create-ai-project 1.20.8 → 1.20.9

package/CHANGELOG.md

@@ -5,6 +5,50 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [1.20.9] - 2026-05-01
+
+### Added
+
+- **task-executor / -frontend: Fix Mode** (agents) — New optional `requiredFixes` and `incompleteImplementations` input parameters trigger Fix Mode. In Fix Mode the executor drives work from fix items instead of task file checkboxes, skips the `task_already_completed` gate, extends the allowed file list with each item's `file_path` (path) or `location` (parsed as `file[:line]`, file part only), and leaves task checkboxes unchanged. Outcomes recorded in `changeSummary`. Exit Gate validates Fresh Mode checkbox completion AND Fix Mode item coverage
+- **task-executor / -frontend: File Scope Constraint** (agents) — Allowed file list built explicitly from Target Files entries plus the task file, the referenced work plan, `Provides:` deliverable paths, and (in Fix Mode) fix item file paths. Modifications outside this set escalate as `out_of_scope_file` with `details.file_path` / `allowed_list` / `modification_reason`
+- **task-executor / -frontend: Investigation Notes persistence** (agents) — Step 2 appends observations from each Investigation Target to the task file's `## Investigation Notes` section before implementation. Exit Gate references the recorded notes as the consistency check baseline
+- **task-executor / -frontend: Phase Entry Gate / Step Completion Gates separation** (agents) — Phase Entry Gate now covers only true preconditions (skills loaded, task_file resolution input). Mid-flow conditions (task file content readable, uncompleted checkboxes, Target Files extracted, Investigation Targets read) move to Step 1 / Step 2 Completion Gates. New Exit Gate runs immediately before final JSON output
+- **task-executor / -frontend: granular `escalation_type` enum** (agents) — `task_file_not_found`, `task_already_completed`, `target_files_missing`, and `out_of_scope_file` added alongside existing types. Each carries a typed details payload via the common envelope and per-type contract table in Structured Response Specification
+- **security-reviewer: `suspected_risk` category and confidence-aware status routing** (agents) — Findings whose exploitability is uncertain or partially mitigated downgrade from `confirmed_risk` to `suspected_risk` (with `confidence` and rationale) instead of being discarded. High-confidence `suspected_risk` on primary input boundaries (auth, input boundaries, data persistence) routes the response to `blocked` for human investigation. `requiredFixes` remains code-level only
+- **security-reviewer: structured `requiredFixes` items** (agents) — Each entry now carries `{location, issue, fix}` with `location` parseable as `file[:line]` so downstream Fix Mode can extend its allowed file list correctly
+- **quality-fixer / -frontend: `filesModified` primary scope** (agents) — New optional input parameter passes the implementation step's write set as the primary scope for stub-detection. Falls back to `git diff HEAD` when absent. Prevents cross-task false positives when multiple commits accumulate in the working tree
+- **technical-designer / -frontend: Gate 0 / 1 / 2 / 3 ordering** (agents) — Mandatory Process Before Design Doc Creation reorganized into four serial gates (Inputs and Standards / Existing State Analysis / Design Decisions / Impact Documentation). Each subsection annotated with its gate number; subsection order matches gate order
+- **subagents-orchestration-guide: scale-aware Small Scale flow** (skills) — Small Scale (1-2 files) flow now declares that work-planner emits a single task-template-format file directly under `docs/plans/tasks/` and that path is what task-executor receives as `task_file`. Document Requirement table Small row carries the same contract
+- **work-planner: `scale` input parameter and Output Mode by Scale table** (agents) — `scale: small | medium | large` controls output: `small` produces a single task-template-format file under `docs/plans/tasks/` (no separate plan + decomposition); `medium` / `large` produce a plan-template-format work plan under `docs/plans/`. Step 7 branches accordingly
+- **All JSON-returning agents: Output Protocol section** (agents) — Single-sentence directive that the final message is exactly one JSON object matching the schema below (begins with `{`, ends with `}`, no code fence) and that progress text appears only in earlier messages
+- **Self-Validation [BLOCKING — before output] sections** (agents) — Replaces former `## Output Self-Check` blocks across JSON-returning agents. Includes return-to-step recovery rule when any item is unsatisfied. Applied to both EN and JA mirrors
+- **subagents-orchestration-guide: explicit subagent prompt construction rules** (skills) — Two new rules: subagents see only the Agent tool prompt and files they read (so paths, prior JSON, parameters, scope constraints must be injected explicitly); every `[placeholder]` in the call examples must be replaced with concrete values before invoking the Agent tool
+- **task-template: `## Investigation Notes` section** (skills) — New section between Investigation Targets and Implementation Steps. Implementation observations are appended here before implementation begins
+
+### Changed
+
+- **Recipes: Fix Mode wiring across the orchestration cycle** (commands) — `needs_revision` and `stub_detected` re-runs now route through executor Fix Mode with the same `task_file` and `requiredFixes[]` / `incompleteImplementations[]` array. quality-fixer invocations always pass the upstream step's `filesModified` array (with `git diff HEAD` fallback when omitted). Applied to build / front-build / implement / review / front-review / add-integration-tests
+- **Recipes: consolidated fix task file pattern for cross-task verifier outputs** (commands) — Post-implementation Fix cycle and security-reviewer needs_revision route now create a consolidated fix task file (`docs/plans/tasks/post-impl-fixes-YYYYMMDD.md` / `security-fixes-YYYYMMDD.md`) whose Target Files cover every file referenced by all verifier outputs (parsed as `file[:line]`, file part only). Verifier outputs normalized into a unified `requiredFixes[]` before invoking task-executor: `security-reviewer.requiredFixes[]` passes through; `code-verifier.discrepancies[]` converts to `{location, issue, fix}` with planned-target fallback when `codeLocation` is `null`
+- **task-executor / -frontend: progress checkbox update conditional on file existence** (agents) — Task file is always updated; work plan and overall design are updated only when the corresponding files exist. At small scale only the task file exists, so the other two are skipped without escalation
+- **technical-designer / -frontend: Fact Disposition prose to 4-row table** (agents) — Disposition selection criteria (preserve / transform / remove / out-of-scope) reorganized into a table with columns for "when to use", "rationale must state", and "review-time mismatch flag". All four categories preserved
+- **technical-designer / -frontend: Implementation Approach Decision and AC guidelines tables** (agents) — Strategy selection (Vertical / Horizontal / Hybrid) and Verification Strategy by `design_type` rendered as tables. Acceptance Criteria Include / Exclude lists rendered as a 7-row table
+- **quality-fixer / -frontend: Output Format compression** (agents) — Four full per-status JSON examples replaced by a common envelope plus per-status field table plus one minimal example. Mermaid Fix Determination Flow removed (duplicated Fix Execution Policy). Debugging Hints and Correct Fix Patterns consolidated into a single Anti-patterns table
+- **task-executor / -frontend: Escalation Response common trunk** (agents) — Six full JSON examples replaced by a common envelope plus per-type contract table plus one minimal example
+- **All agents: Task Registration wording** (agents) — "first 'Confirm skill constraints', final 'Verify skill fidelity'" replaced with "first task 'Map preloaded skills to applicable concrete rules' and final task 'Verify the mapped rules before final JSON' (or 'before producing the final output' for document-producing agents)"
+- **Recipes: Scope Boundary for Subagents replaces MANDATORY suffix** (commands) — Old "[SYSTEM CONSTRAINT] within X skill scope" suffix replaced with a positive scope-boundary block (operate within task scope and referenced files; use loaded skills; escalate when out of scope). Inline per command, no centralization
+- **subagents-orchestration-guide: Subagent I/O table updates** (skills) — task-executor row carries the full `escalation_type` enum and Fix Mode signal description; quality-fixer row documents `filesModified`; integration-test-reviewer / security-reviewer rows describe Fix Mode routing (security-reviewer routes through a consolidated fix task file)
+- **task-decomposer: English canonical headings in generated tasks** (agents) — Task Structuring step explicitly instructs use of task-template's English headings (`## Target Files`, `## Investigation Targets`, `## Implementation Steps (TDD: Red-Green-Refactor)`, `## Quality Assurance Mechanisms`, `## Operation Verification Methods`, `## Completion Criteria`) with a note that headings are not translated, so executor extraction stays consistent across language environments
+- **task-template: machine-read identifiers preserved as English** (skills) — Section headings and `Metadata:` keys (`Dependencies:`, `Provides:`, `Size:`, `Deliverable:`) kept as English so executor parsing remains language-independent. Body prose remains natural Japanese
+- **JA mirror: routing markers naturalized in frontmatter** (agents) — "Use when / Use PROACTIVELY / MUST BE USED PROACTIVELY" replaced with natural Japanese ("使用するシーン: / 積極的に使用するシーン: / 必ず積極的に使用するシーン:"); keyword pairs that aid routing preserved
+- **JA mirror: terminology unified** (commands, skills) — `step N` → ステップN; `cross-layer` → レイヤー横断; `Layer-Aware Agent Routing` → レイヤー別エージェントルーティング; `INVOKE` → 呼び出す; `consolidated fix task file` → 統合修正タスクファイル; `stub-detection` → 未完成実装検出
+- **Compression of six largest agent files** (agents) — task-executor.md 459 → 315; task-executor-frontend.md 428 → 303; quality-fixer.md 409 → 275; quality-fixer-frontend.md 453 → 316; technical-designer.md 471 → 429; technical-designer-frontend.md 471 → 409. All decision criteria, escalation_types, Fix Mode contract, Gate ordering, and JSON schemas preserved
+
+### Fixed
+
+- **subagents-orchestration-guide: acceptance-test-generator I/O row information loss** (skills) — Restored fields lost in an earlier translation pass: `budgetUsage`, `integration: path|null` (was `string`), `e2eAbsenceReason` enum values (`no_multi_step_journey | below_threshold_user_confirmed`), and the "verify files exist" instruction
+- **task-executor / -frontend: typo** (agents) — テスト嚗空化 → テスト空洞化 in Step 2 Quality Standard Violation Check
+- **JA mirror: stale references and English residue** (agents, commands, skills) — Task Registration old wording, "JSON format is mandatory" boilerplate, "Output Self-Check" sections, and "Final response is the JSON output" checklist items removed in favor of the new Output Protocol / Self-Validation [BLOCKING] structure. Awkward spacing around contract terms (`ルーティング先のexecutor`, `executorをFix Mode`, `task-template を用いて 統合`) corrected
+
 ## [1.20.8] - 2026-04-17
 
 ### Added

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "create-ai-project",
-  "version": "1.20.8",
+  "version": "1.20.9",
   "packageManager": "npm@10.8.2",
   "description": "TypeScript boilerplate with skills and sub-agents for Claude Code. Prevents context exhaustion through role-based task splitting.",
   "keywords": [