Skip to content

CHANGELOG.md

Latest commit

 

History

History
283 lines (195 loc) · 35.1 KB

CHANGELOG.md

File metadata and controls

283 lines (195 loc) · 35.1 KB
Raw

Changelog

All notable changes to this project are documented in this file.

Format follows Keep a Changelog; versioning follows Semantic Versioning. Versions are recorded in git tags, this document, the README badge, docs/VERSIONS.md and the authoritative skills/flow-comet/INSTALLED_VERSION; bundle.yaml version stays 1.0.0 (decoupled from release versioning).

[1.5.0-rc.1] - 2026-08-23

Release candidate consolidating the mechanism and skill robustness batch: artifact template fidelity, skill-load declaration front gate, the two-layer skill-loading model, formal zero-commit semantics, and verify-exit sandbox compatibility, plus dsh bridge idle-state containment and plan wave-grouping validation. (#64) (#65) (#66) (#67)

Changed

  • dsh write containment now applies only while the flow is running: in the idle state (no state / no activeChange / completed) writes outside the project root are no longer intercepted by the bridge, matching Claude Code / Codex semantics; parse failures or unknown status stay fail-closed deny. (#65)
  • System test suite expanded to 61 items with dsh bridge flow-state-gate assertions (idle pass / parse-failure & unknown-status fail-closed). (#65)
  • Regression suite gained wave-grouping consistency validation and contract payload parse-failure detection scenarios. (#66)
  • Wave-grouping consistency check added at plan exit (interleaved [P]/serial mixing warns on legacy changes and blocks on new changes, with recovery guidance); record / workflow-handoff now fail closed on unparseable contract payloads (error with a --json-file hint, nothing written to state). (#66)
  • Artifact template fidelity and skill-load front gate: new changes must keep the template title, header fields, and section order for SUMMARY / TASK / CHANGE / REQUIREMENT / DESIGN; handoff requests and records now require the node's skill-load declaration, made after loading the skill with the Skill tool (reading the SKILL.md file is not loading). (#67)
  • Zero-commit task formal semantics: a task with an empty write_files list and an explicit noCommit contract skips the commit-subset check with an auditable prompt; a non-empty task claiming noCommit still runs the full check. (#67)
  • Two-layer skill-loading model guard scenarios added to the regression suite; two-layer skill-loading model across workflow skills: entry guidance now loads the routed node's implementation skill with the Skill tool before its required protocol calls; node skills state they are already loaded by routing (same-name required calls are declaration-only); the legacy auto-fill wording is qualified to legacy changes behind the declaration front gate; two regression scenarios pin declaration-command consistency and forbid self-loading and conflated phrasing. (#67)
  • Verify-exit command execution compatible with dsh headless sandboxed sessions: when a piped spawn is denied (EPERM, observed in dsh headless one-shot runner sessions), the guard retries the same command with inherited stdio — real execution and exit-code judgment unchanged, a VERIFY-DEGRADED line marks the degraded capture; non-EPERM failures never degrade. (#67)
  • Zero-commit results may not carry tracked commits: a result declaring noCommit whose commit contains tracked files is rejected on new changes (warn on legacy), closing an escape hatch around the commit-subset boundary; intake documents enforce the Change ID header on new changes (previously warn-only). Regression suite expanded to 171 scenarios. (#67)

[1.4.2] - 2026-08-18

DeepSeek Harness (dsh) platform support with bridge-based write enforcement, expanded regression coverage (144 scenarios / 60 system tests), and dsh write-guard hardening. (#60)

Added

  • DeepSeek Harness (dsh) platform support: flow-comet now installs on dsh through the installer — prepare-env --platform dsh installs the skill tree project-locally at <project>/.dsh/skills/flow-comet (dsh auto-discovers skills there at rank 100 without a restart, and projects without that directory cannot see the skill — naturally project-level), injects the orchestration rules into an AGENTS.md managed block (non-destructive merge), and mounts a thin bridge loader globally at $DSH_HOME/plugins/dsh-flow-comet-bridge.mjs with a managed block in $DSH_HOME/cordis.patch.yml (read-merge-write, preserves existing blocks such as dsh-skin, effective for all profiles). The bridge listens on the native tools/pre-execute waterfall event, maps dsh tool calls to the project-local guard decision script (allow / deny with BLOCK message and recovery guidance / fail-closed on shape mismatch or abnormal exit, Windows 8.3 short paths normalized), and only engages when the session's project root contains .dsh/skills/flow-comet (narrow listening — non flow-comet projects are untouched). Platform selection supports interactive multi-select (arrow keys, pre-checked from project traces) and explicit comma-separated platforms (--platform dsh / claude-code,dsh / all; the previous both option is removed). Version anchor: dsh 0.1.0-rc.6. (#PR)

Changed

  • Regression suite expanded to 144 scenarios with installer platform-selection chains (explicit single platform, comma-separated multi-platform, all, TTY multi-select with trace pre-checking, trace detection, unknown-platform errors, both rejection, idempotent re-install); the dsh adaptation keeps the engine untouched — the guard decision core is reused unchanged via subprocess calls. (#PR)
  • System test suite expanded from 55 to 60 items with dsh platform assertions (project-level skill install with path replacement and version marker, AGENTS.md managed block with user-content preservation, bridge loader syntax and home patch injection with read-merge-write, purge recovery). (#PR)

Fixed

  • Write containment on dsh: writes whose target resolves outside the project root are now denied before reaching the guard decision (previously a non-resolved target skipped the whitelist judgment — fail-open); Windows 8.3 short paths are normalized before judgment. (#PR)
  • Fail-closed on dsh tool arguments: argument shape mismatches (missing file_path / command, non-object arguments) now produce a deny with WARN instead of being silently passed through. (#PR)
  • Subagent execution on dsh: when a workflow delegates tasks to subagents, delegated agents (identified by the session's subagent delegation depth) can write source code as the executor — matching the worktree-isolation semantics of the other platforms — while the coordinating agent remains subject to the phase write whitelist. Out-of-project writes and malformed tool arguments stay denied for both. (#PR)
  • 8.3 path normalization before the guard decision on dsh: the project root is canonicalized to its long path form before the project-local guard is invoked, closing a fail-open where an 8.3 short-form project root made the guard's lexical path resolution skip the phase write whitelist (out-of-whitelist coordinator writes were allowed). (#PR)

[1.4.1] - 2026-08-16

Installer platform-selection fix and release-model change (release PRs now merge). (#56, #57)

Fixed

  • Installer platform selection: the interactive prompt now offers 3) both (install to both Claude Code and Codex — an interactive-only option; --platform still accepts only claude-code or codex); no-TTY auto-detection no longer arbitrarily picks Codex when the target project has both .claude/ and .codex/ traces — it defaults to Claude Code (primary) with a hint; Codex remains reachable via --platform codex, and both via an interactive terminal.

Changed

  • Release model: release PRs (dev → main) now merge instead of squash — dev's change-level commits enter main permanently, and dev stops accumulating a lead over main after each release.

[1.4.0] - 2026-08-14

Multi-platform installer framework, platform modularization, and real-artifact examples. (#45, #46, #47, #48, #50, #52)

Added

  • skill-load declaration mechanism: subagents declare loaded workflow skills per node; record validates declarations against the protocol's required skill calls; exit checks protocol declaration markers; cross-consistency timestamp checks; backward compatible with legacy changes.
  • Installed version marker: prepare-env writes <project>/.claude/skills/flow-comet/INSTALLED_VERSION from the source repo's git state (1.4.0 on a release tag; 1.4.0-N-g<hash> on accumulated dev) so issues and PRs can state the exact version — including how far dev has accumulated since the last release.
  • Local commit/push hooks: install-commit-hook.mjs sets up commit-msg and pre-push hooks that reject messages carrying process codes (this project's own convention, not a universal list).
  • Multi-platform installer framework: prepare-env now targets Claude Code (default, unchanged) or Codex via a platform-descriptor table — interactive platform selection (TTY) with an explicit --platform override and automatic detection of an existing .claude/ / .codex/; skills install to each platform's native location (.agents/skills/ for Codex, auto-discovered); SKILL/GUIDANCE command paths are rewritten at install time for non-default platforms (authoritative source stays in .claude form); Codex rules are injected into an AGENTS.md managed block (Codex's rules/ directory serves command-approval policies, not instruction files); the write-guard hook gains full Codex adaptation — Codex PreToolUse intercepts Bash tool calls, the hook parses write targets from the command (PowerShell cmdlets, .NET File API, redirection) and denies out-of-scope writes via {"decision":"block"} (measured on Codex CLI 0.146.0; trust the hook on first use via /hooks), while the Claude Code output stays unchanged.

Changed

  • Regression suite expanded to 137 scenarios: init rejects unknown flag-like arguments (e.g. --help) instead of treating them as a change name; covering skill-load declarations, record validation, exit protocol checks, cross-consistency timestamps, legacy compatibility, review finding disposition, artifact completeness, delegation attribution, recovery guidance, wave-wording consistency, self-check-method section enforcement for new changes, and per-change verification-failure isolation.
  • CI: process-code checks moved from the server-side PR policy to local hooks; PR/issue templates reworked for practice (deduplicated checkboxes, related-issue section, based-on version, protocol and installed-version fields).
  • Commit history made jargon-free: 52 historical commit messages rewritten to plain descriptions (tree unchanged); duplicate commits deduplicated.
  • Docs: README reorganized (quick start moved up) with recognizable anchors (GSD link, pain-point intro, fit boundary); installation guide gained an uninstall section; release checklist deduplicated to a single source; terminology unified across docs.
  • System test suite expanded to 55 items (installer version-marker check, multi-platform installer scenarios: Codex install smoke, hook platform contract, platform selection chain, purge semantics, platform-descriptor-driven install smoke, per-change verification-failure isolation).
  • Merge gate changed to CI status checks: branch protection no longer requires an approving review (single-account repo cannot self-approve); required checks are the CI jobs; bot reviewers (CodeRabbit / Sourcery) are advisory — contributing guide gains a bot-reviewers section (advisory-only, threaded replies, resolve before merge).
  • Examples rebuilt from a real archived run: docs/examples/ now carries the complete artifact set of a real 8-node change (processor-pipeline, run in the e2e fake project) — six-section summaries, review findings with disposition markers, skill-load declaration markers, actually-executed verify; the simulated example and outdated artifact screenshots were removed, and the README showcase now points at the real artifacts.
  • Documentation overhaul: skill instructions deduplicated (generator-template remnants and mid-file frontmatter removed), Comet-positioning claims replaced with flow-comet's own mechanism descriptions, per-node guardrail tables aligned with the actual guard implementation (unimplemented items now marked as review-checked execution discipline), dual-platform (Claude Code / Codex) adaptation for brooks-lint invocation, user entries, and installation docs, the regression baseline promoted to the two-tier suite (guard self-test + system test) across all docs, and timeliness updates (roadmap state, design-doc backfill, archived handover notes).
  • Strict mode for new changes: changes created via init are marked new (newChange) and enforce all content-level checks as blocking (disposition markers, built-in self-check evidence, wave-wording consistency, overreach delegation, append placement, entry evidence, self-check-method section); legacy changes keep the progressive warnings.
  • Execution-omission protection: node entry is now recorded (exit blocks on new changes when a node was exited without entering it); completed tasks require their summary for new changes (blocked — legacy changes keep the progressive warning); handoff results require TDD RED evidence for new changes; record auto-fills skill-load declaration markers; execute gains an explicit empty-exit exemption; init detects commit-less repositories; the installer cleans up stale empty hook groups when the hook matcher evolves.

Fixed

  • Artifact-path derivation respects protocol pathBase (custom protocols with project-root artifacts now advance correctly); fail-fast for unsupported roots.
  • Done tasks require matching per-task summaries (progressive warning, not block).
  • Overreach detection for delegated parallel tasks (execute and verify exits).
  • Recovery guidance added to blocked messages (advance / select / record).
  • Wave-wording consistency: prose marking a task parallel without the matching task attribute warns progressively.
  • init command rejects arguments starting with -- (e.g. --help) — previously treated as the change name, which auto-created a change, a branch, and state.
  • dev-main sync check: files deliberately deleted on dev (e.g. the template md → forms migration) no longer counted as drift.
  • Skill commands that lacked the install-path prefix now carry the authoritative-source path — executable on both platforms after installation.
  • Process-code detection regex now covers 1-3 digit scenario numbers (previously only two digits); the doc-scanner regex is back in sync with the single source; POSIX hook files get their executable bits set at install time.
  • CI bilingual-mirror check now covers SECURITY (CoC excluded by design, matching the local checker); the version-expected extraction no longer relies on a dead fallback.
  • Bundle metadata aligned: skills list, references, and script side effects match the actual distribution.
  • Mechanism docs and the skill guidance now state the strict-mode rules explicitly (entry enforcement, summary enforcement, and command-level write interception in the Red Flags).
  • Verification-failure counting is now isolated per change — switching changes no longer carries over another change's failure count (legacy states migrate automatically).
  • Archived changes now require the leftover-issues list (explicitly stating "no leftovers" when none); declaration-marker auto-fill no longer rebuilds an archived change's active directory.
  • The empty-repository branch hint now matches behavior — no branch is claimed when none can be created.
  • Review field-label exemption matches full labels only (a finding titled like "Source maps expose paths" is no longer mis-exempted from disposition validation).
  • --json-file reads are restricted to paths inside the project root (record and handoff).
  • The hook installer fails explicitly when POSIX executable bits cannot be set (previously silent — git would skip non-executable hooks).
  • Delegation writeFiles matching supports partial-segment globs (e.g. src/*.mjs) with anchored per-segment matching — * never crosses / and literal segments stay exact (previously a partial glob was compared literally and blocked valid delegations).
  • --json-file with a missing or empty value now reports a usage error instead of an internal type error (record and handoff).
  • Stale scenario-count references (136) in the suite headers and the test-matrix / known-issues templates are synced to 137.
  • PowerShell output-encoding guidance uses the explicit [System.Text.Encoding]::UTF8 expression (bilingual).
  • The archive-stage hook whitelist includes .specs/<change-id>/ so the leftover-issues list can be written into the change directory before archiving (previously blocked).
  • Archive exit warns when the project CHANGELOG has no entry for the change being archived (progressive, both new and legacy changes).

[1.3.1] - 2026-08-11

Documentation and CI maintenance release — no behavior changes. (#31, #33, #34, #35, #36, #37)

Added

  • README run-demo section with real screenshots (bilingual); flow-kit introduction and comparison with alternatives (bilingual).
  • TROUBLESHOOTING entries for the 1.3.0 initialization messages (INIT-GENERATE / VALIDATE-FAILED / DONE, bilingual).

Changed

  • CI: dev-main sync check is now tree-level and one-directional (only files main added count as dev behind — fixes false positives after squashed releases); branch-prefix allowlist gains ci/; commit-message convention check applies to development PRs only (release PRs carry accumulated history by design); actionlint upgrade with weekly canary job.
  • CHANGELOG scenario count corrected 95 → 97 to match the implementation; bilingual wording cleanups across CONTRIBUTING / INSTALLATION / USAGE.
  • Regression suite self-check list now includes CHANGELOG (prevents stale scenario counts in entries); regression mapping comment corrected to the actual count.

[1.3.0] - 2026-08-10

Automatic project-context initialization (init pre-step). (#30, #32)

Added

  • Automatic initialization detection: on first use in a project, the workflow automatically detects whether a project context (CONTEXT.md) exists and prompts to initialize it when missing — existing AI-context documents (such as CLAUDE.md / AGENTS.md) are read and integrated with source attribution (existing files are never modified); projects with a fresh context run silently. No separate command to remember.
  • Agent-assisted generation protocol: init <id> --init-context is now a collaboration — the script performs deterministic detection, decision, prompting, and validation, while the agent reads existing documents and probes the codebase to generate a template-aligned CONTEXT.md (seven sections, source-attribution citations, accumulated glossary/decisions/defaults preserved); the script validates the seven sections plus key template formats (dated decision entries, metadata fields, glossary table) and records the scan timestamp only after validation passes. A re-run after generation completes the handoff.
  • Template-aware guidance: the generation prompt reports whether the flow-kit CONTEXT template is detected and validates section names against it (built-in fallback when the template is missing).
  • Scenario-count consistency check: the regression suite now fails when public docs' scenario count drifts from the actual suite size.
  • guard self-test suite expanded to 97 scenarios covering detection prompting, generation guidance, validation pass and fail, placeholder tolerance, and guidance wording.

Fixed

  • The freshness hint no longer shows "null days" when no scan record exists (legacy projects) — it points to the exact next action instead.
  • Refreshing project context preserves accumulated CONTEXT content (glossary / locked decisions / defaults) instead of overwriting it.
  • New-project skeleton CONTEXT (placeholder sections) passes validation instead of being rejected.
  • Re-running init with an existing change id warns before resetting progress (protection against accidental progress loss).
  • Agent-generated CONTEXT entries in the wrong format (e.g. backtick-wrapped dates) are caught by format validation with a precise rewrite hint.

[1.2.4] - 2026-08-09

Documentation and message cleanup — public-facing content uses plain descriptive language throughout.

Changed

  • Public-facing content cleanup: commit-message convention now requires plain descriptions (no codes, numbers, or jargon — commit messages, PR titles, and CHANGELOG follow the same standard); script comments, scenario names, and runtime warning messages cleaned of process codes; behavior unchanged (82/82 regression + real-run verification).
  • CHANGELOG entries now link their release PRs (Keep a Changelog practice — public traceability; v1.2.1~v1.2.3 entries backfilled).
  • Contributing guide gains the plain-language convention for public artifacts. (#14, #15)

[1.2.3] - 2026-08-09

Worktree delegation chain fixes (real-run report). (#12, #13, #14)

Fixed

  • route diagnostic: when TASK.md task tags carry parallel="true" without status="pending" (old template shape), the router now emits a ROUTE WARN explaining the required attribute order — structural validation stays strict, failures are visible instead of silently stalling in execute.
  • C4 check visible: the worktree dirty-check catch block now prints C4-CHECK SKIP: <reason> (previously silently swallowed — Windows git-command failures were invisible).
  • WARN COUNT summary: entry/exit output ends with a WARN COUNT: N summary line (appended; existing output unchanged).
  • Empty-exit documentation aligned with implementation: the execute SKILL no longer claims an "empty exit" path (guard blocks it in three layers: evidence → serial pending → output-schema artifacts); documented correct paths (direct routing to subagent-execute / explicit parallelTakeoverApproved exemption).
  • Pre-delegation checklist: subagent-execute SKILL now mandates a pre-delegation checklist (git status / HEAD / inline context) with a Red Flag; dirty-worktree protocol marked as mandatory before delegation.
  • guard self-test suite expanded to 82 scenarios (RED→GREEN with correct failure reasons, verified by an independent reviewer with 24 independently constructed assertions).

Changed

  • Troubleshooting gains new entries (ROUTE WARN / C4-CHECK SKIP / WARN COUNT, bilingual); scenario counts 77→82 synced across all public docs (historical changelog entries unchanged).

[1.2.2] - 2026-08-08

Self-review two-tier fallback for the 6-dimension check (brooks-lint). (#8, #9)

Fixed

  • Two-tier fallback for brooks-lint: when the Skill tool returns only a "Launching skill" placeholder (worktree subagent skill routing is unstable — plugin content may not be injected), the subagent now Reads the plugin-cache protocol files and executes the full brooks review manually (selfReview: cache-brooks) before falling back to the built-in R1~R6 quick check. Review quality is preserved even when loading fails.
  • Guard validates fallback evidence: a builtin-quickcheck declaration must now state the unavailability reason and the cache-attempt evidence (missing either → progressive BROOKS-LINT WARN, never BLOCK).
  • Guard recognizes cache-brooks: method/whole-content/six-dimension patterns updated — a clean cache-brooks declaration passes without WARN (previously mis-blocked).
  • guard self-test suite expanded to 77 scenarios (builtin evidence checks + cache-brooks acceptance; RED→GREEN with correct failure reasons, verified by two independent reviewers).

Changed

  • Troubleshooting gains the new WARN entry (bilingual); public MECHANISM (bilingual) documents the two-tier fallback behavior layer; scenario counts 74→77 synced across all public docs (README/INSTALLATION/MECHANISM/VERSIONS/CONTRIBUTING — historical changelog entries unchanged).

[1.2.1] - 2026-08-08

Installation-guide fixes (init/hook/state fixes + README guidance fixes + independently verified fixes). (#1, #4, #6, #10)

Fixed

  • init state gains status: 'running' + three-tier hook semantics : init previously did not write status, and the hook allowed writes on undefined-status state — the first defense layer was ineffective during open (before the first guard exit); archived completed states were also blocked from all writes. After: running (including legacy states without status but with activeChange — fail-closed backward compatible) → whitelist validation; completed → allowed
  • init creates .specs/<id>/ directory : previously next/status reported "No active change. Run: init" after init (findActiveChange requires the directory), contradicting the SKILL startup protocol
  • findActiveChange skips fallback scan on archived-complete state : leftover archive directories (with TASK.md copies) are no longer misdetected as active changes
  • init currentNode follows the protocol's first node : no longer hardcodes open when a custom protocol starts elsewhere
  • record strips --protocol from payload : stripped before JSON parsing, preventing structural fields from being lost on parse failure
  • Custom-node coordinator-default whitelist : undeclared non-built-in nodes default to ['.specs/'] (writing source code requires explicit declaration) — closes the fail-open gap
  • Legacy state without activeChange is allowed : legacy states (no status + no activeChange) are no longer fully blocked by the hook
  • writeWhitelist supports the <change-id> placeholder : protocols reuse across changes with automatic adaptation
  • init output NODE follows the protocol's first node : consistent with state.currentNode (previously hardcoded open)
  • findActiveChange checks completed first : archived-complete states with a residual activeChange are no longer misdetected
  • hook statePath falls back to default : minimal-schema protocols (no state.statePath) no longer crash
  • Three scripts tolerate UTF-8 BOM in JSON.parse : state/evidence files written by external tools (e.g. session Write) with BOM parse normally
  • guard self-test suite expanded to 74 scenarios (full positive/negative cases + independent-verification scenarios + BOM tolerance)

Changed

  • README installation verification now has four steps (structure / config loadability / authoritative-source diff consistency / real-environment smoke test) — guard-self-test marked as the author regression baseline (script-logic self-test; does not depend on installation completeness; not an installation verification criterion)
  • README Quick Start gains the flow-kit prerequisite hint and a new-session hint; Requirements gains a flow-kit verification step; settings injection documents first-time vs existing-file cases; hook command documents the project-root resolution base
  • SKILL startup protocol documentation corrected : init output is the first route (NODE: open / protocol first node); next is used after a node exits
  • README restructured into multi-document layout (bilingual): README index + docs/ (INSTALLATION/USAGE/PROTOCOL/MECHANISM/TROUBLESHOOTING/VERSIONS) + compose example; VERIFICATION.md removed from the public repo (verification records are out of scope for the public repository)
  • hook blocking semantics confirmed by measurement: main TUI session blocks writes (physical interception); claude -p (SDK CLI mode) downgrades to non-blocking (log-only)

[1.2.0] - 2026-08-08

Historical version from before the PR workflow existed — no PR link available.

Custom skill composition (flow-comet-compose) + protocol parameterization + non-destructive installer.

Added

  • flow-comet-compose guidance skill: interactively composes any installed skill (superpowers/brooks-lint/custom) into a custom protocol JSON, driven by the same engine (state routing + guard validation + hook interception); does not replace the built-in 8-node protocol
  • Protocol parameterization: resolveProtocol priority (--protocol CLI → FLOW_COMET_PROTOCOL env → default reference/workflow-protocol.json); determineNode data-driven — node declarations come from the protocol JSON (ADR-001)
  • Specialization validation bound by node id: the general defense layer is protocol-agnostic (all protocols physically validated); specialization fires only for built-in node ids, never for custom ids (ADR-002)
  • writeWhitelist declaration: protocols may declare hook write whitelists; parse failures fall back to the built-in table (fail-closed, no whitelist gaps)
  • prepare-env installer: generates/overwrites the target project's .claude/ (rules + skills + settings injection) from .comet/bundle-drafts/flow-comet/; settings use read-merge-write idempotent injection (preserving permissions etc.); --purge --yes for explicit destructive rebuild
  • Configurable branch prefix: --branch-prefix customizes the change branch prefix (default change/)
  • guard self-test suite expanded to 54 scenarios (custom protocols / composition / node-element validation / exemption tightening / hook-declaration fail-closed / branch-prefix positive-negative)

Changed

  • Single authoritative source: removed the .comet/bundles/ dual directory (comet needs eval/publish distribution; flow-comet is copy-install only — single source bundle-drafts/); README installation now uses prepare-env (option A) + manual copy (option B), comet bundle distribution flow removed
  • prepare-env non-destructive: overwrites only generated files (rules/skills) + settings injection by default; no longer unconditionally deletes the target .claude/

Fixed

  • record overwriting handoff triggers takeover BLOCK : record that replaces an evidence key wholesale is BLOCKED
  • Node ordering/completion markers: completedChecks validation, skipped-node BLOCK
  • redEvidence ordering : redEvidence must precede greenEvidence
  • next false-block fix : exempt-node mis-block corrected
  • hook declaration fail-closed: protocol hook-declaration parse failures fall back to the built-in table (audit supplement)

[1.1.0] - 2026-08-05

Historical version from before the PR workflow existed — no PR link available.

Change branches + PR review + append-placement discipline + documentation rewrite.

Added

  • Change branch mode: init auto-creates the change/<id> branch; the whole flow runs on it; archive wraps up with merge + branch deletion
  • PR review: config set enablePrReview true; push branch + create PR before archiving, merge after approve
  • Branch-state consistency check: status/next detect branch/activeChange mismatch → WARN (not BLOCK)
  • Append-placement discipline and structural detection: CONTEXT terms/decisions inserted into existing sections, LESSONS inserted by L-NNN into the entries section, STATE decision log inserted at top (reverse order), CHANGELOG table inserted at top, rollback fixes appended to ## Fix 任务; guard detects (orphan sections / numbering disorder / outside entries / non-reverse order) as progressive WARN
  • guard self-test suite expanded to 23 scenarios (branch checks + append-placement positive-negative)

Changed

  • README fully rewritten as a complete product document (ecosystem / quick start / workflow overview / artifacts / core mechanisms / design principles / troubleshooting / versions)
  • state-schema gains branchMode / enablePrReview field validation (boolean)

Fixed

  • LESSONS disorder detection split by section: multi-section numbering (## 活跃条目 / ## 已解决条目 independent numbering) no longer false-positives

[1.0.0] - 2026-08-04

Historical version from before the PR workflow existed — no PR link available.

First stable release (8-node workflow + three defense layers + guard validation, verified end-to-end in real projects).

Added

  • 8-node auto-routed workflow (open → design → plan → execute ⇄ subagent-execute → review → verify → archive)
  • Own state machine (.comet/flow-comet-state.json) + determineNode file derivation + auto-correction
  • Three defense layers (hook phase whitelist / coordinator prohibition / exit takeover detection)
  • Guard validation system: template-derived section names, SUMMARY six sections + mandatory self-check method, TASK signature hash, real verify execution, verifyFailures counting, state schema validation (fail-closed)
  • Subagent-based execution engine + executionMode (subagent default / direct escape hatch) + Return Contract + handoff hash provenance
  • Side commands (flow-comet-evolve / flow-comet-health) + express downgrade path
  • guard self-test suite (17 scenarios) + end-to-end real-project verification (new and existing projects)