intentdna 1.5.16 → 1.5.18

package/dist/templates/flutter-rewrite.dna.yaml

@@ -134,25 +134,33 @@ context_files:
       - ".dna/specs/diagnosis-{{ARGUMENTS}}.md"
     surgeon:
       - ".dna/specs/diagnosis-{{ARGUMENTS}}.md"
+    fix_reviewer:
+      - "docs/behavior/{{ARGUMENTS}}.md"
+      - ".dna/specs/diagnosis-{{ARGUMENTS}}.md"
     test_runner:
       - "docs/behavior/{{ARGUMENTS}}.md"
+      - ".dna/specs/diagnosis-{{ARGUMENTS}}.md"
 
+verifier_policy:
+  allow_builtin_asserts:
+    - clean_working_tree
 
 roles:
   # ── Phase 0/3: behavior-lock 角色 ──
   scanner:
-    description: Scans v1 source code, outputs behavior document. Read-only.
+    description: Scans v1 source code and writes the behavior document artifact.
     tool_permissions:
-      allow: [Read, Grep, Glob, Bash]
-      deny: [Edit, Write, NotebookEdit]
+      allow: [Read, Grep, Glob, Write, Edit]
+      deny: [Bash, NotebookEdit]
     scope:
       read: ["**/*"]
-      write: []
+      write: ["docs/behavior/**"]
     instructions:
       - Scan ONE module only
       - "For each page, list user-visible actions: 动作 → 函数调用 → 返回值"
-      - Output as markdown with checkboxes
+      - Write the behavior document artifact under docs/behavior/
       - Do NOT read v2 code — only v1
+      - Do NOT write tests or app code
 
   test_writer:
     description: Reads behavior document, writes tests for all layers (logic + widget).
@@ -170,29 +178,7 @@ roles:
       - Run tests after writing — record baseline
       - Red tests are expected (implementation doesn't exist yet)
 
-  # ── Phase 4: rescue 角色 ──
-  investigator:
-    description: Traces broken chain in v1 and v2, identifies breakpoint. Read-only.
-    tool_permissions:
-      allow: [Read, Grep, Glob, Bash]
-      deny: [Edit, Write, NotebookEdit]
-    scope:
-      read: ["**/*"]
-      write: []
-    instructions:
-      - Trace the call chain in v1 for the target feature
-      - Find where v2 diverges
-      - Report with exact file paths and line numbers
-      - Never suggest code changes
-      - "If round > 1: review previous round's git diff first, judge if direction is correct"
-      - "If previous fix produced 0 red→green transitions: warn 'no progress'"
-      - "Categorize by severity and type:"
-      - "  CRITICAL: compile errors, import failures — blocks everything"
-      - "  HIGH-INFRA: mock incomplete causing test hang — blocks behavior verification, fix mock infrastructure first"
-      - "  HIGH-LOGIC: logic test failures (state/notifier) — behavior inconsistency"
-      - "  MEDIUM: widget test failures (rendering/navigation)"
-      - "hung tests ≠ failed tests. hung = mock infrastructure problem (needs mock fix), failed = behavior inconsistency (needs v2 code fix)"
-
+  # ── Phase 4: fix 角色 ──
   surgeon:
     description: Fixes breakpoints and builds missing layers by understanding v1 intent and rewriting in v2 style.
     tool_permissions:
@@ -244,10 +230,13 @@ roles:
       write: [".dna/specs/**"]
     instructions:
       - "REQUIRED FIRST: Read all context files listed in SKILL.md"
+      - "Treat docs/behavior/blocked_items.md as historical context only, never as the current run verdict"
       - "For each failing, skipped, and hung test: read test code + v1 impl + v2 impl"
       - "Classify each as BUG / UNIMPLEMENTED / INFRA / REMOVED / TEST_BUG (skipped tests are usually UNIMPLEMENTED)"
-      - "Output diagnosis spec with v1 code snippets + v2 current state"
-      - "DO NOT run tests. DO NOT edit code. Analysis only."
+      - "Output diagnosis spec with v1 code snippets + v2 current state + evidence paths"
+      - "If reanalyzing after REQUEST_REANALYSIS, update every reviewer sections_to_fix item and report sections_updated"
+      - "DO NOT include fix prescriptions, implementation order, call-site instructions, or Notes for Surgeon"
+      - "DO NOT run tests. DO NOT edit app code. Analysis only."
 
   analysis_reviewer:
     description: "Reviews analyzer output quality. Read-only."
@@ -261,7 +250,23 @@ roles:
       - "REQUIRED FIRST: Read all context files listed in SKILL.md"
       - "Verify analyzer spec: v1 references exist? Classifications sound?"
       - "Check missing: any failing test not covered?"
-      - "Output: APPROVE or REQUEST_REANALYSIS with specific feedback"
+      - "Verify diagnosis remains evidence-only: no fix prescriptions, implementation order, call-site instructions, or Notes for Surgeon"
+      - "Output a structured verdict block with verdict, artifact_reviewed, sections_to_fix, evidence_paths, confidence, and summary"
+
+  fix_reviewer:
+    description: "Reviews surgeon changes against approved diagnosis, v1 evidence, and v2 architecture. Read-only."
+    tool_permissions:
+      allow: [Read, Grep, Glob, Bash]
+      deny: [Edit, Write, NotebookEdit]
+    scope:
+      read: ["**/*"]
+      write: []
+    instructions:
+      - "Read diagnosis spec, behavior doc, and git diff before judging"
+      - "Verify changes are minimal and limited to the approved diagnosis classifications"
+      - "Verify v1 evidence was preserved and v2 patterns are followed"
+      - "Reject fake mocks, unrelated refactors, scope creep, or unapproved REMOVED handling"
+      - "Output structured verdict: APPROVE or REQUEST_CHANGES with issues and evidence paths"
 
   test_runner:
     description: "Runs tests and reports progress delta."
@@ -340,159 +345,6 @@ workflows:
             - type: git_commit
               description: "Behavior lock commit"
 
-  rescue:
-    name: Rescue
-    description: "Fix v2 module $ARGUMENTS — investigate, fix, review, verify, report. Max 10 rounds with convergence protection."
-    max_rounds: 10
-    convergence_rule: "2 consecutive rounds with 0 test progress (green count not increasing) → STOP. Output blocked items + analysis."
-    round_budget: "Max 5 files per round. Each round must produce at least 1 test transition (red/skip/hung → green), otherwise counted as no progress."
-    priority_order: |
-      Phase 1: Fix CRITICAL (compile errors) — unblocks everything
-      Phase 2: Fix HIGH-INFRA (mock infrastructure, make hung tests runnable) — unblocks behavior verification
-      Phase 3: Fix HIGH-LOGIC (logic tests, red → green) — behavior alignment
-      Phase 4: Fix MEDIUM (widget tests, red → green) — UI alignment
-      Complete each phase before moving to the next.
-    steps:
-      - id: investigate
-        role: investigator
-        description: "Run tests, assess current state, pick next targets by severity."
-        prompt: |
-          Round context:
-          - If round > 1: review previous round's git diff first
-          - If previous round had 0 test progress (no red→green or skip→green): warn "no progress" and consider changing approach
-
-          Run tests in {{test_path}}/$ARGUMENTS/ --timeout 30s. Categorize all non-passing tests by severity:
-          CRITICAL: compile errors, import failures — blocks everything, fix first
-          HIGH-INFRA: mock incomplete causing test hang (timed out) — blocks behavior verification, fix mock infrastructure
-          HIGH-LOGIC: logic test failures (state/notifier/service) — behavior inconsistency, fix after infra
-          MEDIUM: widget test failures (rendering/navigation) — fix after logic
-
-          IMPORTANT: hung ≠ failed. A test that times out (hung) = mock infrastructure problem, NOT a v2 behavior issue. Classify separately.
-
-          Pick highest severity batch. Trace: what does v1 do vs what does v2 do? Find the breakpoints.
-          Report findings and the plan for this round.
-        handoff:
-          produces:
-            - type: summary
-              description: "Investigation findings and breakpoint analysis"
-            - type: test_result
-              path: "{{test_path}}/$ARGUMENTS/"
-              description: "Current test state assessment"
-      - id: fix
-        role: surgeon
-        depends_on: [investigate]
-        description: "Fix identified issues. Max 5 files per round."
-        max_attempts: 3
-        on_fail: handoff
-        handoff_to: investigate
-        max_handoffs: 3
-        on_handoff_exhausted: skip
-        blocked_items_path: "docs/behavior/blocked_items.md"
-        checkpoints:
-          - assert: clean_working_tree
-            message: "Commit all changes before proceeding"
-        prompt: |
-          Fix the identified breakpoints. Rules:
-          - Read v1 intent in {{v1_path}}/, rewrite in v2 style in {{v2_path}}/ (not copy v1 verbatim)
-          - Maximum 5 files per round — if more needed, split the scope
-          - Run `flutter analyze` after each file change
-          - Same issue failed 3 times → STOP, report as blocked, do not retry
-          - Logic tests: implement notifier/state/service code
-          - Widget tests: copy widget from v1, change bindings (Obx→Consumer, Get.to→context.go), set up infra (ProviderScope, mock providers, GoRouter) if needed
-          - Run tests after each fix. Red→green or Skip→green = progress. Still failing = revert and re-analyze
-          - Commit: "rescue($ARGUMENTS): round N — what changed, why, which tests targeted"
-        handoff:
-          consumes:
-            - type: summary
-              from: investigate
-              description: "Investigation findings from investigate step"
-          produces:
-            - type: git_commit
-              description: "Rescue round commit"
-      - id: progress_check
-        role: investigator
-        depends_on: [fix]
-        description: "Check test progress after surgeon's fix."
-        prompt: |
-          Run tests. Compare green count with previous round.
-          If green count increased: PROGRESS — proceed to review.
-          If green count unchanged or decreased: NO_PROGRESS — record what surgeon tried
-          and why it didn't work, for the experience chain.
-        handoff:
-          consumes:
-            - type: git_commit
-              from: fix
-              description: "Fix commit from surgeon"
-          produces:
-            - type: summary
-              description: "Progress check result (PROGRESS or NO_PROGRESS)"
-      - id: review
-        role: investigator
-        depends_on: [progress_check]
-        description: "Read-only review of surgeon's changes."
-        prompt: |
-          Review surgeon's git diff (read-only, do NOT modify any files):
-          1. Are changes minimal? No unnecessary files touched?
-          2. Does the code match v2 patterns (Riverpod, GoRouter)?
-          3. Are mocks correct (not faked just to make tests pass)?
-          4. Any new issues introduced?
-
-          Verdict: APPROVE → proceed to verify
-          Verdict: REQUEST_CHANGES → describe specific problems. Next round's investigate step will include this feedback.
-        handoff:
-          consumes:
-            - type: summary
-              from: progress_check
-              description: "Progress check result"
-            - type: git_commit
-              from: fix
-              description: "Committed fix from surgeon"
-          produces:
-            - type: summary
-              description: "Review verdict (APPROVE or REQUEST_CHANGES)"
-      - id: verify
-        role: investigator
-        depends_on: [review]
-        description: "Independent test verification + regression check."
-        prompt: |
-          Run tests independently (do not trust surgeon's reported results):
-          1. `flutter test {{test_path}}/$ARGUMENTS/ --timeout 30s` (per-test safety net; hung = mock infra issue)
-          2. `flutter analyze` (compilation check)
-          3. Check for regressions in core module tests if applicable
-
-          Record test delta vs previous round.
-          Each acceptance criterion: VERIFIED / PARTIAL / MISSING
-          Verdict: PASS or FAIL
-        handoff:
-          consumes:
-            - type: summary
-              from: review
-              description: "Review verdict"
-          produces:
-            - type: test_result
-              path: "{{test_path}}/$ARGUMENTS/"
-              description: "Verified test results"
-      - id: report
-        role: investigator
-        depends_on: [verify]
-        description: "Round summary with convergence judgment."
-        prompt: |
-          Summary:
-          1. Test delta: +N green, -M red, ±K skipped vs last round
-          2. Remaining tests by category (logic vs widget vs platform)
-          3. Convergence check:
-             - If 2 consecutive rounds with no test progress → STOP, output blocked items + analysis
-             - If all green → "Module $ARGUMENTS rescue complete"
-             - Otherwise → "Run /rescue $ARGUMENTS to continue (round N+1 of max 10)"
-          4. If review verdict was REQUEST_CHANGES: include the specific feedback for next round
-        handoff:
-          consumes:
-            - type: test_result
-              from: verify
-              description: "Verified test results from verify step"
-          produces:
-            - type: summary
-              description: "Round summary with test delta and convergence status"
 
   core-align:
     name: Core Align
@@ -517,16 +369,19 @@ workflows:
 
           For module $ARGUMENTS, analyze failing tests in baseline:
           1. Read behavior doc: docs/behavior/$ARGUMENTS.md
-          2. Read test results from last behavior-lock run
-          3. For each failing, skipped, and hung test:
+          2. Read test results from last behavior-lock run. If no structured baseline artifact exists, state exactly which file(s) you used as the baseline source.
+          3. Treat docs/behavior/blocked_items.md as historical context only. It must never be used as the current run verdict.
+          4. For each failing, skipped, and hung test:
              a. Read test code (what behavior does it expect?)
              b. Read v1 implementation (how did v1 do this?)
              c. Read v2 current state (what's missing/wrong?)
              d. Classify: BUG / UNIMPLEMENTED / INFRA / REMOVED / TEST_BUG
-          4. Write .dna/specs/diagnosis-$ARGUMENTS.md with:
+          5. Write .dna/specs/diagnosis-$ARGUMENTS.md with:
              - For each failure: v1 code snippet + v2 current state + classification
-             - NO fix prescriptions (surgeon decides how to implement)
+             - Evidence paths for every classification
+             - NO fix prescriptions, implementation order, call-site instructions, or "Notes for Surgeon"
              - NO running tests (analysis only)
+          6. If this is a reanalysis after REQUEST_REANALYSIS, read the reviewer verdict first, update every section listed in sections_to_fix plus the summary table, and report sections_updated in your final response. Reanalysis is not successful unless the spec artifact is rewritten.
         handoff:
           produces:
             - type: file
@@ -545,10 +400,23 @@ workflows:
           1. All failing tests from baseline covered?
           2. v1 code references actually exist at claimed locations?
           3. Classifications reasonable? (INFRA not mistaken for BUG)
-          4. Enough detail for surgeon to act?
-
-          APPROVE → diagnosis complete
-          REQUEST_REANALYSIS → list specific gaps, back to analyze
+          4. Enough evidence detail for surgeon to act without adding fix prescriptions?
+          5. Diagnosis remains evidence-only: no implementation order, call-site instructions, or "Notes for Surgeon"?
+
+          Output exactly one structured verdict block:
+          ```json
+          {
+            "verdict": "APPROVE" | "REQUEST_REANALYSIS",
+            "artifact_reviewed": ".dna/specs/diagnosis-$ARGUMENTS.md",
+            "sections_to_fix": ["section ids or titles; empty when approved"],
+            "evidence_paths": ["paths that justify the verdict"],
+            "confidence": "high" | "medium" | "low",
+            "summary": "brief reason"
+          }
+          ```
+
+          APPROVE → diagnosis complete.
+          REQUEST_REANALYSIS → sections_to_fix must be passed back to analyze. docs/behavior/blocked_items.md is historical context only and must not be treated as the current run verdict.
         handoff:
           consumes:
             - type: file
@@ -560,36 +428,32 @@ workflows:
 
   fix:
     name: Fix
-    description: "Read diagnosis spec and fix v2 code for module $ARGUMENTS with reflection gate"
+    description: "Read diagnosis spec and fix v2 code for module $ARGUMENTS with review and verification gates"
     steps:
       - id: fix_bugs
         role: surgeon
-        max_attempts: 3
-        on_fail: handoff
-        handoff_to: fix_bugs
-        max_handoffs: 3
-        on_handoff_exhausted: skip
         blocked_items_path: "docs/behavior/blocked_items.md"
         description: "Fix issues according to diagnosis spec classifications"
         prompt: |
-          Read .dna/state/workflow/experience.md first if it exists. It contains previous failed approaches — do NOT repeat them.
+          This invocation is one fix round. If prior verify_report or experience_chain context is provided, read it first and do not repeat failed approaches.
 
           Read context + diagnosis spec first.
 
-          Process issues in priority order: INFRA → BUG → UNIMPLEMENTED → TEST_BUG
-          (REMOVED skipped — needs user confirmation)
+          Count diagnosis classifications before editing. If UNIMPLEMENTED count is greater than 100, this round is implementation-first: fix UNIMPLEMENTED items by priority only, and leave TEST_BUG / BUG / INFRA / REMOVED for later rounds unless they block implementation.
+
+          Otherwise process issues in priority order: INFRA → BUG → TEST_BUG → UNIMPLEMENTED.
+          REMOVED is always skipped unless the user explicitly approved handling it.
 
           For each issue:
           - BUG: read v1 file, edit v2 (use Riverpod per refactoring-workflow-v2.md)
-          - UNIMPLEMENTED: read v1 impl, implement in v2 style, remove skip marker from test
-          - INFRA: edit test_helpers only, do NOT touch v2/lib
           - TEST_BUG: re-read v1, edit test to match v1 behavior
+          - INFRA: edit test_helpers only, do NOT touch v2/lib
+          - UNIMPLEMENTED: read v1 impl, implement in v2 style, remove skip marker from test
 
-          After each fix: run tests ONCE for that specific file.
+          Continue within the round while changes are coherent and reviewable. If the remaining work is large, stop at a natural boundary and leave the rest for the next fix round.
 
-          If experience_chain is non-empty:
-          - You have tried approaches that failed
-          - Read them, use a DIFFERENT approach
+          After each fix: run tests ONCE for that specific file.
+          Commit only the round's coherent changes, with message including what changed, why, and which tests were targeted.
         handoff:
           consumes:
             - type: file
@@ -599,24 +463,87 @@ workflows:
             - type: git_commit
               description: "Fix commit"
 
-      - id: progress_check
-        role: test_runner
+      - id: review_changes
+        role: fix_reviewer
         depends_on: [fix_bugs]
+        description: "Review surgeon changes before verification"
+        prompt: |
+          Read diagnosis spec, behavior doc, and git diff.
+
+          Review only the surgeon's changes:
+          1. Are changes limited to BUG / UNIMPLEMENTED / INFRA / TEST_BUG items in the approved diagnosis?
+          2. Is REMOVED handling absent unless explicitly approved by user?
+          3. Does implementation preserve v1 behavior and follow v2 architecture?
+          4. Are test infra changes honest, not fake mocks?
+          5. Is there unrelated refactor or scope creep?
+
+          Output exactly one structured verdict block:
+          ```json
+          {
+            "verdict": "APPROVE" | "REQUEST_CHANGES",
+            "issues": ["specific issue with evidence path; empty when approved"],
+            "evidence_paths": ["paths that justify the verdict"],
+            "failure_reason": "why changes need another round; empty when approved",
+            "next_round_focus": "specific priority/classification/test file for the next fix invocation; empty when approved",
+            "summary": "brief reason"
+          }
+          ```
+
+          REQUEST_CHANGES → include failure_reason and concrete next_round_focus for the next fix invocation.
+        handoff:
+          consumes:
+            - type: git_commit
+              from: fix_bugs
+              description: "Fix commit from surgeon"
+            - type: file
+              path: ".dna/specs/diagnosis-$ARGUMENTS.md"
+              description: "Diagnosis spec"
+            - type: file
+              path: "docs/behavior/$ARGUMENTS.md"
+              description: "Behavior document"
+          produces:
+            - type: summary
+              description: "Review verdict (APPROVE or REQUEST_CHANGES)"
+
+      - id: verify_report
+        role: test_runner
+        depends_on: [review_changes]
         description: "Verify progress and output report"
         prompt: |
-          Run full module test suite for $ARGUMENTS.
-          Compare with diagnosis spec baseline.
+          Read the review verdict and diagnosis spec first.
+
+          If the review verdict is REQUEST_CHANGES, set Verdict: BLOCKED, set Failure reason: BLOCKED_BY_REVIEW plus reviewer details, and do not claim verification.
+
+          If the review verdict is APPROVE:
+          - Run full module test suite for $ARGUMENTS.
+          - Compare with diagnosis spec baseline.
 
           Output report:
+          - Verdict: PASS / CONTINUE / REQUEST_CHANGES / BLOCKED
+          - Review: APPROVE / REQUEST_CHANGES
           - Fixed: N tests now green
           - New red: M tests that regressed
           - Blocked: K tests skipped (see blocked_items.md)
           - Remaining: R tests still failing
+          - Failure reason: why progress stopped, if any
+          - Next round focus: the priority/classification/test file to continue with, if verdict is CONTINUE or REQUEST_CHANGES
+
+          Verdict meanings:
+          - PASS: all approved diagnosis items are fixed or intentionally skipped
+          - CONTINUE: this round made progress and remaining approved items should continue in the next fix invocation
+          - REQUEST_CHANGES: review or verification found issues in this round's changes
+          - BLOCKED: user confirmation is needed or repeated no-progress prevents safe continuation
         handoff:
           consumes:
+            - type: summary
+              from: review_changes
+              description: "Review verdict from fix_reviewer"
             - type: git_commit
               from: fix_bugs
               description: "Fix commit from surgeon"
+            - type: file
+              path: ".dna/specs/diagnosis-$ARGUMENTS.md"
+              description: "Diagnosis spec"
           produces:
             - type: summary
-              description: "Progress report with test delta"
+              description: "Verification report with test delta"

package/dist/templates/full-pipeline.dna.yaml

@@ -100,6 +100,10 @@ roles:
 
 contexts: {}
 
+verifier_policy:
+  allow_command_prefixes:
+    - npm test
+    - npx tsc --noEmit
 
 workflow:
   name: dev-pipeline

package/dist/templates/mobile-dev.dna.yaml

@@ -52,6 +52,11 @@ genes:
 
 contexts: {}
 
+verifier_policy:
+  allow_command_prefixes:
+    - dart analyze
+    - flutter test
+
 roles:
   mobile_dev:
     description: Mobile developer

package/hooks/hooks.json

@@ -6,7 +6,7 @@
     "SubagentStop": [{ "matcher": "*", "hooks": [{ "type": "command", "command": "dna-hook SubagentStop", "timeout": 3 }] }],
     "PreCompact": [{ "matcher": "*", "hooks": [{ "type": "command", "command": "dna-hook PreCompact", "timeout": 3 }] }],
     "Notification": [{ "matcher": "*", "hooks": [{ "type": "command", "command": "dna-hook Notification", "timeout": 3 }] }],
-    "Stop": [{ "matcher": "*", "hooks": [{ "type": "command", "command": "dna-hook Stop", "timeout": 3 }] }],
+    "Stop": [{ "matcher": "*", "hooks": [{ "type": "command", "command": "dna-hook Stop", "timeout": 45 }] }],
     "SessionStart": [{ "matcher": "*", "hooks": [{ "type": "command", "command": "dna-hook SessionStart", "timeout": 5 }] }]
   }
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "intentdna",
-  "version": "1.5.16",
+  "version": "1.5.18",
   "description": "Intent DNA — Declarative policy layer for AI agent behavior",
   "type": "module",
   "main": "dist/index.js",