-## [3.0.0] - 2026-03-25
-
-### Changed
-
-- Upgraded generated repositories from a tasks-first scaffold to a shared long-running harness model.
-- Added `docs/spec.md`, `tasks/reviews/`, and `.ai/harness/{checks,handoff}` as first-class generated artifacts.
-- Reworked hook behavior around artifact-aware execution gates, contract scope enforcement, structured checks, and mandatory handoff generation.
-- Upgraded the initializer question pack to `v2` and added stack-aware orchestration, evaluation, and handoff defaults.
-- Updated helper scripts, templates, CLAUDE/AGENTS routing output, and tests to the shared harness model.