CHANGELOG.md

Changelog

All notable changes to Forge are documented here.

[v1.51.0] - 2026-07-18

Added

• agents-cli Quality Flywheel (local). pnpm skill-eval compare|analyze (+ --judge) on run-history.json / judge-history.json; each run appends a snapshot. (scripts/skill-eval-flywheel.ts, scripts/skill-eval.ts)
• Global Shortcuts to Resist + 三类证据. core/skills/_shared/shortcuts-to-resist.md; bootstrap Iron Law 10 (re-read Skill at phase entry); quickref / skill-eval docs. Comparison: core/docs/agents-cli-comparison.md.

Changed

• Adapter sync catch-up: evolution-engine editable-surface / model-staleness prose, dev-planner generate-contracts.md, check-evolution.bat surface hint — aligned to core via pnpm sync.

Fixed

• Removed invalid claude --clear call from yolo-driver. claude CLI has no --clear flag (verified via claude --help); the call in scripts/yolo-driver.sh:78 and yolo-driver.bat:90 silently failed (2>/dev/null || true / 2>nul). Each claude -p invocation is already a fresh session (no --resume/--continue), so context is naturally clean — the call was redundant and misleading. Comments updated to explain why no explicit clear is needed.

Changed

• forge-verify: STALE evidence detection implemented (was README-only). evidenceFileMeta() 记录 mtime 已有，但从未与预读 mtime 比对。新增：(1) EG 管道前捕获预读 mtime；(2) 验证结束后重 stat 并比较，不一致时设 evidence_file.stale: true + 顶层 trace.stale: true。修复 pre-existing bug：trace 在 const 定义前被 early return 引用 → ReferenceError。README 更新描述。 (5dc6165)
• forge-verify: C2→C1 feedback loop 闭环. pattern_suggestions 此前仅输出到 stdout/trace，不持久化。新增：(1) 写入 .forge/verify-pattern-suggestions.json；(2) --apply-suggestions 标志将建议自动合并到 .forge/content-verify.json（去重、仅添加真正新的 pattern）。README 新增反馈环文档。 (5dc6165)
• forge-verify: 新增集成测试 + export layeredVerify. test-verify-integration.mjs（13 检查点）验证 trace evidence_files 结构、FILE_NOT_FOUND 路径、无 evidence_gates 路径。导出 layeredVerify 供测试 import。 (5dc6165+)
• package.json: 注册全部 forge-verify 测试脚本. 补充 test-evidence-gate、test-c2c1-feedback、test-dfv2-failure-class。新增 test-forge-verify-all 一键全跑。 (5dc6165+)
• dogfood-05-tracking.md: "实现落地" + Run #3 plan. Backfilled post-fc88a2d documentation: marked the YOLO forward-driver as implemented (2026-07-04 commit fc88a2d had shipped the driver + dev-builder Step 5a/6 changes but tracking doc still said "next: design the driver"), recorded static review findings (6 items, only claude --clear was a real bug), and wrote the full Run #3 execution checklist (rebuild target dir, run driver, 5-question observation table, decision branches). yolo-driver was retroactively attributed to v1.49 (fc88a2d predates 1af96ac by 5 days) — see v1.49 entry below.

[v1.50.0] - 2026-07-11

Added

• forge-verify LLM protocol-adaptive layer. L2 and C2 route through a unified llmComplete() that picks protocol from ANTHROPIC_BASE_URL — containing /anthropic → Anthropic Messages (e.g. Zhipu GLM's /api/anthropic gateway, the same path Claude Code uses), otherwise → OpenAI Chat Completions (e.g. deepseek). One ANTHROPIC_* env works across both endpoint families with no code change. Previously every call was hardcoded to OpenAI and stripped /anthropic, so under a GLM env every L2/C2 call hit HTTP 403. (bd0a912)

Fixed

• C2 honest degradation. C2 (per-requirement LLM) now records API/parse errors (non-2xx, empty body, network) as non-votes → UNCLEAR (C2_api_errors), never a false REJECT (execution-lapse). An unreachable model no longer reads as "the agent did sloppy work". L2 already behaved this way (API_ERROR/API_PARSE_ERROR non-votes → L3 UNCLEAR); C2 now matches. (ffff4cc)
• --runs/--task/--files honored under --from-config. The CLI override block sat inside parseConfig's else branch, so --from-config --runs 1 was silently ignored (always 3). Overrides now apply in both modes. (2ff784c)
• content-verify unloadable since v1.49.0. A duplicate const trace (SyntaxError) left by the chain-of-evidence commit meant content-verify.mjs could not parse or run at all. Removed. (98d8a36)

Changed

• Tests import production pipeline. test-evidence-gate.mjs imports evidenceGateCheck/contractRegexCheck/perRequirementLlmCheck from content-verify.mjs (now exported; main() guarded by import.meta.url) instead of inline copies that had drifted — the inline C2 was still on the old OpenAI-only protocol. −171 lines, zero drift. (cdcdaa7, 3157316)

[v1.49.0] - 2026-07-09

Added

• forge-verify four-layer pipeline (L0-L3). Refactored content-verify.mjs from ad-hoc checks into a layered pipeline: L0 shape (file exists, non-empty, non-placeholder, has test cases) → L0e Re-Stat (catches future-tense / stub / meta-comment "restating the requirement" instead of demonstrating it) → L1 contract regex (min/keywords) → L2 LLM thin audit → L3 disagreement escalation. Inspired by nexus-lab-zen design. (164c991, c1200ce)
• Evidence Gate pipeline (EG→C1→C2). Optional pipeline for projects with .forge/content-verify.json: evidence files exist + non-empty (EG, deterministic, zero-cost) → C1 contract regex per-requirement → C2 per-requirement LLM. Replaces the traditional L2 free-text audit when configured. (050cc35)
• content-quality check (cross-model semantic verification). New pnpm forge-verify check that runs the pipeline across Phase output files. Catches semantic garbage that symbolic checks (file-exists / exit-code) miss. (8279de1)
• YOLO forward-driver for cross-Phase continuation. External loop drives /dev-builder across all Phases without relying on skill prose. Dogfood #5 Run #1 + Run #2 both proved the prior design (agent reads "Continue to the next Phase automatically" prose and self-re-invokes) fails: in two independent sessions the agent stopped after Phase 1 and never re-invoked, regardless of context headroom. The Stop-hook machine-layer fix (see Fixed) only addressed erroneous stop-waiting; it does not drive continuation. New design: dev-builder YOLO Step 5a writes .forge/.yolo-continue JSON (completedPhase / nextPhase / timestamp) at Phase completion; scripts/yolo-driver.sh + .bat loops claude -p "/dev-builder" → read signal → delete → re-invoke. One Phase per invocation stays non-negotiable, even in YOLO mode. Not installed into user projects by forge-install — run from a Forge checkout. Scope: claude-code only (native Stop hook). Pending Run #3 validation in an independent session. (fc88a2d: scripts/yolo-driver.sh, scripts/yolo-driver.bat, core/skills/dev-builder/references/workflow.md § YOLO Mode, .forge/dogfood-05-tracking.md)
• CSS Sanity Check in dev-builder. Catches AI blind-spot layout bugs (z-index stacking, overflow clipping, missing display rules) that compile clean but render wrong. (861514f)
• Editable surface. .forge/editable-surface.json defines explicit read/write paths for evolution engine, excluding evaluator code (scripts/forge-verify/) and gate config from evolution scope. Inspired by DGM (Zhang et al. 2025): evaluator and permission control must live outside the loop. (b223577)
• failure_class on forge-verify verdicts. Each REJECT/UNCLEAR verdict carries failure_class (execution-lapse, skill-defect, unset), bridging verification pipeline to feedback-observer for automatic evolution triggering. (b223577)
• Held-in/held-out proposal validation. Evolution proposals carry held_in_files (target failures) and held_out_files (regression check). Both splits must PASS before proposal is finalized. Inspired by Self-Harness (Zhang et al. 2026). (b223577)
• Chain-of-evidence trace. Each forge-verify stage output includes evidence field referencing the source file or evidence file that produced the verdict. Final output includes trace.chain with per-stage evidence trail. Inspired by ScientistOne (Meng et al. 2026). (b888e19)
• Model staleness detection. Rules graduated under older model versions can be proposed for retirement when current model no longer needs them. Prevents circular evolution.
• Harness search space. .forge/harnesses/ directory with candidate registry and template for config-level crossover. Evolution-engine [Harness Search] section defines the pattern for generating and evaluating config variants. Inspired by Meta-Harness (Lee et al. 2026). (8813bb6)
• Tests: test-dfv2-failure-class.mjs (20 DF v2 scenarios × deterministic layers), test-evidence-gate.mjs (EG→C1→C2 pipeline, 12/12 pass).

Fixed

• Stop-time gates wired to Claude Code Stop lifecycle. phase-exit-guard, stop-gate, retry-gate were registered under PreToolUse, but their block files (phase-exit-block, .verify-block) are written later — so the hooks always exited 0 (no-op). Moved to the Stop array in adapters/claude-code/.claude/settings.json + settings.windows.json. detect-feedback-signal stays PreToolUse. phase-exit-guard.bat also gained .verify-block fallback (was phase-exit-block-only, drift from .sh) and a YOLO-detection fix (if not errorlevel 1 inside if (...) block, which had made YOLO forever read 0). New stop-gate-wired smoke asserts the wiring. Found via dogfood #5 Run #1. (60e6e2d, de70852)
• Cursor death loop (4 commits). Cursor's BeforeCommand hook chain triggered a feedback-signal / stop-gate infinite loop on every prompt. Cleared the chain (bc8549f), killed the loop in detect-feedback-signal (129838b), backfilled adapter drift from 60e6e2d (da83a58), closed dangling refs in install + Windows death-loop in context-compaction.bat (1d9419b, f7515e1).

[v1.48.6] - 2026-06-27

Added

• DESIGN.md token freeze (Google design.md). design-maker v1.1.0 merges Design-Brief direction with design-tool values into root DESIGN.md after mockup verification (Step 3e); templates + workflow in core/skills/design-maker/templates/design-md-template.md and references/design-md-freeze.md. Optional npx -p @google/design.md designmd lint DESIGN.md (Windows: designmd alias).
• dev-builder style priority: DESIGN.md > design tool MCP > Design-Brief.md when implementing UI.
• quickref design chain: core/templates/forge-quickref.md documents Brief → mockup → UI-Spec.md + DESIGN.md vs OpenSpec changes/<name>/design.md.
• code-review UI baseline: When root DESIGN.md exists, design reviewer cross-references frozen tokens and component map before Brief/MCP (same priority as dev-builder).
• Nature Gate size pre-check. Small Phases (≤3 key files + ≤5 deliverables) skip implementer isolation regardless of Nature — cold-start overhead exceeds actual coding time on small Backend/Data phases. Verified in Dogfood #3 (@forge/ecosystem-cache). (core/skills/dev-builder/references/workflow.md § Step 1.5, references/sub-agent-isolation.md, SKILL.md HARD-GATE)
• Phase-boundary detector (procedural version). implementer receives files_to_modify as a strict scope boundary; off-scope writes trigger DONE_WITH_CONCERNS. After implementer returns, dev-builder Step 8.5 cross-references file_changes against files_to_modify — violations escalate to user (A: merge, B: reject, C: replan). This is Overstepping Gate's first real enforcement. (core/agents/implementer.md, core/skills/dev-builder/references/workflow.md § Step 8.5, .forge/deferred-ideas.md)
• package-integrity forge-smoke. Validates package.json structure (fields, scripts targets, forge-hooks file references, bin entries, files whitelist). 39 checks, no network calls. (scripts/forge-smoke/package-integrity.mjs, forge-smoke now 14/14)

Fixed

• Machine Gates contract aligned with enforcement reality. Sloppiness Gate's actual enforcement is forge-verify → .verify-block → phase-exit-guard (stop-time), not the PreToolUse write chain. Overstepping Gate marked "declared, not yet machine-enforced". core/hooks/AGENTS.md false claim ("Plus Sloppiness/Overstepping gates in the same hook chain") removed. (CLAUDE.md, core/hooks/AGENTS.md, core/templates/forge-quickref.md, .forge/quickref.md)
• machine-gates-doc smoke upgraded from "assert names present" → "assert enforcement files exist" (25 checks). A future ghost gate now fails CI. (scripts/forge-smoke/machine-gates-doc.mjs)
• design-maker stall prevention. Per-step and per-page HARD-GATE checkpoints write to .forge/design-maker/ — token checkpoint (00-tokens-done.md), component checkpoint (01-components-done.md), per-page checkpoint (page-checkpoint.md). Prevents 15-20 minute silent stalls by forcing concrete output before proceeding. (core/skills/design-maker/SKILL.md [Design Phase] + [Gotchas])
• product-spec-builder size-aware default routing. Catch-all "Otherwise → 0-to-1" replaced with size detection: small products (CLI, ≤4 features, no auth, no DB) default to Quick mode; medium/large default to 0-to-1. (core/skills/product-spec-builder/references/startup-check.md)
• forge-install Windows auto-detection since v1.14.2 (repeated entry — already fixed in v1.48.4, included here for v1.48.6 branch consistency)

Changed

• README / README.zh-CN: v1.48.6 notes, Skill table, project layout, design priority, and workflow step 4 updated for DESIGN.md.
• deferred-ideas.md: Parked 4 entries with concrete trigger conditions (CI-watch, gate-prose-generation, phase-boundary-detector hook upgrade, double-gate→single). Phase-boundary-detector procedural version shipped; hook-level entry preserved.

[v1.48.5] - 2026-06-27

Added

• Per-finding action taxonomy (auto-fix / ask-user / no-op) in code-review. Borrowed from no-mistakes: every review finding now carries an action field answering "who fixes this", orthogonal to severity / risk_rank / the Must-fix·Should-fix·Insight buckets / Priority. The four specialists (code-reviewer-{bug,design,security,types}) assign it; the aggregator (code-reviewer) propagates + counts it unchanged. Routing in dev-builder Step 14.6: only auto-fix findings flow to bug-fixer/dev-builder; ask-user (intent / product-behavior / pre-existing dead-code / S5-aesthetic decisions) trips immediate escalation — sets .forge/.retry-counter.json state=escalated without incrementing retries (does not consume a retry round) and surfaces the existing A/B/C options; no-op is logged only. Reuses the existing retry-gate.sh — no hook or loop code added. New shared pointer core/skills/_shared/finding-actions.md + full text core/docs/finding-actions.md. Non-breaking: a missing action is treated as auto-fix (fail-open = prior behavior). (core/agents/code-reviewer*.md, core/skills/{code-review,bug-fixer,dev-builder}/)

[v1.48.4] - 2026-06-25

Fixed

• forge-install Windows auto-detection broken since v1.14.2. parseInstallArgs defaulted windows: false, but installForge used options.windows ?? process.platform === "win32" for fallback — since false is not undefined, the ?? never triggered and Windows users got Unix-style settings.json (.sh + bash) unless they passed --windows explicitly. All PreToolUse/PostToolUse hooks were silently dead in fresh Windows installs for ~1 month. Default is now process.platform === "win32" at the CLI parse layer. Found via dogfood #2 setup. (scripts/install.ts, scripts/__tests__/install.test.ts)

[v1.48.3] - 2026-06-19

Added

• Spec-Distillation Mode (issue #7): new optional mode in product-spec-builder — the spec-stage anti-sycophancy analog of critique-gate (post-spec) and plan-critique-check (planning). On a fuzzy one-liner, runs 4-path inference (Real Need / Competitor Approaches / Domain Patterns / Tech Feasibility) — not user questioning — cross-validates against what the user said, and outputs a ✅/⚠️/❓ Distillation Map that hands off to 0-to-1/Quick (❓ → first clarifying questions; ✅/⚠️ → pre-filled Spec). Keyword-gated (distill / 蒸馏 / infer what I need), so the default spec path is unchanged. Consumes domain-mapper outputs when present; greenfield-safe. (core/skills/product-spec-builder/references/distillation-mode.md, +5 wiring touchpoints)
• forge-spec-distill: pnpm forge-spec-distill <Distillation-Map> verifier — checks 4 inference paths present, ⚠️ findings carry §P/URL basis, ❓ section non-empty, density quota (≥3 sourced), 0-⚠️ = sycophantic. Non-zero exit on sycophantic/malformed. Reuses forge-spec-critique.mjs primitives. (scripts/forge-spec-distill.mjs)

[v1.48.2] - 2026-06-19

Fixed

• Documented forge-* commands now resolve. forge-scaffold, forge-coverage, and forge-skill-retrieve were advertised in README/CHANGELOG as pnpm forge-* but were absent from package.json scripts, so they failed with "Missing script". All three are now registered. (package.json)
• Two shallow relative links in skill references resolved one directory level too shallow (../../ where ../../../ was needed) and pointed at nonexistent paths: core/skills/product-spec-builder/references/workflow-0-to-1.md → core/docs/llm-council-comparison.md; core/skills/dev-planner/references/workflow.md → core/templates/github-issues-slices-template.md.
• DEV-PLAN.md intro phase count corrected from "Phase 1–13" to "Phase 1–16" to match the table body.
• docs/github-wiki/Home.md current-version bumped v1.44.0 → v1.48.1 and date refreshed to 2026-06-19.

Added

• Unit tests for sync.ts agent-transform logic (scripts/__tests__/sync.test.ts): adaptAgentContent (opus/sonnet/haiku → inherit for non-Claude adapters; claude-code preserves pinning; whitespace preserved; non-model/comment lines untouched) and AGENT_DIR_SKIP (excludes AGENTS.md, inherits SKIP_FILES), plus a syncDir integration test verifying both the skip and the transform wire-up. Suite 154 → 162.

Changed

• scripts/split-skill-references.mjs → scripts/archive/: one-off v1.20.4 tool (referenced only in CHANGELOG) archived rather than deleted, preserving the changelog cross-reference. Historical CHANGELOG entry left intact as a point-in-time record.
• CONTEXT.md rewritten from v1.35.7-era stale status to current verified state (v1.48.1 / 495cbed, audit reconciliation summary, test commands).

[v1.48.1] - 2026-06-19

Fixed

• Cross-platform agent dispatch (Mode A) fully delivered. The platform→Mode A/B mapping had overstated delivered state: Cursor/Gemini CLI were classed Mode B against pre-subagent versions, and the Cursor adapter shipped agents to the wrong path. Re-verified against current versions — all four target platforms (Claude Code, OpenCode, Cursor 2.4+, Gemini CLI v0.38.1+) support native isolated subagents. Cursor adapter now ships to .cursor/agents/; the 4 code-review reviewers gained cross-platform frontmatter (model: inherit). (core/skills/code-review/references/multi-perspective-dispatch.md, core/skills/dev-builder/references/sub-agent-isolation.md, core/agents/code-reviewer-*.md)
• Per-platform sub-agent model normalization. Claude-specific model aliases (opus/sonnet/haiku) are invalid on Cursor/Gemini/OpenCode and could fail validation or fall back silently. scripts/sync.ts (adaptAgentContent) normalizes them to inherit for non-Claude adapters while preserving the opus pinning on Claude Code. syncDir and fileHash are both transform-aware, so sync:discover stays 0-drift.
• AGENTS.md index no longer shipped into sub-agent scan dirs (.cursor/agents/, .gemini/agents/, ...) where it would surface as a bogus agent entry. New AGENT_DIR_SKIP. (scripts/sync.ts)

Changed

• code-review/SKILL.md review-checklist summary reworded to platform-neutral "4-dimension parallel review (Mode A dispatch)".

[v1.48.0] - 2026-06-18

Added

• Critique Gate hardening: Density quota (3 evidence-backed findings required, below → re-scan), Evidence schema (all three signal tables now require spec citation column), critique-of-critique mode (forge-spec-critique.mjs --critique detects fake-critique patterns). (core/skills/product-spec-builder/references/critique-gate.md, scripts/forge-spec-critique.mjs)
• Plan Critique Check: Anti-sycophancy gate for dev-planner — challenges Phase Order, MVP Scope, and Tech Stack decisions before writing DEV-PLAN.md. Generation Mode only. (core/skills/dev-planner/references/plan-critique-check.md)
• Product size detection: pnpm forge-size-detect reads Product-Spec.md and auto-recommends gate level — small products (CLI, ≤4 features, no auth/DB) get light, medium/large get full. Writes to .forge/gate-config.json with --write-gate-config. Runs after Spec confirmation in both 0-to-1 and Quick Mode. (scripts/forge-size-detect.mjs)
• TBD overflow gate: Hook blocks app code writes when Product-Spec.md has >5 [TBD]/[待确认] markers outside Idea Stage Exit Criteria. Excessive TBD markers indicate key decisions being deferred — a sycophancy pattern. (scripts/hooks/spec-before-code-gate.mjs)
• Quick Mode density check: Quick Mode now requires at least 1 substantive concern before presenting the Spec. Zero concerns triggers re-examination. (core/skills/product-spec-builder/references/workflow-quick-mode.md)
• SKILL.md inline evidence rules: Critique Gate evidence format (§section or "spec quote") and verdict thresholds now inlined in product-spec-builder SKILL.md, ensuring anti-sycophancy mechanism works even if references/ aren't read. Based on progressive disclosure experiment.
• Pre-commit sync hook: .git/hooks/pre-commit runs pnpm sync:discover to detect core↔adapter drift. Blocks commit if drift found. (scripts/sync.ts)

Changed

• Gate downgrade risk warning: The recovery message for missing Spec now presents light/none as a risk-bearing option with ⚠️ warning, not a neutral skip instruction. (scripts/hooks/spec-before-code-gate.mjs)
• dev-planner workflow.md: Analysis Phase Step 4 now routes through Plan Critique Check before Output Phase.
• dev-planner SKILL.md: Added Plan Critique Check section with skip conditions.

Fixed

• Skill eval: 14/14 skills now pass eval (0 errors, 0 warnings). Previously: change-manager (missing archive dir), domain-mapper (missing artifacts), bug-fixer (ref-lint "四阶段" warning).
• Adapter drift: plan-hard-gate-rationalization.md was out of sync across adapters. Fixed by pnpm sync.

[v1.47.0] - 2026-06-18

Added

• Simplification intensity dial: FORGE_SIMPLIFY=off|lite|full|ultra in .forge/config controls how aggressively the agent applies YAGNI and simplification. Safety boundaries documented — input validation, data-loss prevention, security, accessibility never simplified. (core/skills/dev-builder/references/first-principles.md, .forge/config.example)
• forge-simplify-check: pnpm forge-simplify-check — light over-engineering review (git diff or --full scan). Detects single-use interfaces, single-method classes, speculative generics, bloated files, commented-out code. Lighter than /code-review. (scripts/forge-simplify-check.mjs)
• forge-debt: pnpm forge-debt — harvests // NOTE: simplification markers into a structured debt ledger at .forge/debt-ledger.md. Supports custom pattern via --pattern. (scripts/forge-debt.mjs)

Changed

• first-principles.md: YAGNI check in action summary now conditioned on FORGE_SIMPLIFY level. New Safety Boundaries section. (core/skills/dev-builder/references/first-principles.md)

[v1.46.0] - 2026-06-18

Added

• forge-ecosystem — Ecosystem library cache: New pnpm forge-ecosystem CLI managing a global, per-language cache of curated library recommendations at ~/.forge/ecosystem/<lang>.json. Subcommands: search (search cache by query), add (manually add entries), refresh (populate from cold-start defaults), get (dump language data), status (cache summary). Search-on-miss pattern — check cache before Context7/WebSearch. (scripts/forge-ecosystem.mjs, scripts/forge-ecosystem/)
• Cold-start defaults for 5 languages: TypeScript (12 entries across 7 categories), Python (10/7), Go (9/6), Rust (8/6), Java (7/5) — curated library recommendations for testing, HTTP, CLI, DB, validation, logging, and config categories. (scripts/forge-ecosystem/cold-start.mjs)
• Ecosystem Cache First principle: dev-builder first-principles.md now guides agents to run pnpm forge-ecosystem search before any online library lookup. Pinned selections go to .forge/ecoresult.json. (core/skills/dev-builder/references/first-principles.md, core/skills/dev-builder/references/development-strategies.md)
• Project-level overrides: .forge/ecoresult.json supports pins (force-include a library) and bans (block a library). Template installed by forge-install. (core/templates/ecoresult-template.json, scripts/install.ts)

Docs

• forge-quickref.md: Added ecosystem command to Skill reference table. (core/templates/forge-quickref.md)

Technical

• Forge scripts: New pnpm forge-ecosystem registered in package.json. All new scripts follow existing .mjs convention with exported functions and test isolation via setCacheDirForTest(). (package.json, scripts/forge-ecosystem/)

[v1.45.0] - 2026-06-14

Added

• forge-bug-fix diagnose --scenario: Three new scenario-specific diagnostic modes — compile (tsc+imports+tsconfig), config (git+env+deps+ports), data (encoding+JSON+CRLF+large files). (scripts/forge-bug-fix.mjs)
• ReqForge-specific debugging strategies: debugging-strategy.md enriched beyond generic 4-stage methodology — Stage 1 now leverages dep-graph blast radius, trace history matching, project-memory known pitfalls, and scenario-specific diagnose; Stage 2 uses forge-bug-fix classify/bisect; Stage 3 adds trace capture and three-strikes stall; Stage 4 adds verify automation and three-layer diagnosis reporting. (core/skills/bug-fixer/references/debugging-strategy.md)

Changed

• bug-fixer workflow.md: During Stage 1, Agent now guided to run scenario-specific diagnose by problem type. (core/skills/bug-fixer/references/workflow.md)
• bug-fixer SKILL.md: forge-bug-fix diagnose reference updated with --scenario flag. (core/skills/bug-fixer/SKILL.md)
• README / README.zh-CN: forge-bug-fix entries updated to include diagnose --scenario.

Technical

• Cross-platform findFiles(): New portable recursive file finder replaces execSync("find ...") calls — resolves find command unavailability in execSync on Windows (cmd.exe default shell). (scripts/forge-bug-fix.mjs)

[v1.44.0] - 2026-06-09

Added

• Critique Gate phase: New pre-Spec adversarial review in product-spec-builder 0-to-1 workflow, after Multi-Stakeholder Review and before Document Generation. Three structural signals (Hidden Assumptions, Unchallenged Decisions, Scope That Should Be Cut) counteract LLM sycophancy bias. Validated by three-round experiment: 5:0 blind eval preference, +5.2 risk visibility, +4.2 rework resistance. Default on for 0-to-1, skipped in Quick/Iteration, user may opt out. (core/skills/product-spec-builder/references/critique-gate.md)
• forge-spec-critique: pnpm forge-spec-critique — automated spec critique scoring tool measuring assumption count, challenge count, and scope decisions. (scripts/forge-spec-critique.mjs)
• forge-spec-blind-eval: pnpm forge-spec-blind-eval — automated A/B blind evaluation experiment for validating spec quality differences. Randomized order, independent LLM session evaluation on 5 dimensions. Supports DeepSeek, OpenAI, and Anthropic providers. (scripts/forge-spec-blind-eval.mjs)

Changed

• Multi-Stakeholder Review stop rule: "Proceed" now transitions to Critique Gate (was directly to Document Generation). (core/skills/product-spec-builder/references/multi-stakeholder-review.md)
• product-spec-builder workflow: 0-to-1 now runs Discovery → Exploration → Clarifying → Refinement → MS Review → Critique Gate → Document Generation. (references/workflow-0-to-1.md)

Docs

• Spec Critique Gate experiment report: forge-spec-experiment/result.md — three-round experiment with manual A/B, dogfood (novel-writing app), and automated blind evaluation.
• Architecture & usage sync: README, README.zh-CN, CHANGELOG, Skill table updated for Critique Gate + three-gate relationship (MS Review / Critique Gate / Step 7 Council).

Sync

• Adapter sync: Propagated critique-gate.md, workflow changes, MS Review stop rule update, SKILL.md gotchas, commands phase index, and product-spec-template Critique Gate Summary section to all 4 adapters.

Added (measurement infrastructure)

• Grovel Index baseline: First-ever run of Position-Swap test across 5 scenarios. Avg GI=0.21 (moderate catering). Key finding: catering is asymmetrical — model pushes back on "don't want" more than it follows "want". (.forge/skills/product-spec-builder/eval/grovel/baseline-2026-06-09.json)
• Conversational Catering Test (CCT v2): Free-form dialogue sycophancy measurement across 3 interventions. Baseline sycophancy 0.8/5 → 0 with anti-cater. Blind spot detection 33% → 67%. (cct-v2-baseline-2026-06-09.json)
• Structured review ceiling test: Structured review format achieves 93% blind spot detection unaided — Critique Gate value is in conversational phase. (catering-baseline-2026-06-09.json)
• Cross-provider validation: Claude Sonnet/Opus match DeepSeek pattern: sycophancy is scenario-specific, not model-specific. "Don't cater" works universally. (cross-provider-2026-06-09.json)
• forge-smoke #13 grovel-baseline: Structural integrity check for measurement framework (no LLM calls). (scripts/forge-smoke/grovel-baseline.mjs)

Docs (supplementary evidence)

• Spec Critique Gate technical report: Added §6.4 with Grovel Index, conversational catering, structural ceiling data; §6.4.7 cross-provider validation. (docs/spec-critique-gate-technical-report.md)

[v1.43.0] - 2026-06-08

Added

• Multi-Stakeholder Review phase: New pre-Spec structured review in product-spec-builder 0-to-1 workflow. Four perspectives (Business, Technical, Experience, Scope/Risk) each return verdict (ok/clarify/blocked). Hard cap at 1 re-scan. Default on for 0-to-1, skipped in Quick/Iteration, user may opt out. (core/skills/product-spec-builder/references/multi-stakeholder-review.md)
• forge-bug-fix bisect + classify: Auto git bisect to locate first bad commit; error classification by pattern matching (compile/runtime/logic/data categories) with severity detection and fix recommendations. (scripts/forge-bug-fix.mjs)
• forge-bug-fix wired into bug-fixer workflow: Dependency Check + Workflow now reference forge-bug-fix diagnose/trace/bisect/classify/verify.

Fixed

• package.json version alignment: Bumped to v1.43.0 to match README/CHANGELOG (was 1.42.0).
• forge-change.test.ts: Added execSync timeout, DRY cleanup logic.
• forge-bug-fix.test.ts: Fixed vitest timeout shorter than execSync timeout (Windows ETIMEDOUT).

Docs

• Architecture & usage sync: Mermaid diagram, 14-Skill tables, four-client install guide, domain-mapper / Multi-Stakeholder Review workflow across README (EN/ZH), Product-Spec, llms.txt, file-structure, dev-map.
• GitHub Wiki source: docs/github-wiki/Home.md updated to v1.43 (14 Skills, Gemini CLI, forge-install quick start).

Sync

• Adapter sync: Propagated bug-fixer workflow changes, forge-bug-fix references, and multi-stakeholder-review to all 4 adapters (claude-code, cursor, gemini-cli, opencode).

[v1.42.0] - 2026-06-08

Added

• /domain-mapper skill: New Interactive-tier skill for transforming unstructured domain knowledge into a structured Markdown database (domain-map.md). 5-step guided pipeline: scope definition → domain snapshot → competitor deep-dive → social analysis → synthesis. Depth levels L1 (snapshot) / L2 (standard) / L3 (deep). Platform-agnostic — maps industries, technologies, codebases, and markets equally. (core/skills/domain-mapper/)

Sync

• gc-audit-routing.md + first-principles.md updates: pnpm sync propagated Build Speed principle, GC audit routing, and all first-principles changes across 4 adapters (claude-code, cursor, gemini-cli, opencode).

[v1.41.0] - 2026-06-07

Added

• Workflow cookbook: core/docs/workflow-cookbook.md — 7 recipes (熔断循环/分头扫/对抗验证/Judge Panel/Pipeline/完整性审查/重试Guard) + 3 assembly examples (review→fix, spec→ship, feedback→evolve). Platform compatibility table included.
• GC audit routing: core/skills/dev-builder/references/gc-audit-routing.md — platform-agnostic decision table for audit depth allocation (Minor/Major/Full GC based on change impact). Change tracking card (Card Table) and assumption registry formats defined.
• dev-builder first-principles.md: Step 5 added to 行动摘要 — Phase complete → read gc-audit-routing.md → execute audit per decision table.

Docs

• Agent dispatch platform-agnostic proposal: Issue #6 — moving hardcoded code-review agent dispatch (code-reviewer-bug/security/types/design) from workflow.md to platform-agnostic reference docs.
• Published dev.to: "Smarter Resource Allocation Beats Stronger Models" — English article on attention allocation (GC audit) + sample allocation (2.5 layer anchors).
• Zhihu drafts: 3 Chinese articles prepared (2.5-layer manifesto, triple allocation, benchmark).

[v1.40.0] - 2026-06-07

Added

• UI-Spec.md — model-to-model 2.5 layer: design-maker auto-generates a structured UI spec (component hierarchy, states, responsive rules, reusable components) during verification. dev-builder reads it at startup, skipping the guess-from-pixels trap. Ephemeral (regenerated each design cycle), not committed. (core/templates/ui-spec-template.md, core/skills/design-maker/SKILL.md, core/skills/dev-builder/references/first-principles.md)
• Machine gate recovery options: When spec-before-code, plan-before-build, idea-validation, implementer, or hallucination gates block a write, the block message now includes numbered recovery steps ("1. Run /product-spec-builder → 2. Fill criteria → 3. Confirm → 4. Retry"). (scripts/hooks/spec-before-code-gate.mjs, core/hooks/hallucination-gate.sh, core/hooks/hallucination-gate.bat)
• Attention summaries across skills: 7 skills (bug-fixer, code-review, change-manager, release-builder, dev-planner, product-spec-builder, design-brief-builder) now have ⚠️ 当前 Task 行动摘要 in their first-principles.md, leveraging recency/primacy bias. (core/skills/*/references/first-principles.md)
• Self-review step for bug-fixer and release-builder: Hot-context self-review between fix generation and verification (bug-fixer) / between smoke test and release confirmation (release-builder). (core/skills/bug-fixer/references/workflow.md, core/skills/release-builder/references/workflow.md)
• design-maker first-principles.md: New file with 出图前行动摘要 covering component hierarchy, state coverage, responsive rules, priority labeling, and UI-Spec generation. (core/skills/design-maker/references/first-principles.md)

Changed

• Anti-ai-slop-checklist guardrails restored: 7 removed check items restored across change-manager, code-review, bug-fixer, release-builder, and design-brief-builder. (core/skills/*/references/anti-ai-slop-checklist.md)
• bug-fixer anti-ai-slop-checklist: Added "跳过复现" check. (core/skills/bug-fixer/references/anti-ai-slop-checklist.md)
• design-maker SKILL.md: Fixed typo "Anti-ai-slip" → "Anti-ai-slop". (core/skills/design-maker/SKILL.md)

Docs

• 2.5-layer-manifesto extension: New §5.6 "Model-to-Model 2.5 Layer" added to both Chinese and English versions, explaining how the 2.5-layer principle extends from person→model anchors to model→model intermediate representations. (docs/2.5-layer-manifesto.md, docs/2.5-layer-manifesto.en.md)

[v1.39.0] - 2026-06-06

Added

• Spec difficulty markers: Product-Spec.md now has ## Known Difficult Spots section with 🔴/🟡/🟢 levels. dev-planner propagates difficulty to DEV-PLAN phases. dev-builder adjusts execution speed accordingly (🔴 = 怵然为戒, slow down + self-review; 🟢 = fast pass). New difficulties discovered during a Phase are auto-appended via 善刀而藏之. (core/templates/product-spec-template.md, core/templates/dev-plan-template.md, core/skills/dev-planner/references/workflow.md, core/skills/dev-builder/SKILL.md)
• Anti-slop anchor reform: Replaced 9 "don't" rules in dev-builder anti-slop checklist with 3 perfect code anchors (error handling / API endpoint / test patterns). Old checklist demoted to pre-delivery safety net. Based on the insight that LLMs are pattern matchers, not rule followers. (core/skills/dev-builder/references/anti-ai-slop-checklist.md)
• 放下骨架 (Phase 1 catalyst): Phase 1 explicitly lays down domain models, types, interfaces, and conventions as the project's structural anchor. All subsequent Phases naturally continue this anchor. (core/skills/dev-builder/references/first-principles.md, core/skills/dev-builder/SKILL.md, core/skills/dev-builder/references/workflow.md)
• 善刀而藏之 closing ritual: Mandatory 5-step Phase completion: review → celebrate → append hidden difficulties to Spec → log decisions → clear context. Inspired by Zhuangzi's butcher who sharpens his knife by sheathing it after use. (core/skills/dev-builder/references/phase-completion-assessment.md)
• Change-manager auto-rollback: On verify failure, pnpm forge-change restore <name> reverts files from pre-apply snapshot. New snapshot and restore commands. (core/skills/change-manager/references/workflow.md, scripts/forge-change.mjs)
• Security rules template: core/templates/security-rules-template.md — 3 hard rules (no hardcoded secrets, always validate inputs, audit logs for sensitive operations) + PII-in-logs constraint. Installable as .claude/rules/security.md. (core/templates/security-rules-template.md)
• Benchmark suite: benchmark/ directory with side-by-side comparison of old anti-slop (9 rules) vs new approach (3 anchors) on todo-cli. Results: both pass all tests; new approach produces 15% shorter code. (benchmark/RESULTS.md)
• Design-maker multi-alternative + gradual refinement: --alternatives N mode generates N design variants with cross-comparison table. Gradual refinement mode delivers in 3 tiers (structure → interaction → edge cases). (core/skills/design-maker/SKILL.md)

Changed

• No dashboard Web UI: Decision made not to build it — Forge users live in CLI, not web.
• Skill evolution P1/P2: Deferred until feedback data accumulates to justify the work.

Docs

• 2.5层设计文章: docs/2.5-layer-manifesto.md (Chinese) and .en.md (English) — full article explaining the design philosophy shift from shackles to anchors, with benchmark results, anchor selection guide, and troubleshooting.

[v1.38.0] - 2026-06-06

Added

• Template marketplace: core/templates/project-starters/ — curated project starters with metadata + skeleton files. 4 starter templates: next-fullstack, cli-tool, express-api, electron-app. forge-scaffold init <template> [dir] — one-command init with placeholder replacement, post-init steps, and loadout recommendations. forge-scaffold list-templates — browse available starters.
• Gemini CLI adapter: adapters/gemini-cli/ — 4th AI client adapter. Uses .gemini/ directory with GEMINI.md as control file, agents/ for subagent definitions, skills/ for Forge skills. Synced from core via pnpm sync. All 944 files in sync across all 4 adapters.

[v1.35.12] - 2026-06-01

Added

• forge-ops: new pnpm forge-ops <url> — production monitoring loop: health check → verify suite → baseline comparison → regression detection → fix brief → report. Supports --interval <sec>, --baseline save|compare, --fix. Output: .forge/ops/report.md. (Phase 16 candidate)
• forge-loop --fde: Forward Deployed Engineer mode — reads Product-Spec.md + DEV-PLAN.md context before execution, generates structured evidence report (.forge/evidence/phase-N-report.md) on completion. pnpm forge-fde convenience alias.
• Evidence chain: forge-loop --fde outputs machine-readable evidence report with test pass rate, delivery checklist completion, UI check results, file change list, and pass/fail verdict per deliverable.
• skill-eval judge v2 rubric: expanded from 5 to 6 dimensions — added "工作流质量与可重复性" (workflow quality & repeatability, 30% weight) focusing on result consistency, verification loops, and real-project usability. Baseline comparison dimension added. judge-config template bumped to version 2.

Fixed

• forge-ui-check: phases with progress-table entry but no detailed ## Phase N: section no longer cause false-positive failures (exit 0 instead of 1, so forge-loop doesn't get stuck in dead loop).
• forge-loop checkUi: catches "Phase not found"/"无 UI" errors and treats them as pass, preventing iteration deadlock on non-UI phases.

[v1.36.0] - 2026-06-04

Added

• Quality Rubric for all 13 skills: every Skill now has a domain-specific 8-16 item scoring rubric with ship threshold and critical-item-zero rule. Run pnpm validate-skill --score core/skills/<name> to compute.
• Dimension checklists for 9 skills: formal references/dimension-checklist.md with Must-Have / Recommended / Optional tiers. New: bug-fixer, change-manager, dev-builder, release-builder, design-maker, evolution-engine, feedback-writer, request-dispatcher, skill-builder.
• Anti-rationalization for 6 skills: change-manager, design-brief-builder, design-maker, evolution-engine, feedback-writer, release-builder — each with domain-specific Rationalization|Reality tables to prevent AI from skipping critical steps.
• Anti-ai-slop checklists for 5 skills: bug-fixer, change-manager, dev-builder, code-review, release-builder — 7-9 item Chinese pass/fail checklists for pre-delivery self-check. Wired into all workflows.
• references/ filled for request-dispatcher and skill-builder: request-dispatcher gained first-principles, output-style, workflow, anti-rationalization files. skill-builder gained first-principles, output-style (new), workflow, anti-rationalization (new) — previously these sections were inline-only or absent.
• Eval pack verification: all 13 skills pass pnpm skill-eval <name> static checks (triggers.json + cases.json).

Changed

• request-dispatcher SKILL.md -66 lines: extracted inline content to references/ files.
• skill-builder SKILL.md -104 lines: extracted inline content to references/ files; added Output Style section (previously absent).

[v1.35.14] - 2026-06-04

Added

• Skill eval packs for all 13 core skills: every framework Skill now has .forge/skills/<name>/eval/ with triggers.json (2 positive + 2 near-miss negative trigger cases) and cases.json (output assertions: fileExists + maxBytes + regexChecks). pnpm skill-eval <name> static check passes on all skills. Enables regression detection and trigger accuracy verification.

[v1.35.13] - 2026-06-03

Added

• forge-hashline verify-brief: pnpm forge-hashline verify-brief <brief-path> — parses fix-brief.md hashline entries, verifies all hash anchors match current files (before-fix) or detects whether files were actually modified (after-fix with --after-fix). Reports OK/STALE/UNCHANGED/MISSING per entry. --json for machine-readable output.
• forge-hashline apply-brief: pnpm forge-hashline apply-brief <brief-path> — auto-creates files marked as (新文件) in fix-brief.md.
• Hashline verification loop: forge-loop and forge-phase-loop now run verify-brief on brief generation (freshness check) and on next iteration (application check), plus apply-brief for auto-creating new files. Breaks the string-matching retry deadlock in fix→verify cycles.
• forge-hashline: pnpm forge-hashline hash|verify|edit <file> — SHA256 content-hash anchored editing. Atomic write via tmp+rename, CRLF→LF normalization, block hashing with --lines N:M. Integrated into forge-loop/forge-phase-loop fix-brief.md as **Hashline**: anchor entries.
• sync --discover: pnpm sync:discover — drift detection between core/ and adapters/. Reports drifted (hash mismatch), orphan (adapter-only), missing (core-only).
• forge-loop --strict/--linear: --strict stops on first test failure with review.md; --linear runs single pass without iteration.
• skill-eval trigger: auto-generates 20 diverse queries from SKILL.md, runs static checker, logs results. Judge pipeline: judge-prep → judge → judge-record.

[v1.37.0] - 2026-06-05

Added

• forge-coverage: pnpm forge-coverage <specifier> — test coverage gap detection and auto-stub generation. Scans vitest/luma coverage JSON, maps uncovered functions, generates stub files. Naming-convention fallback when coverage JSON unavailable.
• forge-scaffold: pnpm forge-scaffold <specifier> — project scaffolding from Product-Spec.md. Generates DEV-PLAN.md stubs, directory structure, and package.json from feature descriptions.
• forge-step-capture: trajectory capture and evolution proposal loop for forge-step-capture. Records step-level execution traces for failure attribution.
• forge-skill-retrieve: skill dynamic retrieval engine — context-aware section selection from skill references based on user intent matching.
• forge-evidence: pnpm forge-evidence generate|aggregate|list — three-tier evidence grading reports: dev (coverage, git diff, traces), lead (gate pass rates, trends), client (deliverables, acceptance criteria). Wired into forge-loop as always-on.
• forge-ops 完整化: pnpm forge-ops <url> — production monitoring loop. Added sendSlack/sendFeishu webhook notifications, auto-deploy mode, baseline comparison. forge-loop sends ops alerts on completion paths.
• forge-release: pnpm forge-release version|changelog|tag|check|all — release automation. Version bump, auto-changelog entry, git tag, pre-release checklist.
• forge-evolve: pnpm forge-evolve status|scan|propose|apply — evolution engine loop closure. Scans .claude/feedback/ entries, identifies Level 2/3 candidates, generates evolution proposals.
• forge-change: pnpm forge-change init|list|check|archive — change management CLI. Scaffolds change directories from templates, tracks active/archived status. 8 E2E tests.
• forge-skill-eval: pnpm forge-skill-eval status — centralized eval dashboard across all 13 skills. Shows triggers/cases/judge-config/judge-history per skill.
• skill-eval judge-all: batch judge-config.json deployment to all .forge/skills/ packages. All 13 skills now have judge-config.json deployed.

Changed

• forge-loop: coverage scan after auto-fix, ops alerts at allOk/max-reached, evidence always-on (no longer FDE-only), forge-evolve scan on completion paths.
• Feedback backfill: 3 feedback entries backfilled with failure_class and scores for evolution engine processing.

[v1.36.0] - 2026-06-04

Changed

• Prompt slimming: change-manager SKILL.md index-only (11k→4.7k); propose/apply/verify/archive workflow → references/; must read references/workflow.md.

[v1.35.7] - 2026-05-26

Changed

• Prompt slimming (P6): CLAUDE.md General Rules 指针化（7.9k→5.5k）；tool-call / web-first / pin / forge-install / preflight / skill-eval / CLI / loadout → forge-quickref §通用规则；移除重复的 [Available Skills] 段（保留 [Skill Dispatch]）。
• forge-quickref: 新增 §通用规则（CLAUDE.md 指针目标），含 --loadout 安装说明。
• skill-builder: 新 Skill 注册步骤改为只更新 [Skill Dispatch]。

[v1.35.6] - 2026-05-26

Added

• forge-install --loadout <name> (-l): install only skills/agents from a loadout (+ _shared), merge loadout hooks into adapter settings, write .forge/loadout-active.json. Shared logic in scripts/loadout.ts; apply-loadout refactored to use it.

Changed

• loadout-scenarios.md: document --loadout install path; skills are filtered on install when loadout is specified.

[v1.35.5] - 2026-05-30

Changed

• Prompt slimming (P4): bug-fixer SKILL.md index-only (~5.2k chars); four-stage debugging, CoT, three-layer model → references/; must read references/workflow.md; retry-gate unchanged.
• Prompt slimming (P4): code-review SKILL.md index-only (~4.5k chars); principles/judgment-spectrum/anti-rationalization → references/; parallel reviewer workflow unchanged in references/workflow.md.

[v1.35.4] - 2026-05-30

Changed

• Prompt slimming (P3): design-brief-builder SKILL.md index-only (~3.8k chars); interview/workflow → references/; must read references/workflow.md.
• Prompt slimming (P3): release-builder SKILL.md index-only (~4.8k chars); checklist/strategy/8-step workflow → references/; privacy audit steps unchanged.

[v1.35.3] - 2026-05-30

Changed

• Prompt slimming (P2): product-spec-builder SKILL.md index-only (~7k chars); principles/output-style/judgment-spectrum/startup routing → references/; Quick Mode isolated in references/workflow-quick-mode.md (do not load full 0-to-1 interview chain); Karpathy → _shared/karpathy-discipline.md.

[v1.35.2] - 2026-05-30

Added

• lite loadout (core/loadouts/lite.json): 8 skills (spec, change-manager, plan, build, bug, review, feedback, dispatcher) + 8 hooks — token-conscious greenfield/brownfield without design/release/evolution/skill-builder.

Changed

• Prompt slimming (P1): dev-planner SKILL.md index-only (~5k chars); workflow/analysis → references/; must read references/workflow.md before planning.
• CLAUDE.md volatile trim: Project State Detection → forge-quickref.md §项目状态路由; bootstrap points to quickref for session routing.

[v1.35.1] - 2026-05-30

Added

• Cross-client handoff: forge-quickref and agents-template — fixed file read-order when switching AI clients (Claude Code ↔ Cursor ↔ OpenCode); update memory/handoff.md before leaving.

Changed

• Prompt slimming (P0): dev-builder SKILL.md index-only (~7k chars); full workflow → references/workflow.md; principles → references/first-principles.md; shared pointers in core/skills/_shared/ (karpathy, hard-gate, output-style). Compressed forge-bootstrap.md; deduped bug-fixer / code-review Karpathy blocks.

[v1.35.0] - 2026-05-29

Added

• AGENTS.md template (core/templates/agents-template.md): enhanced with parallel worktree workflow — branch naming convention (Type/user/short-task), dependency sharing scripts, multi-agent coordination rules, and integration checklist. Written by pnpm forge-install as project root AGENTS.md.
• Structured exploration trace (scripts/forge-trace.mjs): init, decision, dead-end, evidence, summary commands for recording Phase decisions, abandoned approaches, and claim-to-evidence bindings. Stored in .forge/trace/phase-<N>.json. Hooked into dev-builder Loading Phase (init) and Phase Completion Assessment (record). New trace-fresh check in forge-verify.
• Scope filter (巽) (scripts/forge-scope.mjs): init, check, show commands for declaring Phase file boundaries (modify/readonly/outOfScope). Creates .forge/active-scope.json. Enforced by forge-verify scope-check against git diff --name-only HEAD. Inspired by 八卦信息论 protocol audit.
• Evolution proposals (兌): dev-builder Phase Completion After All Pass triggers evolution-engine scan for actionable patterns. Presents Y/N proposals to user before proceeding. Logs deferred proposals to .forge/evolution-proposals.md. Closes the feedback→evolution loop proactively.
• Skill quality judge (pnpm skill-eval judge): independent sub-agent evaluates Skill against 5-dim rubric (structure/specificity/failure-mode/anti-patterns/effectiveness). Commands: judge-prep, judge (prints briefing), judge-record. Results append to judge-history.json. Inspired by darwin-skill's autonomous evaluation loop.
• Skill authoring patterns (core/docs/skill-authoring-patterns.md): practical reference synthesizing 300 prompts patterns, darwin-skill rubric, and ReqForge Harness experience — workflow design, failure-mode encoding, actionable specificity, anti-pattern blacklists, evaluation checkpoints, multi-agent collaboration, and rubric self-check table.

Changed

• dev-builder: Loading Phase Step 5 initializes trace; Phase Completion Assessment After All Pass records decisions/dead-ends/evidence.
• forge-verify: new trace-fresh check — validates trace files have content (decisions or dead-ends); new scope-check — enforces .forge/active-scope.json boundaries against git diff.
• forge-quickref: index rows for AGENTS.md and .forge/trace/; trace command quick reference.
• phase-exit-guard: now checks both .forge/phase-exit-block and .forge/.verify-block (written by forge-verify on baseline regression).

[v1.34.0] - 2026-05-28

Added

• Matt Pocock Skills 对照：mattpocock-skills-comparison.md — 与 skills-main 能力映射；并行安装建议
• P1 Light Grill：product-spec-builder Light Grill Mode + light-grill-mode.md
• P2 Zoom-out / 架构保健：zoom-out-pass.md、architecture-health-pass.md
• P3 GitHub slices：github-issues-slices-template.md + dev-planner optional export
• request-dispatcher routing for grill / zoom-out / architecture health / issue slices

[v1.33.0] - 2026-05-28

Added

• 腾讯 Harness 镜子对照：tencent-harness-mirror-comparison.md — 显形/三块石碑/不可能三角 ↔ Forge
• .forge/project-taste.md：团队口味模板（installProjectTaste + forge-install）
• Judgment Spectrum (S1–S5)：product-spec-builder、code-review；dev-builder Loading Phase 读 project-taste

[v1.32.0] - 2026-05-28

Added

• SkillOpt 对照：skillopt-comparison.md — rollout/反思/有预算编辑/验证门/rejected buffer ↔ Forge
• skill-eval：rejected-edits.json 模板；cases 文档化 split: train | held-out
• evolution-engine：Bounded Skill Edits（≤3 条 add/delete/replace + 失败写入 rejected-edits）
• Template evolution-bounded-edit-template.md

[v1.31.0] - 2026-05-28

Added

• Skill eval for user-project custom Skills: pnpm skill-eval init <name> + pnpm skill-eval <name>; templates in core/templates/skill-eval/; doc skill-eval.md
• skill-builder: deliver eval/ pack (triggers + output assertions); forge-install writes .forge/skills/_template/eval/
• README / DEV-PLAN / CLAUDE.md / quickref documentation for skill-eval

[v1.30.0] - 2026-05-28

Added

• Release preflight: scripts/preflight.ts + pnpm preflight — git/version/build privacy + .forge/preflight.json rules
• Templates: preflight-config.template.json, preflight-wechat.example.json; doc external-publish-preflight.md
• release-builder Step 3b Preflight Gate; forge-install writes .forge/preflight.json + preflight-wechat.example.json
• README / quickref / DEV-PLAN / CLAUDE.md documentation for preflight

[v1.29.0] - 2026-05-28

Added

• .forge/security-guidance.md: team-versioned security rules template; pnpm forge-install (installSecurityGuidance); security-guidance-comparison.md.
• forge-verify security-patterns: lightweight eval / new Function scan when .forge/security-guidance.md exists.

Changed

• code-review / release-builder / dev-builder: read .forge/security-guidance.md for security-sensitive work.
• forge-quickref: index row for security-guidance.
• founders-playbook-comparison: Security Gate links to security-guidance doc.

[v1.28.0] - 2026-05-27

Added

• dev-map: project-level navigation index at .forge/dev-map.md — AI explores module structure, existing patterns, and change impact chains before coding. Template at core/templates/dev-map-template.md; written by pnpm forge-install (installDevMap). Maintained by dev-builder (who changes code updates the map).
• forge-verify: unified post-verification entry point pnpm forge-verify with 5 checks (skill-quality, compile, test, no-placeholders, dev-map-fresh) and baseline comparison (--baseline save|compare|check). Scripts at scripts/forge-verify.mjs + scripts/forge-verify/baseline.mjs.
• request-dispatcher: added missing skill.json metadata and commands/request-dispatcher.md.

Changed

• dev-builder: Loading Phase now runs pnpm forge-verify --baseline save before coding; Phase Completion Assessment runs pnpm forge-verify --baseline compare after four-step verification and updates .forge/dev-map.md. New First Principle: Post-Verification Gate.
• check-evolution hook: Part 2 project state detection now includes dev-map status.
• Smoke tests: adapters-sync and skills-complete updated from 12→13 skills to reflect request-dispatcher count.

[v1.27.0] - 2026-05-27

Added

• CLAUDE.md zone partitioning: Immutable/Stable/Volatile zones — session-varying content at end of prompt preserves prefix for caching.

Changed

• Settings hook format: PreToolUse/PostToolUse now arrays (was objects); removed invalid events (PreCommit, BeforeCommand, AfterCommand, PostCommit, OnInit); merged BeforeCommand/AfterCommand into PreToolUse/PostToolUse arrays.
• Skill quality: 13/13 PASS, 0 FAIL (4 at perfect 33/33). Fixed Gotchas/step-counting cross-contamination, Strategy vs Strategies regex mismatch, TBD text matching \[.*\] sed pattern.
• Feedback cleanup: Converted stray JSON files to standard .md frontmatter format; fixed FEEDBACK-INDEX.md missing entries.

[v1.26.0] - 2026-05-27

Added

• .forge/quickref.md: human one-pager (gates, 4 principles, 8 rules summary, anti-patterns, Skill commands); template core/templates/forge-quickref.md; written by pnpm forge-install (installForgeQuickref).
• Idea Validation Gate: Product-Spec.md § Idea Stage Exit Criteria (3 questions) required before app code (spec-before-code-gate.mjs); template + product-spec-builder devil's advocate step.
• MVP Scope block in DEV-PLAN.md template (in scope / out of scope / scope amendment criteria); dev-planner + dev-builder enforcement.
• founders-playbook-comparison.md — Playbook ↔ Forge mapping.
• shuge-openspec-superpowers-comparison.md — OpenSpec + Superpowers article ↔ Forge; change-manager ↔ dev-builder handoff.
• request-dispatcher meta-skill — routes ambiguous user requests to the correct Forge Skill.
• HTML knowledge boundaries (<!-- begin/end -->) on all 12 skills for progressive disclosure / validator cross-refs.

Changed

• Agent execution discipline: rule 8 is now an explicit verify loop (fix → re-run checks until green; last-run output required); anti-patterns + framework test-placement table in session-execution-discipline.md.
• behavior-rules.md: violation signals link to feedback case IDs; 何时可简化 exemption table (prior).
• forge-bootstrap / implementer: verify-loop summary aligned with discipline doc.
• README / README.zh-CN / docs/github-wiki/Home.md / llms.txt: verify loop + quickref on install.
• Audit alignment: root + test-demo Spec/Plan examples; core/templates/product-spec-template.md; repo .forge/quickref.md; change-proposal scope discipline; machine-gates-doc + core/hooks/AGENTS.md six-step gate docs.
• change-manager / dev-builder: Change-Scoped Mode; Delta Spec + G/W/T in change-specs-template; two-tier Task review (spec compliance → code quality).
• Skills: version metadata standardization; Anti-Rationalization tables unified; progressive disclosure for product-spec-builder, bug-fixer, code-review.

Fixed

• validate-skill: broaden Strategy pattern in cross-reference check; validator score fixes (change-manager, evolution-engine, feedback-writer, design-maker).

[v1.25.0] - 2026-05-25

Added

• Harness hardening (Superpowers-aligned): forge-bootstrap session iron laws; product-spec-builder / dev-planner HARD-GATE; evolution Skill TDD quartet + failure_class; tests/skill-fixtures + forge-smoke 12 smokes; mandatory implementer + worktree; .forge/*-confirmed.json + implementer-session.json PreToolUse chain; drift-map-template.md; forge-smoke skill-bypass (P2); validate-skill requires [HARD-GATE] on spec/plan/build skills.
• product-spec-builder PM frameworks (MIT-adapted from pm-skills): optional OST, JTBD value prop, assumptions, competitive briefs + Spec template sections.
• Chain-of-Thought (CoT): conversation-strategy templates; implementer step 0b + reasoning_summary; bug-fixer diagnostic checklist; forge-bootstrap Iron Law 9 (no user magic phrase required).
• Agent execution discipline (8 rules): session-execution-discipline.md; bootstrap summary; full copy in agents-template.md; README Agent execution discipline section.
• release-builder: Finishing Branch Checklist (merge / PR / keep / discard). bug-fixer: Phase-1-before-fix cross-ref to systematic debugging.

Changed

• hallucination-gate now runs full spec-before-code-gate.mjs (Spec, confirmations, Plan, implementer session). feedback-observer auto failure_class heuristics (P1-lite).
• README / README.zh-CN / llms.txt / GitHub Wiki source: PM frameworks + CoT + Agent execution discipline (8 rules); Requirements depth section; Workflow spec-confirmation steps; version v1.25.0.
• DEV-PLAN / test-demo README: pnpm forge-smoke documented as 12 smokes (~15–30s).

[v1.24.0] - 2026-05-24

Added

• Comparison docs: autoresearch, llm-council, jobs, llm-wiki gist.
• Primary metric per DEV-PLAN Phase; dev-builder Spec/Plan read-only + Task micro-cycle; code-review risk_rank (S×I×C); PROJECT-HEALTH-template.md; product-spec LLM-as-Judge subsection; Spec Step 7 quality council.

Fixed

• OpenCode adapter: pnpm sync now copies root CLAUDE.md → .opencode/AGENTS.md (was empty agents-template.md, breaking Forge dispatch).
• forge-smoke machine-gates-doc: asserts OpenCode AGENTS.md mirrors Machine Gates + Skill Dispatch.

Changed

• memory-system.md: LLM Wiki pattern cross-reference; query filing → ADR rule.

[v1.23.0] - 2026-05-24

Added

• forge-smoke release gate: pnpm forge-smoke runs 9 static smokes (skills, loadouts, adapter sync, hooks, templates, machine gates, platform compliance, workflows); CI workflow .github/workflows/forge-smoke.yml on push/PR (no cron).
• loadout-scenarios.md: scenario → loadout → Skill path picker; scenarios[] tags on each built-in loadout JSON (required in schema).
• platform-compliance.md: GitHub/OSS policies (no cron CI, no central secrets, fork intent); forge-smoke checks workflows + doc presence.

[v1.22.2] - 2026-05-23

Added

• retry-gate in all loadouts and apply-loadout registry (was in settings.json since prior commit).

Changed

• settings.windows.json (repo + claude-code + cursor): align with Unix — phase-exit-guard, retry-gate; remove default PostCommit auto-push.
• README / README.zh-CN / llms.txt: 10 default hooks; document retry-gate; note core/docs/ on GitHub for adapter-only installs.
• Skill doc links to core/docs/* use GitHub URLs (work when only adapter bundle is copied).

[v1.22.1] - 2026-05-23

Added

• agent-harness-seven-layer-map.md: AGENT魔方「七层 Harness」↔ ReqForge mapping (teaching skeleton vs product pipeline).
• phase-exit-guard hook: blocks agent stop while .forge/phase-exit-block exists (Ralph-style Phase completion); wired in all loadouts + adapter settings.
• retry-gate (shipped in b9f5dfc): .forge/.retry-counter.json escalation blocks proceed; dev-builder / bug-fixer retry limits.

Changed

• dev-builder phase-completion-assessment: when/how to write/remove .forge/phase-exit-block.
• evolution-engine: proposals require Predicted effect + Verify by (observability-driven evolution).

[v1.22.0] - 2026-05-23

Added

• context7-comparison.md: vs upstash/context7 — stack with ReqForge; optional CLAUDE.md rule snippet.
• Library Docs Strategy in dev-builder/references/development-strategies.md (Context7 before WebSearch for dependencies).
• Context7 Library ID column in Product-Spec Technical Direction and DEV-PLAN Tech Stack templates.

Changed

• dev-planner SKILL: resolve Context7 IDs when MCP/CLI available.
• web-app loadout: optional Context7 MCP entry.
• README / README.zh-CN: Context7 callout + link.

[v1.21.0] - 2026-05-23

Added

• harness-maturity-checklist.md: P0/P1/P2 self-assessment for harness readiness; positioning vs consumer scheduled-automation products.
• Product-Spec: Integrations, Operations & Scheduling section (integrations, notifications, jobs, observability); Quick Mode infers defaults.

Changed

• README / README.zh-CN: Harness OS analogy + scope boundary (shippable product vs post-chat life automation).
• product-spec-builder: Fix broken 0-to-1 workflow reference in SKILL.md.

[v1.20.9] - 2026-05-23

Added

• open-design-comparison.md: vs nexu-io/open-design; when to use OD vs ReqForge design skills.
• Design P0 from OD: design-discovery-questionnaire, visual-direction-presets, anti-ai-slop-checklist, design-self-critique; Design-Brief template sections.

Changed

• design-brief-builder / design-maker workflows reference new references/ files.

[v1.20.8] - 2026-05-23

Added

• superpowers-comparison.md: ReqForge vs obra/superpowers — workflow mapping, TDD/subagent alignment, when to combine or pick one.

Changed

• README / README.zh-CN: Superpowers callout + reference table row; cross-links in openspec/openhuman comparison docs.

[v1.20.7] - 2026-05-23

Changed

• Brand messaging: README/README.zh-CN lead with「需求→产品」, subtitle Agent Harness; product name unified as ReqForge in hero (repo name unchanged).
• repo-metadata.json description aligned; re-run pnpm set-github-metadata to refresh GitHub About.

[v1.20.6] - 2026-05-23

Added

• GitHub discoverability: README top — OpenSpec one-liner + architecture mermaid; .github/repo-metadata.json + scripts/set-github-metadata.mjs for About description/topics.

Changed

• package.json: repository/homepage/keywords aligned with repo metadata.

[v1.20.5] - 2026-05-23

Changed

• memory-guard hook: PostToolUse runs memory-guard (wraps context-compaction + check-handoff). memory-check stays on AfterCommand. Default hook count 9→8. Legacy scripts retained for direct invoke.

[v1.20.4] - 2026-05-23

Changed

• SKILL progressive disclosure: dev-builder and product-spec-builder oversized sections moved to references/; main SKILL.md under 500 lines (validate-skill warning cleared).
• scripts/split-skill-references.mjs: Utility to repeat the split pattern for other skills.

[v1.20.3] - 2026-05-23

Changed

• Commands layer: All 12 skills' commands/*.md are index-only pointers to SKILL.md.
• auto-push: Removed from default adapter settings.json and from full/web-app/cli-tool loadout hook lists (optional manual enable documented in README).

[v1.20.2] - 2026-05-23

Changed

• P0 — Spec vs change-manager: product-spec-builder iteration no longer creates changes/; moderate scoped features route to /change-manager. Sole owner of changes/ documented in both skills.
• P1 — Review default simple: code-review / code-reviewer / dev-builder default change_complexity=simple; parallel 4-agent review only when escalated.
• P1 — Thin commands: product-spec-builder, change-manager, code-review, dev-builder commands/*.md are index-only (details in SKILL.md).
• Product-Spec.md: Template market and Dashboard moved to Roadmap (not shipped).

[v1.20.1] - 2026-05-23

Added

• openhuman-comparison.md: Forge vs OpenHuman positioning; optional memory backends documented in memory-system.md.
• change-verify-template.md: Template for changes/<name>/verify.md in change-manager.

Changed

• CLAUDE.md: Mission step for brownfield /change-manager; Project State Detection routes active changes/ folders.
• Loadout docs: cli-tool / minimal explicitly omit change-manager (use full or web-app for brownfield).
• behavior-boundaries.md: Cross-links to comparison docs; README architecture diagram 12 Skills.

[v1.20.0] - 2026-05-23

Added

• change-manager Skill: Brownfield workflow for existing projects with Product-Spec.md — changes/<name>/ with propose → apply → verify → archive; templates for proposal, specs, design, tasks; command /change-manager.
• openspec-comparison.md: Forge vs OpenSpec positioning, artifact mapping, and when to use /change-manager.

Changed

• 12 Skills (was 11): change-manager included in full and web-app loadouts; CLAUDE.md dispatch updated.
• README / README.zh-CN: What's New, workflow, install tree, repo structure, version badge; llms.txt, DEV-PLAN.md, file-structure.md aligned.

[v1.19.1] - 2026-05-23

Fixed

• Hallucination Gate wired: PreToolUse registered in all adapter settings.json / settings.windows.json; hook reads tool_name from stdin JSON (was tool only).
• Parallel review docs aligned: code-review/SKILL.md, dev-builder/SKILL.md, bug-fixer/SKILL.md, README workflow diagram, and reqforge-dev-build.mdc now match parallel agent review (removed stale two-stage Stage 1/2 language).
• Confidence scale unified: code-review SKILL uses 0.0–1.0 thresholds (≥0.6 confirmed, 0.3–0.6 suspected) matching agents/commands.
• Missing commands layer: Added commands/*.md for design-brief-builder, design-maker, evolution-engine, feedback-writer.
• Loadout check-sync: Removed check-sync from user-facing loadouts (ReqForge-repo-only hook).
• Version alignment: package.json, README badge, DEV-PLAN.md updated to v1.19.1; Product-Spec and core/docs/ agent lists corrected (10 agents).

Added

• scripts/validate-skill.mjs: Cross-platform SKILL.md validator (default pnpm validate-skill).
• scripts/apply-loadout.ts: Print/validate loadout bundles; merge hook registrations into adapter settings.
• Windows hallucination-gate.bat: Node-based JSON parsing (replaces broken string parsing).

[v1.19] - 2026-05-23

Added

• Loadout mechanism: core/loadouts/ with loadout.schema.json and 4 built-in loadouts (full, web-app, cli-tool, minimal). Each loadout is a reusable bundle of skills, agents, hooks, and MCP server recommendations.
• Sync support: core/loadouts/ added to sync map for all 3 adapters.

Changed

• scripts/sync.ts: added core/loadouts → adapter loadouts mapping.

[v1.18] - 2026-05-23

Added

• skill.json metadata: All 11 skills ship with machine-readable skill.json (name, version, triggers, prerequisites, agents, hooks). JSON Schema at core/skills/skill.schema.json.
• Commands layer: 7 skills get commands/<name>.md with YAML frontmatter + phased workflows (Goal → Actions → Acceptance).
• Parallel agent code review: 4 specialized agents (design, bug, security, types) run concurrently with confidence-scored aggregation (≥0.6 confirmed, 0.3-0.6 suspected, <0.3 suppressed). Cross-agent boost for corroborated findings.
• Hallucination Gate: PreToolUse hook validates Write/Edit target directories exist before file creation.
• Project state injection: check-evolution.sh detects Spec/Plan/Code state on session start and injects routing guidance.
• validate-skill.sh — skill.json validation: Existence check + required fields (name, version, description, triggers.auto/manual/command).

Changed

• Sub-Agent count 6→10 (4 new specialist code reviewers + existing planner/test-writer)
• code-review SKILL.md: serial two-stage review → parallel agent dispatch + confidence aggregation
• code-reviewer agent v2.0: parallel dispatch workflow, aggregation rules, cross-agent boost
• sub-agent-orchestration.md: documented parallel review pattern
• AGENTS.md (both core/agents/ and core/skills/): added skill.json + commands requirements

Fixed

• validate-skill.sh python3/node fallback: Windows Store python3 shim no longer breaks validation

[v1.17] - 2026-05-22

Added

• Decidable Activation — [Not For] section: All 11 skills include [Not For] specifying when NOT to use the skill.
• Three-Layer Diagnostic Model: bug-fixer reports Symptom → Design Flaw → Principle Violation layers.
• Numeric Quality Rubric: skill-builder 16-item, 32-point scoring system (ship threshold ≥ 24).
• create-skill.sh scaffold: CLI tool to scaffold new skills from name.

Changed

• validate-skill.sh: [Not For] is now a required section.
• skill-template.md: updated with [Not For] section.

[v1.16] - 2026-05-21

Added

• Harness Engineering principles: Tool AI-fication Priority, Substitute Don't Mock, Environment-First, Minimum Runnable Subset, Scripted Verification.
• Machine Gates: Hallucination Gate, Sloppiness Gate, Overstepping Gate — enforceable via lint/test/hook/CI.
• Iron Rules: 8 baseline non-negotiable rules (knowledge offloading, no prompt magic, real files, guardrails).
• llms.txt: AI-searchable project summary for LLM discoverability.
• Per-directory AGENTS.md: Local MUST/MUST NOT/SHOULD rules for each core/ subdirectory.
• validate-skill.sh: Formal SKILL.md spec validator (frontmatter, sections, naming, file size, placeholders).
• Claude Code adapter rules migration: AGENTS.md → .claude/rules/*.md with path-scoped globs frontmatter.
• Gotchas in all 11 skills: Domain-specific failure points documented.
• CLI best practices in CLAUDE.md: /model, /compact, /context, /sandbox usage guidance.

Changed

• dev-builder: "SDK-First" → "Glue Code First" priority chain.
• evolution-engine: Generator/Optimizer recursion principle — the engine that evolves rules should itself be evolvable.
• [Anti-Rationalization Checklist] → [Gotchas: Anti-Rationalization] in dev-builder, code-review, bug-fixer.
• code-review: cross-session audit principle for complex reviews.
• feedback template: prompt_remediation field for reusable prevention prompts.

Fixed

• All 11 skills validated: 11 missing-section errors and 19 warnings fixed across design-maker, evolution-engine, feedback-writer, bug-fixer, code-review, dev-builder, dev-planner.

[v1.15] - 2026-05-20

Added

• Dependency graph: scripts/dependency-graph.ts for file-level import analysis and blast-radius.
• Dedicated AGENTS.md template for OpenCode (constraint-focused: tech stack, behavior boundaries, hard constraints).

Changed

• Exact version pinning: every dependency pinned to major.minor.patch — no ranges, no latest.
• dev-builder code review loop: passes affected_files from blast-radius to code-reviewer.

Fixed

• OpenCode control file: now uses dedicated AGENTS.md template instead of CLAUDE.md clone.

[v1.14.2] - 2026-05-20

Added

• forge-install: scripts/install.ts copies adapter bundles into user projects (pnpm forge-install <client> [--target dir] [--force])
• scripts/install.sh / scripts/install.ps1 thin wrappers for one-command install
• Windows: auto-applies settings.windows.json → settings.json on win32 (or --windows)
• Merge mode (--force) preserves existing feedback/ files and settings.local.json

[v1.14.1] - 2026-05-20

Added

• Vitest unit tests for scripts/sync.ts and scripts/dependency-graph.ts (scripts/__tests__/)
• pnpm test / vitest.config.ts in repo root

Fixed

• dependency-graph.ts: TypeScript named imports (import { x } from "./y") now resolve correctly

Changed

• package.json version 1.14.1; devDependencies pinned to exact patch versions (no ^)
• scripts/sync.ts and scripts/dependency-graph.ts export testable functions; CLI guarded with require.main === module
• DEV-PLAN.md: progress table, Phase 10 tests, 6 agents, OpenCode AGENTS.md, sync script docs

[v1.14] - 2026-05-19

Added

• Exact version pinning rule in CLAUDE.md: every dependency must be exact patch, no ranges
• Dedicated AGENTS.md template (core/templates/agents-template.md): constraint-focused format with tech stack, behavior boundaries, hard constraints
• Dependency graph (scripts/dependency-graph.ts): file-level import graph for blast-radius analysis. pnpm dep-graph build | affected | risk | stats

Changed

• OpenCode AGENTS.md now uses dedicated template instead of CLAUDE.md clone
• OpenCode CLAUDE.md removed (AGENTS.md is the only control file)
• sync.ts updated: AGENTS.md template for OpenCode, check-sync excluded from adapter sync
• dev-builder SKILL.md: Blast-Radar principle, dep-graph integration in review loop
• code-reviewer agent: accepts affected_files input from blast-radius analysis for focused review

[v1.13] - 2026-05-19

Added

• Planner sub-agent (core/agents/planner.md): dedicated agent for architecture design and Phase splitting, decoupled from implementer context
• Session handoff mechanism: handoff-template.md + check-handoff hook + dev-builder Step 5 to generate handoff before context reset, preventing lost progress across sessions
• Complexity gate for code-reviewer: change_complexity="simple" skips Stage 1, matching review depth to change scope
• Model version tracking in feedback-observer: records model version with feedback, enabling evolution engine to detect outdated rules

Changed

• Sub-Agent count 5→6, Hook count 9→10

[v1.12] - 2026-05-19

Fixed

• Missing detect-feedback-signal and auto-push hooks in ReqForge's own .claude/settings.json
• ReqForge self-development settings now at parity with adapter config

[v1.11] - 2026-05-19

Changed

• README and Product-Spec updated for test-writer, check-sync, and self-wired settings

[v1.10] - 2026-05-19

Added

• test-writer sub-agent (core/agents/test-writer.md): Vitest-based test generator for sync.ts and core utilities
• check-sync hook (core/hooks/check-sync sh+bat): PostToolUse detection of core/ vs adapters/ divergence
• ReqForge self-wiring: .claude/settings.json + settings.windows.json with all 6 hooks

Changed

• .claude/settings.local.json pruned from 65→32 lines, one-time commands replaced with wildcard patterns

[v1.9] - 2026-05-19

Added

• AI Only for Judgment Tasks: deterministic logic is plain code, not AI busywork
• Fail Loudly: uncertainty must be stated explicitly, never hidden
• Token Budget Awareness: check context headroom after each Task

[v1.8] - 2026-05-18

Added

• Feedback auto-scoring: failures automatically infer Skill scores (Precision/Coverage/Efficiency/Satisfaction)
• Hard-trigger evolution: check-evolution hook injects additionalContext, forcing evolution-runner dispatch on session init
• Structured failure context: dev-builder passes trigger_reason + current_skill + ai_action + failure_detail

Fixed

• Ratchet idling: feedback had records but no score data, preventing evolution proposals — now every failure produces scored data

[v1.7] - 2026-05-18

Changed

• CLAUDE.md trimmed from 309 lines to 59 lines — dispatch map only, details in core/docs/
• Detailed docs moved to core/docs/: file-structure.md, behavior-boundaries.md, memory-system.md, sub-agent-orchestration.md

Added

• Feedback auto-trigger on failure (compile errors, review failures, verification failures)
• Anti-Rationalization Checklist with "skip feedback recording" positive/negative examples

[v1.6] - 2026-05-18

Fixed

• Product-Spec.md cross-platform claim corrected — both .sh and .bat hooks provided
• .gitignore now excludes .claude/worktrees/

Added

• context-compaction hook entry in README hook table

[v1.5] - 2026-05-18

Added

• Agent Harness engineering foundation (Addy Osmani reference)
• Context compaction hook: Post-Tool-Use auto-archive old task-history entries
• Progressive disclosure: CLAUDE.md as dispatch map, SKILL.md loaded on demand
• Tool-call offloading: outputs >2000 lines written to temp files

Changed

• EVOLUTION.md: added Level 0 (Harness Foundation)
• Workflow section simplified — detailed flows live in SKILL.md

[v1.4] - 2026-05-18

Added

• Three-tier memory system: project-memory / decisions-log / task-history
• Traffic light behavior boundaries (green/yellow/red)
• Quick start mode: one-sentence project description
• memory-check hook: reminds to update memory after code changes

[v1.3] - 2026-05-16

Added

• Multi-role Sub-Agent isolation (OpenAI Symphony inspiration)
• Skill-as-prompt design (awesome-chatgpt-prompts inspiration)

[v1.2] - 2026-05-16

Changed

• File-first architecture: copy to project and use, no npm install required
• Simplified from monorepo to core + adapters directory structure

[v1.1] - 2026-05-16

Added

• Pluggable skill architecture with TDD discipline (superpowers inspiration)
• CLI initialization and incremental artifact management (OpenSpec inspiration)
• Multi-client adapter support

[v1.0] - 2026-05-16

• Initial release: Forge product development framework with core + multi-adapter architecture