@hecer/yoke 1.10.0 → 1.11.0

package/CHANGELOG.md

@@ -1,379 +1,398 @@
-# Changelog
-
-## 1.10.0 — 2026-09-06
-
-### Added
-- Adopt the Yoke wordmark in the GitHub and npm README.
-- Prepare task assessments in one bounded planning call with `yoke prd assess`; draft and inbox planning bind assessments to requirements, upstream dependencies and the approved brief.
-- Separate `planning.agent`/`model`/`reasoningEffort` from execution settings. New setups require prepared assessments and block missing worker profiles; existing configurations retain on-demand planning and parent fallback.
-- Add selective reassessment, planning input limits and `routing.maxTier` to bound automatic model selection and escalation.
-- Add attention-first project search and status filters, with separate goal/loop states and explicit stale activity warnings.
-- Restore dashboard views and UTC period controls through URL links, browser history and refresh; cancel obsolete requests and bound project comparison concurrency.
-- Compare recorded tokens, accepted tasks and reported costs against the preceding equal-duration period without inventing missing measurements or percentages from zero baselines.
-
-### Fixed
-- Fix Windows safe-mode Codex execution by screening Store PowerShell/aliases and probing a native shell under the same sandbox before model work; keep permissions intact.
-- Launch Windows provider executables/npm entry points with literal argv; bound preflight and provider lifetime, recognize streamed infrastructure failures, retain failed-worktree/process evidence, and guard against unconfirmed termination before another worker starts.
-- Separate provider liveness, supervisor heartbeat, output and successful-tool progress in status/dashboard; label backlog percentages and preserve explicit bare startup through capability routing. See [Windows runner validation](docs/WINDOWS-RUNNER-VALIDATION.md).
-- Keep local loop locks and dashboard status out of implementation commits and clean-worktree checks.
-- Stage implementation files safely when runtime history directories are already ignored, including literal filenames and tracked deletions, for serial commits and parallel candidate snapshots.
-
-### Migration and validation limits
-- Existing routing configurations retain on-demand assessment and parent fallback. Use `yoke prd assess` to prepare task packages; configure `planning.agent`, `planning.model`, `planning.reasoningEffort` and `routing.maxTier` to separate planning from bounded execution. New setups require prepared assessments and block missing profiles.
-- Update installed packages and restart the dashboard/runner to use the new code. Existing running processes are not upgraded or instrumented retroactively. Windows preflight preserves sandbox permissions and changes only the provider environment.
-- Dashboard comparisons report recorded measurements, not reconstructed history or calibrated savings. Tokens per minute are consumption rates, not generation speed. Live routed Windows validation covered Codex on the documented machine; cross-provider performance and universal Windows compatibility are not established. See the linked dashboard, batch-planning and runner validation records.
-
-## 1.9.0 — 2026-09-06
-
-### Added
-- Add capability-based routing with persisted task assessments, explicit model/effort tiers, role eligibility and conservative use of independent task-class outcomes.
-- Keep planning on the start model; reuse assessments across attempts/worktrees and invalidate them when task requirements change.
-- Add bounded repair and tier escalation after mechanical gate failures, retaining the patch and forwarding failure evidence. Infrastructure failures do not trigger capability escalation.
-- Apply task-based profiles to reviews, quality critics/repairs and goal execution; display implementation selection reasons and next escalation in the dashboard.
-- Add setup options `--routing-strategy=capability` and `--routing-preset` for explicit migration. Preserve existing strategies and custom profiles by default.
-
-### Validation limits
-- Initial profile tiers are configurable hypotheses, not authenticated model benchmarks, price estimates or calibrated success probabilities. See [capability routing](docs/CAPABILITY-ROUTING.md) for defaults and bounds.
-
-## 1.8.0 — 2026-09-06
-
-### Added
-- Add persistent local measurement history and dashboard views for current work, usage/time and results, including UTC day/week/month filters, model history, project comparisons and consumption charts.
-- Display current tasks, worker phases, integration progress and status age, with automatic refresh of the current-work view.
-- Record explicit acceptances and show measured tokens and time per acceptance. Attribute available reviewer, critic and repair usage; distinguish actual reported models, unknown calls and partial costs.
-
-### Changed
-- Enable routing in new setups and automatically select up to three parallel workers when all pending tasks declare write scopes. Dependencies and overlapping scopes still constrain dispatch. Preserve explicit opt-outs and use isolated worktrees by default.
-- Share routing decisions between synchronous and asynchronous runners; support routed parallel workers, stable recovery history and explicit provider affinity.
-- Reserve execution capacity through integration and disable native delegation for loop providers; preserve Gemini system policy in a temporary bounded-execution configuration.
-- Require a dated changelog entry, synchronized version metadata and verified release checks for every new version in the project instructions.
-
-### Fixed
-- Avoid conflicting Codex sandbox arguments and prevent Codex-only options from leaking into Gemini workers.
-- Keep compact measurement history after recent activity expires, deduplicate archived events, and report incomplete history instead of treating missing usage as zero.
-- Preserve reviewer telemetry and worker/model attribution across parallel execution and recovery.
-
-### Migration and validation limits
-- Use `--parallel=N` to choose a worker limit, `--parallel=auto` for automatic selection, `--no-routing` to opt out of routing, and `--no-isolate` to opt out of default isolation. Explicit existing configuration remains authoritative. Unknown write scopes, tool actions and worktree recovery select serial execution in auto mode.
-- Automatic routing needs configured profiles; otherwise it keeps the selected parent provider. Explicit `--routing` without profiles reports a configuration error.
-- Missing historical usage cannot be reconstructed. Tokens per minute describe interval or summed call consumption, not measured generation speed. Live authenticated provider benchmarks, resource-adaptive concurrency and calibrated time/cost predictions are not established by this release.
-
-## 1.7.0 — 2026-09-05
-
-### Added
-- Independent `yoke check` with executable acceptance mapping, protected test infrastructure and content-bound evidence.
-- Durable project goals, provider handoff, checkpoint budgets, interruption accounting and project-scoped recovery.
-- Local project registry and loopback dashboard with goals, task estimates, evidence, consumption and pause controls.
-- Explicit routing rules with persisted gate-driven escalation; bounded tool actions without model calls.
-- Task-aware context packets, advisory write scopes, dependency-depth scheduling and empirical time ranges with prediction error records.
-
-### Fixed
-- Failed serial isolated worktrees are retained and can be explicitly resumed against their original target and PRD.
-- Reviewer fingerprints include untracked contents and acceptance inputs; unsupported nested repository identity fails closed.
-- Gemini always emits streaming telemetry, preserves model identity, honors aggregate token aliases and rejects unsupported selections. Its RTK hook merges native settings without a shell dependency.
-- Runtime evidence is excluded from Git gates and commits in existing projects. Partial usage and costs remain visibly incomplete.
-- Protected acceptance is checked after verification and repair, and linked goal state cannot overwrite unrelated files through pause.
-
-### Validation limits
-- Live authenticated provider comparisons and calibrated development-time/cost estimates are not established by the automated tests. Browser proof and independent model review still require explicit configuration.
-
-## 1.6.2 — 2026-09-02
-
-### Added
-- GitHub Releases can now publish `@hecer/yoke` through npm trusted publishing with short-lived OIDC credentials and automatic provenance, without a long-lived npm token.
-
-### Fixed
-- Codex safe-mode invocations now use the supported `workspace-write` sandbox with `--approve-for-me`; the removed `--full-auto` flag no longer blocks current Codex CLI releases.
-- Windows provider cleanup rechecks termination after process close and accepts an already-absent process as successfully cleaned up, avoiding stale ownership records and unnecessary watchdog waits.
-- Successful stale-loop cleanup removes obsolete runtime status, so `yoke loop status` no longer reports a dead run as `RUNNING`.
-
-## 1.6.1 — 2026-08-21
-
-### Fixed
-- Release metadata counts platform-conditional tests consistently on Windows and Ubuntu.
-- Provider cleanup confirms termination after child close, preventing stale ownership records when Windows reports process exit asynchronously.
-
-## 1.6.0 — 2026-08-20
-
-### Added
-- The canon now includes `no-ai-slop`, `domain-modeling`, `codebase-design`, `resolving-merge-conflicts`, and `writing-for-agents`, with their supporting templates and evaluation material. Source adaptations are credited in `canon/skills/ATTRIBUTION.md`.
-- UI projects receive an automatic design-quality gate with configurable `design.mode` (`off`, `auto`, or `on`) and score budget. Detection uses package dependencies, UI source files, and configured smoke flows.
-- Durable context now includes a project glossary and can expose an optional bounded-context map.
-
-### Changed
-- Retrofit installs complete skill packages for Claude, Codex, and Gemini instead of copying only `SKILL.md`. Local resources, binary files, and executable bits are preserved where supported.
-- Every canon skill declares `invocation: auto|manual`; retrofit maps that policy to Claude frontmatter, Codex `agents/openai.yaml`, and Gemini's generated automatic-skill index.
-- Canon validation rejects missing local Markdown resources, unsafe package entries, symlinks, and provider metadata that conflicts with the manifest.
-
-### Fixed
-- Windows provider cleanup now treats a successful `taskkill` as a request and confirms that the recorded PID has stopped before deleting its ownership record.
-
-## 1.5.1 — 2026-08-17
-
-### Fixed
-- Serena MCP configurations generated by Yoke no longer open the local web dashboard automatically on every startup.
-
-## 1.5.0 — 2026-08-16
-
-### Added
-- Failed verify, executable-criterion, performance, configured custom-audit, and completion gates now produce deterministic byte-bounded previews and preserve large complete stdout/stderr in content-addressed `.yoke/artifacts/` files with SHA-256 references.
-- Projects can tune `output.previewBytes` and `output.artifactThresholdBytes`; existing projects use backward-compatible 2 KiB/8 KiB defaults.
-- A deterministic local benchmark verifies signal retention, preview bounds, compression measurement, and artifact digest round-trips without making provider-token claims.
-
-### Security
-- Output artifact paths sanitize story identifiers, stay below a project-local root, use user-only file modes where supported, and are excluded from Yoke's clean-tree and story-commit operations even in upgraded projects. Raw artifacts are never injected automatically and documentation warns that project commands may emit secrets or personal data.
-- Gate command capture is capped at 16 MiB per stdout/stderr stream. Quota overflow fails closed and labels retained evidence as truncated instead of risking unbounded memory or claiming partial output is complete.
-
-## 1.4.0 — 2026-08-15
-
-### Added
-- `yoke loop run --parallel=N` now executes dependency-ready, non-colliding stories through real provider subprocess workers, isolated worktrees, leased claims, and a FIFO integration queue with fresh integrated-system gates.
-- Reference-driven quality declarations can collect screenshots, files, command output, or benchmark results and run a schema-validated blind critic with bounded repair rounds, elapsed-time limits, blocking or advisory policy, and retained proof.
-- `--candidates=N` can fan out up to five isolated implementations, discard mechanically red candidates, select one green candidate through identity-blind pairwise comparison, and preserve selected/loser evidence before cleanup.
-- Loop status now exposes dispatcher, worker, integrator, candidate lifecycle, worktree, queue, integration, reopen, quality-round, repair-budget, and trusted provider/model provenance data.
-
-### Changed
-- Provider subprocesses use explicit lifecycle contracts and incarnation-aware process records so worker cancellation and cleanup target only the process tree Yoke actually started.
-- Parallel and candidate runs disable adaptive routing, honor story-level provider affinity, latch pause requests across the whole dispatcher, and rerun quality plus review after integration.
-- `yoke loop cleanup` retains Yoke worktrees unless `--remove-worktrees` is explicit, while still reaping recorded orphan runners and stale locks safely.
-
-### Fixed
-- Expired claims, worker crashes, merge conflicts, pause races, and integration failures now release ownership deterministically, retain terminal proof, and reopen stories without leaking worktrees or marking false completion.
-- Quality repair fails closed on malformed critic output, reference drift, provider/model provenance mismatch, candidate identity leakage, unavailable critics, exhausted limits, and mechanically red repairs.
-- The watchdog resolves its TypeScript loader from both source and built npm layouts on Node 20+, and read-only Codex comparisons can run in disposable candidate worktrees without weakening normal repository checks.
-
-### Security
-- Blind comparison requests expose only opaque labels and digests while binding every verdict to the trusted judge provider, model, prompt, rubric, reference, and candidate provenance.
-- Cleanup and cancellation use project-scoped leases, owner tokens, PID birth/incarnation checks, and recorded process handles rather than machine-wide process-name matching.
-
-## 1.3.0 — 2026-08-09
-
-### Added
-- Teams can submit change requests at any time with `yoke change add` and inspect the append-only inbox with `yoke change status`; queued requests are planned at safe loop boundaries by Claude, Codex, or Gemini without interrupting active story work.
-- Acceptance criteria can carry stable IDs and executable verification commands, and Yoke records per-criterion evidence before a story can be marked complete.
-- Projects can configure an integrated completion command that proves the final system flow after all story-level gates pass.
-
-### Changed
-- New projects use strict criterion-proof requirements by default, while existing boolean-only criteria remain readable for compatibility.
-- PRD authoring, schemas, loop guidance, generated configuration, and provider instructions now treat completion as an ephemeral readiness state: new requests create more stories instead of prematurely forcing a release boundary.
-
-### Fixed
-- Broad green test suites can no longer satisfy unrelated structured acceptance criteria: proof commands must use an approved test runner, contain the criterion ID, and avoid shell operators.
-- Change-intake failures block safely; an independent coverage pass rejects omitted requested outcomes, crash recovery retains uncommitted requests, and concurrent PRD edits are detected instead of silently overwriting user work.
-- Integrated completion checks prevent locally finished stories from masking broken cross-component flows such as authentication callbacks or purchase-to-entitlement activation.
-
-### Security
-- Updated the transitive development dependency `nanoid` to a non-vulnerable release so the complete CI dependency audit is clean.
-- Restricted model-authored criterion proof to criterion-targeted test commands and blocked shell control operators and runner-prefix spoofing before host execution.
-
-## 1.2.1 — 2026-08-02
-
-### Fixed
-- `yoke loop run` now executes every remaining planned story by default instead of stopping after an implicit 25-iteration batch. Use `--max=N` only when an intentional bounded batch is wanted.
-- Unlimited runs remain unlimited after a critical-decision answer/resume cycle, while explicit caps remain preserved and accept positive integers only.
-
-## 1.2.0 — 2026-08-02
-
-### Added
-- Opt-in adaptive model routing lets a strong parent orchestrate each bounded story while Yoke selects an available Claude, Codex, or Gemini worker by quality, speed, cost, or balanced strategy.
-- A concurrency-safe local evidence registry learns from independent verification results without sharing mutable state between simultaneous Yoke processes.
-- Routing telemetry, reproducible benchmark fixtures, and analysis tooling make worker selection, token use, timing, and gate outcomes auditable.
-
-### Changed
-- Setup keeps routing disabled unless explicitly enabled and validates provider/model worker pools before execution.
-- Internal provider contract tests cover Claude Code, Codex CLI, and Gemini CLI through the shared adapter. Codex-only authenticated trials completed all 12 stories and 36 hidden checks; median routed runs used 11.0% fewer fresh input tokens, 49.5% fewer output tokens, 78.2% fewer reasoning tokens, and 33.8% less wall time than routing off.
-
-### Fixed
-- Reviewer prompts now permit their required verdict file while continuing to forbid project changes.
-- Failed reviewer subprocesses retain bounded provider stderr in loop status, exposing authentication, quota, sandbox, and startup failures instead of a generic command error.
-
-## 1.1.0 — 2026-07-30
-
-### Added
-- Shared five-question `yoke setup` wizard with provider-aware defaults for Claude, Codex, and Gemini.
-- Persisted default runner selection and `auto|critical` loop decision policies.
-- Provider-neutral `yoke-workflow` skill for planning questions, approved-plan PRD handoff, autonomous story execution, and critical-decision resume.
-- Structured critical-decision requests with `yoke loop decision` and `yoke loop answer`; answers are validated, committed under the configured human identity, and resume the same story.
-- Approved `.yoke/plan.md` context in PRD drafting and a lint gate for unresolved planning placeholders.
-
-### Fixed
-- Retrofit and loop on/off now preserve timeout, decision, runner, and permission settings.
-- Empty projects prefer the active agent host instead of silently installing Claude artifacts in Codex.
-- Loop and PRD runner selection now prefer an explicit flag, then the configured runner, then the active host.
-- Retrofit reports no longer label every provider as Claude Code.
-- Critical-decision resumes retain isolation, review, runner, permissions, timeout, JSON, policy, and iteration settings instead of falling back to an unreviewed default run.
-- Decision answers use an atomic owner-token lock and recoverable request journal, are checked against the active PRD story, bounded as untrusted data, and committed path-by-path so unrelated edits cannot enter the human-owned commit.
-- Decision recovery now binds the exact selected answer to its commit, rolls back only its own interrupted context append, namespaces resume state per project/worktree, and serializes cleanup with loop startup.
-- Active agent session markers now outrank globally configured provider home directories, and setup rejects partially invalid agent lists.
-
-### Changed
-- New setups enable the loop by default and choose `decisionPolicy: auto`; the wizard can select `critical` or disable the loop.
-- Legacy `loop.onAmbiguity` and `--on-ambiguity` remain compatibility aliases.
-
-## 1.0.0 — 2026-07-27
-
-### Added
-- Native Codex skills, project config, hooks, reusable agents, and plugin metadata.
-- Safe provider permission profiles and structured cross-provider telemetry.
-- Schema-validated independent review verdicts with explicit self-review opt-in.
-- Human-owned commit identity enforcement; AI co-author trailers default off.
-- `yoke audit` dependency, secret, and sensitive-diff gate with versioned suppressions.
-- PRD dependency graphs, collision areas, agent affinity, claims, FIFO merge queue, and bounded async dispatcher APIs.
-- Reproducible cross-runner benchmark schema and matrix launcher.
-
-### Changed
-- Dangerous permission bypass is opt-in via `--unsafe`.
-- Worktree cleanup is non-destructive unless `--remove-worktrees` is passed.
-- Reviews no longer trust process exit code alone.
-
-### Security
-- Vitest upgraded to 4.1.10; the dependency tree reports zero known vulnerabilities.
-
-
-## 0.9.0 — 2026-07-22
-
-### Added
-- **Performance budget gate** (`perf.command` in `.yoke/config.yaml`, optional `perf.retries`).
-  A benchmark command with the same contract as verify (exit 0 = within budget) runs **after
-  verify** on every story — new loop phase `perf`, `YOKE_STORY` exposed, worktree-aware in
-  `--isolate` mode. A red benchmark blocks the story
-  (`story S6 exceeded its performance budget: …`) no matter how clean the diff is. When the
-  gate is configured, the implementer prompt names the budget command so agents keep hot
-  paths efficient and never "simplify away" an optimization without re-running the benchmark.
-- **`performance` canon skill** (28 skills now): efficiency as a measured requirement —
-  clean-by-default with the decision ladder (minimal-code → measurable acceptance criterion →
-  project perf gate), profile-first, optimize leaves not boundaries, benchmarks committed as
-  tests, the *why* of every optimization versioned so future agents don't clean fast code
-  back to slow.
-- **`authoring-prd` guidance**: performance requirements belong in acceptance criteria as
-  numbers ("imports 1M rows in < 2s"), and every clarifying question belongs in the planning
-  round — a criterion still needing a decision is not loop-ready.
-
-## 0.8.0 — 2026-07-20
-
-### Added
-- **Live progress + ETA.** Story completions are now first-class events: the console shows
-  `✓ S6 done in 4m28s — 20/45 (44%) · ~1h40m left`, every status (file, NDJSON stream,
-  `yoke loop status`) carries `percent` and an `eta` block. The estimate averages the
-  durations of stories completed **in this run** (current velocity) and falls back to the
-  persisted history of previous runs (`.yoke/story-durations.json`, last 50, gitignored).
-  No data → no estimate, never an invented one.
-- **Ambiguity policy** (`loop.onAmbiguity` / `--on-ambiguity=<resolve|abort>`). The runner
-  prompt now always forbids asking questions (a loop run has nobody to answer). Default
-  `resolve`: the agent settles ambiguous criteria itself, states the interpretation, and the
-  loop never stops. Opt-in `abort`: the agent writes its open questions to
-  `.yoke/ambiguity.md` and stops; the loop consumes the file, skips verify (an unimplemented
-  story would otherwise pass on pre-existing green tests), and blocks with the question as
-  the reason. Companion principle: clarifying questions belong in the planning round, before
-  the loop starts.
-
-## 0.7.0 — 2026-07-17
-
-### Added
-- **Update check + `yoke upgrade`.** Every CLI invocation ends with a non-blocking
-  version hint (npm/gh-style): a detached background refresher caches the registry's
-  latest at most once a day; when it is newer, a one-line stderr hint suggests
-  `yoke upgrade` (which runs `npm install -g @hecer/yoke@latest`). Silent in CI,
-  `--json` runs, non-TTY pipes, and under `YOKE_NO_UPDATE_CHECK=1`.
-- **Opt-in auto-upgrade** (`update.auto: true` in `.yoke/config.yaml`): evaluated at
-  loop START only — never mid-run; the running process finishes on its version and
-  the upgrade applies from the next invocation. Deliberately NOT the default:
-  a gate harness must not change itself mid-project (determinism), and unreviewed
-  auto-installs are a supply-chain hazard.
-
-## 0.6.0 — 2026-07-17
-
-### Added
-- **Project-scoped orphan reaping.** The watchdog now records its pids in the project's
-  `.yoke/runner.pid` (main dir and per-story worktrees; removed on clean exit), and
-  `yoke loop cleanup` kills exactly those recorded process trees — and only while no
-  live loop holds the lock. Background: without a scoped mechanism, users and agents
-  resorted to machine-wide pattern kills (every process matching
-  `dangerously-skip-permissions`), which took down *healthy* runners of other projects
-  mid-story and stalled their loops. Never kill by pattern; `yoke loop cleanup` is the
-  safe path. `.yoke/runner.pid` is gitignored by retrofit.
-
-## 0.5.0 — 2026-07-17
-
-Root-cause fixes for the two "yoke keeps hanging" failure modes observed in the field
-(orphaned `claude.exe` runners piling up, healthy long stories dying at exactly the
-idle window):
-
-### Fixed
-- **Watchdog now kills the whole process tree on Windows** (`taskkill /T /F`).
-  Previously it killed only the spawned shell (`shell: true`), orphaning the actual
-  agent process — which kept writing to the worktree (dirty-tree blocks, failing
-  worktree removal) and kept burning API tokens. Observed in the field as ~10
-  zombie `claude.exe` per machine plus surviving dev servers.
-- **Claude runner always runs in stream-json mode.** Plain `-p` prints nothing until
-  the run finishes, so the idle watchdog mistook healthy >20-minute stories for dead
-  processes and killed them at exactly the idle timeout — while the user saw dead air.
-  The stream doubles as liveness; token usage is now reported on every run (not just
-  `--json` mode).
-
-### Changed
-- README: operating notes for driving the loop from inside an agent session
-  (background execution, small `--max` batches, `yoke loop cleanup` after interrupts) —
-  outer shell-tool timeouts killing a foreground `yoke loop run` were the third
-  observed "hang" pattern.
-
-## 0.4.0 — 2026-07-17
-
-### Added
-- **Hardened runner prompts** — distilled agent-harness patterns for headless runs:
-  scope discipline (nothing beyond the story), no unsolicited summary/plan/analysis
-  documents, root-cause fixes instead of gate bypasses, faithful outcome reporting,
-  bounded final messages (cuts output-token waste). Review prompts now ground verdicts
-  in observed evidence only and keep them brief.
-
-### Fixed
-- `.yoke/loop.pause` is now gitignored by retrofit. Previously the loop's own
-  `git add -A` story commit swept the pause control file into history in
-  un-retrofitted targets; removing it dirtied the tree and the clean-tree gate
-  blocked the resume run — the loop locked itself out.
-
-> Note: 0.3.0 was tagged and released on GitHub but never reached npm (2FA re-login
-> was pending), so for npm users 0.4.0 is the first release with the 0.3.0 changes below.
-
-## 0.3.0 — 2026-07-10
-
-### Added
-- **Claude Code plugin packaging** — the repo is now its own plugin marketplace
-  (`.claude-plugin/plugin.json` + `marketplace.json`): `/plugin marketplace add HECer/yoke`,
-  then `/plugin install yoke@yoke` installs the full canon under the `yoke:` skill namespace.
-- **Gemini CLI extension manifest** (`gemini-extension.json` + `GEMINI-EXTENSION.md`) —
-  installable via `gemini extensions install https://github.com/HECer/yoke`, listed in the
-  daily-crawled extensions gallery.
-- **Benchmark harness** (`bench/`) — reproducible cross-runner benchmark (tokens · speed ·
-  quality) with a fixed fixture, pre-written objective tests, and committed result data.
-- **Companion tool docs** — `canon/tools/claude-mem.md` (persistent memory; interactive
-  sessions only, explicitly kept out of loop runs) and `canon/tools/ui-ux-pro-max.md`
-  (design generation paired with Yoke's design verification gates).
-- **Multi-agent parallel loop design** — evaluation + phased design for distributing PRD
-  stories across parallel workers (`needs` dependency field, claim files, merge queue,
-  heterogeneous cross-agent dispatch): `docs/superpowers/specs/2026-07-10-multi-agent-parallel-loop-design.md`.
-
-### Fixed
-- Agent-availability probe timeout raised 5s → 20s: Gemini CLI cold-starts in ~6s on
-  Windows, so the loop misreported an installed `gemini` as "not found on PATH"
-  (found by the new benchmark harness).
-- Gemini runner invocation: dropped the bare `-p` flag — current Gemini CLI (0.33+)
-  requires a value after `-p` and errored with "Not enough arguments following: p".
-  Piped stdin selects headless mode by itself, so the runner now passes only `--yolo`
-  (also found by the benchmark harness).
-
-### Changed
-- README: npm install is now the primary quickstart path; documented plugin/extension
-  installs and optional companions.
-- npm package now ships `CHANGELOG.md`, `bench/` (harness + result data), and
-  `docs/superpowers/` (all specs and plans, including the multi-agent parallel loop design).
-
-## 0.2.0 — 2026-07-09
-
-- First npm release as `@hecer/yoke`.
-- Hyperflow integration surface: `yoke loop run --json` NDJSON stream, pause signal,
-  token-usage + model-id reporting for the claude runner.
-- `yoke new` greenfield bootstrap, `yoke prd draft`, cross-model `yoke review`,
-  `yoke flow-smoke` browser gate with proof artifacts, `yoke design-scan`.
-- Retrofit planners for Claude Code, Codex CLI, Gemini CLI; canon of 26 skills;
-  loop with worktree isolation, watchdog, single-flight lock, commit integrity.
+# Changelog
+
+## 1.11.0 — 2026-09-07
+
+### Added
+- Add Qwen Code (Alibaba) as fourth supported provider alongside Claude, Codex and Gemini.
+- Detect Qwen host environment via `QWEN_CLI` and `QWEN_CLI_HOME` environment variables.
+- Add Qwen routing workers with four capability tiers: `qwen-turbo-latest` (light), `qwen3-coder-plus` (standard/strong), `qwen3-235b-a22b` (frontier).
+- Parse Qwen streaming telemetry for token usage and model reporting.
+- Support Qwen in all CLI commands: `setup`, `loop`, `review`, `prd draft`, `prd assess`, `goal run`.
+
+### Changed
+- Update project description from "three agents" to "four agents" to reflect Qwen support.
+- Extend review resolution order to include Qwen for cross-model reviews.
+- Update setup prompts to offer Qwen as agent and runner option.
+
+### Migration and validation limits
+- Existing configurations remain compatible. New setups can select Qwen as agent/runner.
+- Qwen CLI uses Gemini-style arguments (`--approval-mode`, `--output-format stream-json`). `bare`, `reasoningEffort` and `nativeMultiAgent` selections are not supported (like Gemini).
+- Qwen routing profiles are configurable hypotheses, not authenticated benchmarks. Update installed packages and restart the dashboard/runner.
+
+## 1.10.0 — 2026-09-06
+
+### Added
+- Adopt the Yoke wordmark in the GitHub and npm README.
+- Prepare task assessments in one bounded planning call with `yoke prd assess`; draft and inbox planning bind assessments to requirements, upstream dependencies and the approved brief.
+- Separate `planning.agent`/`model`/`reasoningEffort` from execution settings. New setups require prepared assessments and block missing worker profiles; existing configurations retain on-demand planning and parent fallback.
+- Add selective reassessment, planning input limits and `routing.maxTier` to bound automatic model selection and escalation.
+- Add attention-first project search and status filters, with separate goal/loop states and explicit stale activity warnings.
+- Restore dashboard views and UTC period controls through URL links, browser history and refresh; cancel obsolete requests and bound project comparison concurrency.
+- Compare recorded tokens, accepted tasks and reported costs against the preceding equal-duration period without inventing missing measurements or percentages from zero baselines.
+
+### Fixed
+- Fix Windows safe-mode Codex execution by screening Store PowerShell/aliases and probing a native shell under the same sandbox before model work; keep permissions intact.
+- Launch Windows provider executables/npm entry points with literal argv; bound preflight and provider lifetime, recognize streamed infrastructure failures, retain failed-worktree/process evidence, and guard against unconfirmed termination before another worker starts.
+- Separate provider liveness, supervisor heartbeat, output and successful-tool progress in status/dashboard; label backlog percentages and preserve explicit bare startup through capability routing. See [Windows runner validation](docs/WINDOWS-RUNNER-VALIDATION.md).
+- Keep local loop locks and dashboard status out of implementation commits and clean-worktree checks.
+- Stage implementation files safely when runtime history directories are already ignored, including literal filenames and tracked deletions, for serial commits and parallel candidate snapshots.
+
+### Migration and validation limits
+- Existing routing configurations retain on-demand assessment and parent fallback. Use `yoke prd assess` to prepare task packages; configure `planning.agent`, `planning.model`, `planning.reasoningEffort` and `routing.maxTier` to separate planning from bounded execution. New setups require prepared assessments and block missing profiles.
+- Update installed packages and restart the dashboard/runner to use the new code. Existing running processes are not upgraded or instrumented retroactively. Windows preflight preserves sandbox permissions and changes only the provider environment.
+- Dashboard comparisons report recorded measurements, not reconstructed history or calibrated savings. Tokens per minute are consumption rates, not generation speed. Live routed Windows validation covered Codex on the documented machine; cross-provider performance and universal Windows compatibility are not established. See the linked dashboard, batch-planning and runner validation records.
+
+## 1.9.0 — 2026-09-06
+
+### Added
+- Add capability-based routing with persisted task assessments, explicit model/effort tiers, role eligibility and conservative use of independent task-class outcomes.
+- Keep planning on the start model; reuse assessments across attempts/worktrees and invalidate them when task requirements change.
+- Add bounded repair and tier escalation after mechanical gate failures, retaining the patch and forwarding failure evidence. Infrastructure failures do not trigger capability escalation.
+- Apply task-based profiles to reviews, quality critics/repairs and goal execution; display implementation selection reasons and next escalation in the dashboard.
+- Add setup options `--routing-strategy=capability` and `--routing-preset` for explicit migration. Preserve existing strategies and custom profiles by default.
+
+### Validation limits
+- Initial profile tiers are configurable hypotheses, not authenticated model benchmarks, price estimates or calibrated success probabilities. See [capability routing](docs/CAPABILITY-ROUTING.md) for defaults and bounds.
+
+## 1.8.0 — 2026-09-06
+
+### Added
+- Add persistent local measurement history and dashboard views for current work, usage/time and results, including UTC day/week/month filters, model history, project comparisons and consumption charts.
+- Display current tasks, worker phases, integration progress and status age, with automatic refresh of the current-work view.
+- Record explicit acceptances and show measured tokens and time per acceptance. Attribute available reviewer, critic and repair usage; distinguish actual reported models, unknown calls and partial costs.
+
+### Changed
+- Enable routing in new setups and automatically select up to three parallel workers when all pending tasks declare write scopes. Dependencies and overlapping scopes still constrain dispatch. Preserve explicit opt-outs and use isolated worktrees by default.
+- Share routing decisions between synchronous and asynchronous runners; support routed parallel workers, stable recovery history and explicit provider affinity.
+- Reserve execution capacity through integration and disable native delegation for loop providers; preserve Gemini system policy in a temporary bounded-execution configuration.
+- Require a dated changelog entry, synchronized version metadata and verified release checks for every new version in the project instructions.
+
+### Fixed
+- Avoid conflicting Codex sandbox arguments and prevent Codex-only options from leaking into Gemini workers.
+- Keep compact measurement history after recent activity expires, deduplicate archived events, and report incomplete history instead of treating missing usage as zero.
+- Preserve reviewer telemetry and worker/model attribution across parallel execution and recovery.
+
+### Migration and validation limits
+- Use `--parallel=N` to choose a worker limit, `--parallel=auto` for automatic selection, `--no-routing` to opt out of routing, and `--no-isolate` to opt out of default isolation. Explicit existing configuration remains authoritative. Unknown write scopes, tool actions and worktree recovery select serial execution in auto mode.
+- Automatic routing needs configured profiles; otherwise it keeps the selected parent provider. Explicit `--routing` without profiles reports a configuration error.
+- Missing historical usage cannot be reconstructed. Tokens per minute describe interval or summed call consumption, not measured generation speed. Live authenticated provider benchmarks, resource-adaptive concurrency and calibrated time/cost predictions are not established by this release.
+
+## 1.7.0 — 2026-09-05
+
+### Added
+- Independent `yoke check` with executable acceptance mapping, protected test infrastructure and content-bound evidence.
+- Durable project goals, provider handoff, checkpoint budgets, interruption accounting and project-scoped recovery.
+- Local project registry and loopback dashboard with goals, task estimates, evidence, consumption and pause controls.
+- Explicit routing rules with persisted gate-driven escalation; bounded tool actions without model calls.
+- Task-aware context packets, advisory write scopes, dependency-depth scheduling and empirical time ranges with prediction error records.
+
+### Fixed
+- Failed serial isolated worktrees are retained and can be explicitly resumed against their original target and PRD.
+- Reviewer fingerprints include untracked contents and acceptance inputs; unsupported nested repository identity fails closed.
+- Gemini always emits streaming telemetry, preserves model identity, honors aggregate token aliases and rejects unsupported selections. Its RTK hook merges native settings without a shell dependency.
+- Runtime evidence is excluded from Git gates and commits in existing projects. Partial usage and costs remain visibly incomplete.
+- Protected acceptance is checked after verification and repair, and linked goal state cannot overwrite unrelated files through pause.
+
+### Validation limits
+- Live authenticated provider comparisons and calibrated development-time/cost estimates are not established by the automated tests. Browser proof and independent model review still require explicit configuration.
+
+## 1.6.2 — 2026-09-02
+
+### Added
+- GitHub Releases can now publish `@hecer/yoke` through npm trusted publishing with short-lived OIDC credentials and automatic provenance, without a long-lived npm token.
+
+### Fixed
+- Codex safe-mode invocations now use the supported `workspace-write` sandbox with `--approve-for-me`; the removed `--full-auto` flag no longer blocks current Codex CLI releases.
+- Windows provider cleanup rechecks termination after process close and accepts an already-absent process as successfully cleaned up, avoiding stale ownership records and unnecessary watchdog waits.
+- Successful stale-loop cleanup removes obsolete runtime status, so `yoke loop status` no longer reports a dead run as `RUNNING`.
+
+## 1.6.1 — 2026-08-21
+
+### Fixed
+- Release metadata counts platform-conditional tests consistently on Windows and Ubuntu.
+- Provider cleanup confirms termination after child close, preventing stale ownership records when Windows reports process exit asynchronously.
+
+## 1.6.0 — 2026-08-20
+
+### Added
+- The canon now includes `no-ai-slop`, `domain-modeling`, `codebase-design`, `resolving-merge-conflicts`, and `writing-for-agents`, with their supporting templates and evaluation material. Source adaptations are credited in `canon/skills/ATTRIBUTION.md`.
+- UI projects receive an automatic design-quality gate with configurable `design.mode` (`off`, `auto`, or `on`) and score budget. Detection uses package dependencies, UI source files, and configured smoke flows.
+- Durable context now includes a project glossary and can expose an optional bounded-context map.
+
+### Changed
+- Retrofit installs complete skill packages for Claude, Codex, and Gemini instead of copying only `SKILL.md`. Local resources, binary files, and executable bits are preserved where supported.
+- Every canon skill declares `invocation: auto|manual`; retrofit maps that policy to Claude frontmatter, Codex `agents/openai.yaml`, and Gemini's generated automatic-skill index.
+- Canon validation rejects missing local Markdown resources, unsafe package entries, symlinks, and provider metadata that conflicts with the manifest.
+
+### Fixed
+- Windows provider cleanup now treats a successful `taskkill` as a request and confirms that the recorded PID has stopped before deleting its ownership record.
+
+## 1.5.1 — 2026-08-17
+
+### Fixed
+- Serena MCP configurations generated by Yoke no longer open the local web dashboard automatically on every startup.
+
+## 1.5.0 — 2026-08-16
+
+### Added
+- Failed verify, executable-criterion, performance, configured custom-audit, and completion gates now produce deterministic byte-bounded previews and preserve large complete stdout/stderr in content-addressed `.yoke/artifacts/` files with SHA-256 references.
+- Projects can tune `output.previewBytes` and `output.artifactThresholdBytes`; existing projects use backward-compatible 2 KiB/8 KiB defaults.
+- A deterministic local benchmark verifies signal retention, preview bounds, compression measurement, and artifact digest round-trips without making provider-token claims.
+
+### Security
+- Output artifact paths sanitize story identifiers, stay below a project-local root, use user-only file modes where supported, and are excluded from Yoke's clean-tree and story-commit operations even in upgraded projects. Raw artifacts are never injected automatically and documentation warns that project commands may emit secrets or personal data.
+- Gate command capture is capped at 16 MiB per stdout/stderr stream. Quota overflow fails closed and labels retained evidence as truncated instead of risking unbounded memory or claiming partial output is complete.
+
+## 1.4.0 — 2026-08-15
+
+### Added
+- `yoke loop run --parallel=N` now executes dependency-ready, non-colliding stories through real provider subprocess workers, isolated worktrees, leased claims, and a FIFO integration queue with fresh integrated-system gates.
+- Reference-driven quality declarations can collect screenshots, files, command output, or benchmark results and run a schema-validated blind critic with bounded repair rounds, elapsed-time limits, blocking or advisory policy, and retained proof.
+- `--candidates=N` can fan out up to five isolated implementations, discard mechanically red candidates, select one green candidate through identity-blind pairwise comparison, and preserve selected/loser evidence before cleanup.
+- Loop status now exposes dispatcher, worker, integrator, candidate lifecycle, worktree, queue, integration, reopen, quality-round, repair-budget, and trusted provider/model provenance data.
+
+### Changed
+- Provider subprocesses use explicit lifecycle contracts and incarnation-aware process records so worker cancellation and cleanup target only the process tree Yoke actually started.
+- Parallel and candidate runs disable adaptive routing, honor story-level provider affinity, latch pause requests across the whole dispatcher, and rerun quality plus review after integration.
+- `yoke loop cleanup` retains Yoke worktrees unless `--remove-worktrees` is explicit, while still reaping recorded orphan runners and stale locks safely.
+
+### Fixed
+- Expired claims, worker crashes, merge conflicts, pause races, and integration failures now release ownership deterministically, retain terminal proof, and reopen stories without leaking worktrees or marking false completion.
+- Quality repair fails closed on malformed critic output, reference drift, provider/model provenance mismatch, candidate identity leakage, unavailable critics, exhausted limits, and mechanically red repairs.
+- The watchdog resolves its TypeScript loader from both source and built npm layouts on Node 20+, and read-only Codex comparisons can run in disposable candidate worktrees without weakening normal repository checks.
+
+### Security
+- Blind comparison requests expose only opaque labels and digests while binding every verdict to the trusted judge provider, model, prompt, rubric, reference, and candidate provenance.
+- Cleanup and cancellation use project-scoped leases, owner tokens, PID birth/incarnation checks, and recorded process handles rather than machine-wide process-name matching.
+
+## 1.3.0 — 2026-08-09
+
+### Added
+- Teams can submit change requests at any time with `yoke change add` and inspect the append-only inbox with `yoke change status`; queued requests are planned at safe loop boundaries by Claude, Codex, or Gemini without interrupting active story work.
+- Acceptance criteria can carry stable IDs and executable verification commands, and Yoke records per-criterion evidence before a story can be marked complete.
+- Projects can configure an integrated completion command that proves the final system flow after all story-level gates pass.
+
+### Changed
+- New projects use strict criterion-proof requirements by default, while existing boolean-only criteria remain readable for compatibility.
+- PRD authoring, schemas, loop guidance, generated configuration, and provider instructions now treat completion as an ephemeral readiness state: new requests create more stories instead of prematurely forcing a release boundary.
+
+### Fixed
+- Broad green test suites can no longer satisfy unrelated structured acceptance criteria: proof commands must use an approved test runner, contain the criterion ID, and avoid shell operators.
+- Change-intake failures block safely; an independent coverage pass rejects omitted requested outcomes, crash recovery retains uncommitted requests, and concurrent PRD edits are detected instead of silently overwriting user work.
+- Integrated completion checks prevent locally finished stories from masking broken cross-component flows such as authentication callbacks or purchase-to-entitlement activation.
+
+### Security
+- Updated the transitive development dependency `nanoid` to a non-vulnerable release so the complete CI dependency audit is clean.
+- Restricted model-authored criterion proof to criterion-targeted test commands and blocked shell control operators and runner-prefix spoofing before host execution.
+
+## 1.2.1 — 2026-08-02
+
+### Fixed
+- `yoke loop run` now executes every remaining planned story by default instead of stopping after an implicit 25-iteration batch. Use `--max=N` only when an intentional bounded batch is wanted.
+- Unlimited runs remain unlimited after a critical-decision answer/resume cycle, while explicit caps remain preserved and accept positive integers only.
+
+## 1.2.0 — 2026-08-02
+
+### Added
+- Opt-in adaptive model routing lets a strong parent orchestrate each bounded story while Yoke selects an available Claude, Codex, or Gemini worker by quality, speed, cost, or balanced strategy.
+- A concurrency-safe local evidence registry learns from independent verification results without sharing mutable state between simultaneous Yoke processes.
+- Routing telemetry, reproducible benchmark fixtures, and analysis tooling make worker selection, token use, timing, and gate outcomes auditable.
+
+### Changed
+- Setup keeps routing disabled unless explicitly enabled and validates provider/model worker pools before execution.
+- Internal provider contract tests cover Claude Code, Codex CLI, and Gemini CLI through the shared adapter. Codex-only authenticated trials completed all 12 stories and 36 hidden checks; median routed runs used 11.0% fewer fresh input tokens, 49.5% fewer output tokens, 78.2% fewer reasoning tokens, and 33.8% less wall time than routing off.
+
+### Fixed
+- Reviewer prompts now permit their required verdict file while continuing to forbid project changes.
+- Failed reviewer subprocesses retain bounded provider stderr in loop status, exposing authentication, quota, sandbox, and startup failures instead of a generic command error.
+
+## 1.1.0 — 2026-07-30
+
+### Added
+- Shared five-question `yoke setup` wizard with provider-aware defaults for Claude, Codex, and Gemini.
+- Persisted default runner selection and `auto|critical` loop decision policies.
+- Provider-neutral `yoke-workflow` skill for planning questions, approved-plan PRD handoff, autonomous story execution, and critical-decision resume.
+- Structured critical-decision requests with `yoke loop decision` and `yoke loop answer`; answers are validated, committed under the configured human identity, and resume the same story.
+- Approved `.yoke/plan.md` context in PRD drafting and a lint gate for unresolved planning placeholders.
+
+### Fixed
+- Retrofit and loop on/off now preserve timeout, decision, runner, and permission settings.
+- Empty projects prefer the active agent host instead of silently installing Claude artifacts in Codex.
+- Loop and PRD runner selection now prefer an explicit flag, then the configured runner, then the active host.
+- Retrofit reports no longer label every provider as Claude Code.
+- Critical-decision resumes retain isolation, review, runner, permissions, timeout, JSON, policy, and iteration settings instead of falling back to an unreviewed default run.
+- Decision answers use an atomic owner-token lock and recoverable request journal, are checked against the active PRD story, bounded as untrusted data, and committed path-by-path so unrelated edits cannot enter the human-owned commit.
+- Decision recovery now binds the exact selected answer to its commit, rolls back only its own interrupted context append, namespaces resume state per project/worktree, and serializes cleanup with loop startup.
+- Active agent session markers now outrank globally configured provider home directories, and setup rejects partially invalid agent lists.
+
+### Changed
+- New setups enable the loop by default and choose `decisionPolicy: auto`; the wizard can select `critical` or disable the loop.
+- Legacy `loop.onAmbiguity` and `--on-ambiguity` remain compatibility aliases.
+
+## 1.0.0 — 2026-07-27
+
+### Added
+- Native Codex skills, project config, hooks, reusable agents, and plugin metadata.
+- Safe provider permission profiles and structured cross-provider telemetry.
+- Schema-validated independent review verdicts with explicit self-review opt-in.
+- Human-owned commit identity enforcement; AI co-author trailers default off.
+- `yoke audit` dependency, secret, and sensitive-diff gate with versioned suppressions.
+- PRD dependency graphs, collision areas, agent affinity, claims, FIFO merge queue, and bounded async dispatcher APIs.
+- Reproducible cross-runner benchmark schema and matrix launcher.
+
+### Changed
+- Dangerous permission bypass is opt-in via `--unsafe`.
+- Worktree cleanup is non-destructive unless `--remove-worktrees` is passed.
+- Reviews no longer trust process exit code alone.
+
+### Security
+- Vitest upgraded to 4.1.10; the dependency tree reports zero known vulnerabilities.
+
+
+## 0.9.0 — 2026-07-22
+
+### Added
+- **Performance budget gate** (`perf.command` in `.yoke/config.yaml`, optional `perf.retries`).
+  A benchmark command with the same contract as verify (exit 0 = within budget) runs **after
+  verify** on every story — new loop phase `perf`, `YOKE_STORY` exposed, worktree-aware in
+  `--isolate` mode. A red benchmark blocks the story
+  (`story S6 exceeded its performance budget: …`) no matter how clean the diff is. When the
+  gate is configured, the implementer prompt names the budget command so agents keep hot
+  paths efficient and never "simplify away" an optimization without re-running the benchmark.
+- **`performance` canon skill** (28 skills now): efficiency as a measured requirement —
+  clean-by-default with the decision ladder (minimal-code → measurable acceptance criterion →
+  project perf gate), profile-first, optimize leaves not boundaries, benchmarks committed as
+  tests, the *why* of every optimization versioned so future agents don't clean fast code
+  back to slow.
+- **`authoring-prd` guidance**: performance requirements belong in acceptance criteria as
+  numbers ("imports 1M rows in < 2s"), and every clarifying question belongs in the planning
+  round — a criterion still needing a decision is not loop-ready.
+
+## 0.8.0 — 2026-07-20
+
+### Added
+- **Live progress + ETA.** Story completions are now first-class events: the console shows
+  `✓ S6 done in 4m28s — 20/45 (44%) · ~1h40m left`, every status (file, NDJSON stream,
+  `yoke loop status`) carries `percent` and an `eta` block. The estimate averages the
+  durations of stories completed **in this run** (current velocity) and falls back to the
+  persisted history of previous runs (`.yoke/story-durations.json`, last 50, gitignored).
+  No data → no estimate, never an invented one.
+- **Ambiguity policy** (`loop.onAmbiguity` / `--on-ambiguity=<resolve|abort>`). The runner
+  prompt now always forbids asking questions (a loop run has nobody to answer). Default
+  `resolve`: the agent settles ambiguous criteria itself, states the interpretation, and the
+  loop never stops. Opt-in `abort`: the agent writes its open questions to
+  `.yoke/ambiguity.md` and stops; the loop consumes the file, skips verify (an unimplemented
+  story would otherwise pass on pre-existing green tests), and blocks with the question as
+  the reason. Companion principle: clarifying questions belong in the planning round, before
+  the loop starts.
+
+## 0.7.0 — 2026-07-17
+
+### Added
+- **Update check + `yoke upgrade`.** Every CLI invocation ends with a non-blocking
+  version hint (npm/gh-style): a detached background refresher caches the registry's
+  latest at most once a day; when it is newer, a one-line stderr hint suggests
+  `yoke upgrade` (which runs `npm install -g @hecer/yoke@latest`). Silent in CI,
+  `--json` runs, non-TTY pipes, and under `YOKE_NO_UPDATE_CHECK=1`.
+- **Opt-in auto-upgrade** (`update.auto: true` in `.yoke/config.yaml`): evaluated at
+  loop START only — never mid-run; the running process finishes on its version and
+  the upgrade applies from the next invocation. Deliberately NOT the default:
+  a gate harness must not change itself mid-project (determinism), and unreviewed
+  auto-installs are a supply-chain hazard.
+
+## 0.6.0 — 2026-07-17
+
+### Added
+- **Project-scoped orphan reaping.** The watchdog now records its pids in the project's
+  `.yoke/runner.pid` (main dir and per-story worktrees; removed on clean exit), and
+  `yoke loop cleanup` kills exactly those recorded process trees — and only while no
+  live loop holds the lock. Background: without a scoped mechanism, users and agents
+  resorted to machine-wide pattern kills (every process matching
+  `dangerously-skip-permissions`), which took down *healthy* runners of other projects
+  mid-story and stalled their loops. Never kill by pattern; `yoke loop cleanup` is the
+  safe path. `.yoke/runner.pid` is gitignored by retrofit.
+
+## 0.5.0 — 2026-07-17
+
+Root-cause fixes for the two "yoke keeps hanging" failure modes observed in the field
+(orphaned `claude.exe` runners piling up, healthy long stories dying at exactly the
+idle window):
+
+### Fixed
+- **Watchdog now kills the whole process tree on Windows** (`taskkill /T /F`).
+  Previously it killed only the spawned shell (`shell: true`), orphaning the actual
+  agent process — which kept writing to the worktree (dirty-tree blocks, failing
+  worktree removal) and kept burning API tokens. Observed in the field as ~10
+  zombie `claude.exe` per machine plus surviving dev servers.
+- **Claude runner always runs in stream-json mode.** Plain `-p` prints nothing until
+  the run finishes, so the idle watchdog mistook healthy >20-minute stories for dead
+  processes and killed them at exactly the idle timeout — while the user saw dead air.
+  The stream doubles as liveness; token usage is now reported on every run (not just
+  `--json` mode).
+
+### Changed
+- README: operating notes for driving the loop from inside an agent session
+  (background execution, small `--max` batches, `yoke loop cleanup` after interrupts) —
+  outer shell-tool timeouts killing a foreground `yoke loop run` were the third
+  observed "hang" pattern.
+
+## 0.4.0 — 2026-07-17
+
+### Added
+- **Hardened runner prompts** — distilled agent-harness patterns for headless runs:
+  scope discipline (nothing beyond the story), no unsolicited summary/plan/analysis
+  documents, root-cause fixes instead of gate bypasses, faithful outcome reporting,
+  bounded final messages (cuts output-token waste). Review prompts now ground verdicts
+  in observed evidence only and keep them brief.
+
+### Fixed
+- `.yoke/loop.pause` is now gitignored by retrofit. Previously the loop's own
+  `git add -A` story commit swept the pause control file into history in
+  un-retrofitted targets; removing it dirtied the tree and the clean-tree gate
+  blocked the resume run — the loop locked itself out.
+
+> Note: 0.3.0 was tagged and released on GitHub but never reached npm (2FA re-login
+> was pending), so for npm users 0.4.0 is the first release with the 0.3.0 changes below.
+
+## 0.3.0 — 2026-07-10
+
+### Added
+- **Claude Code plugin packaging** — the repo is now its own plugin marketplace
+  (`.claude-plugin/plugin.json` + `marketplace.json`): `/plugin marketplace add HECer/yoke`,
+  then `/plugin install yoke@yoke` installs the full canon under the `yoke:` skill namespace.
+- **Gemini CLI extension manifest** (`gemini-extension.json` + `GEMINI-EXTENSION.md`) —
+  installable via `gemini extensions install https://github.com/HECer/yoke`, listed in the
+  daily-crawled extensions gallery.
+- **Benchmark harness** (`bench/`) — reproducible cross-runner benchmark (tokens · speed ·
+  quality) with a fixed fixture, pre-written objective tests, and committed result data.
+- **Companion tool docs** — `canon/tools/claude-mem.md` (persistent memory; interactive
+  sessions only, explicitly kept out of loop runs) and `canon/tools/ui-ux-pro-max.md`
+  (design generation paired with Yoke's design verification gates).
+- **Multi-agent parallel loop design** — evaluation + phased design for distributing PRD
+  stories across parallel workers (`needs` dependency field, claim files, merge queue,
+  heterogeneous cross-agent dispatch): `docs/superpowers/specs/2026-07-10-multi-agent-parallel-loop-design.md`.
+
+### Fixed
+- Agent-availability probe timeout raised 5s → 20s: Gemini CLI cold-starts in ~6s on
+  Windows, so the loop misreported an installed `gemini` as "not found on PATH"
+  (found by the new benchmark harness).
+- Gemini runner invocation: dropped the bare `-p` flag — current Gemini CLI (0.33+)
+  requires a value after `-p` and errored with "Not enough arguments following: p".
+  Piped stdin selects headless mode by itself, so the runner now passes only `--yolo`
+  (also found by the benchmark harness).
+
+### Changed
+- README: npm install is now the primary quickstart path; documented plugin/extension
+  installs and optional companions.
+- npm package now ships `CHANGELOG.md`, `bench/` (harness + result data), and
+  `docs/superpowers/` (all specs and plans, including the multi-agent parallel loop design).
+
+## 0.2.0 — 2026-07-09
+
+- First npm release as `@hecer/yoke`.
+- Hyperflow integration surface: `yoke loop run --json` NDJSON stream, pause signal,
+  token-usage + model-id reporting for the claude runner.
+- `yoke new` greenfield bootstrap, `yoke prd draft`, cross-model `yoke review`,
+  `yoke flow-smoke` browser gate with proof artifacts, `yoke design-scan`.
+- Retrofit planners for Claude Code, Codex CLI, Gemini CLI; canon of 26 skills;
+  loop with worktree isolation, watchdog, single-flight lock, commit integrity.