maestro-flow 0.3.39 → 0.3.41

package/.codex/skills/maestro-ralph/SKILL.md

@@ -279,9 +279,15 @@ Write `.workflow/.maestro/ralph-{YYYYMMDD-HHmmss}/status.json`:
 }
 ```
 
-### 1.7: Initialize plan + confirm
+### 1.7: Initialize tracking + confirm
 
 ```
+// Goal = outer constraint — ensures entire lifecycle chain completes
+functions.create_goal({
+  objective: `Ralph lifecycle: ${lifecycle_position} → milestone-complete | ${steps.length} steps (${decision_count} decisions) | quality=${quality_mode}`
+})
+
+// Plan = inner tracking — sub-step progress
 functions.update_plan({
   explanation: "Ralph lifecycle: {position} → milestone-complete",
   plan: steps.map(step => ({ step: stepLabel(step), status: "pending" }))
@@ -355,8 +361,10 @@ STATUS: proceed | fix | escalate
 REASON: 一句话解释
 GAP_SUMMARY: 问题描述（fix/escalate 时填写）
 CONFIDENCE: high | medium | low
+CONFIDENCE_SCORE: 0-100（从结果文件中读取置信度分数，无则估算）
+WEAKEST_DIMENSION: 最弱维度名称
 ---END---
-CONSTRAINTS: 只评估 | STATUS 三选一 | retry ${meta.retry_count}/${meta.max_retries} 达上限必须 escalate" --role analyze --mode analysis`,
+CONSTRAINTS: 只评估 | STATUS 三选一 | 置信度 < 60% 倾向 fix | retry ${meta.retry_count}/${meta.max_retries} 达上限必须 escalate" --role analyze --mode analysis`,
   yield_time_ms: 30000,
   max_output_tokens: 6000
 })
@@ -366,17 +374,25 @@ CONSTRAINTS: 只评估 | STATUS 三选一 | retry ${meta.retry_count}/${meta.max
 
 **Parse verdict** (on callback):
 ```
-Extract STATUS / REASON / GAP_SUMMARY / CONFIDENCE from output.
+Extract STATUS / REASON / GAP_SUMMARY / CONFIDENCE / CONFIDENCE_SCORE / WEAKEST_DIMENSION from output.
 If parse fails → fallback: STATUS = "fix", GAP_SUMMARY = generic
+
+Confidence-based verdict adjustment (after parse, before apply):
+  If CONFIDENCE_SCORE < 60 AND STATUS == "proceed":
+    → Override to "fix", REASON += " (置信度不足: {score}%，{weakest} 需加强)"
+  If CONFIDENCE_SCORE > 95 AND STATUS == "fix" AND retry_count > 0:
+    → Suggest "proceed" override, REASON += " (置信度充分: {score}%，建议通过)"
 ```
 
+**Confidence-aware evaluation**: Before delegating, check if artifact contains confidence section (added by downstream commands). If found, include `已有置信度评估: 整体 {overall}%, 最弱维度: {weakest} ({score}%)` in delegate prompt as additional signal.
+
 **Apply verdict:**
 
 | Mode | Behavior |
 |------|----------|
-| `-y` (auto_mode) | Follow verdict directly, no user prompt |
-| Interactive + confidence == "high" | Display recommendation, prompt user with options below |
-| Interactive + confidence != "high" | Display recommendation **with warning**, prompt user with options below |
+| `-y` (auto_mode) | Follow adjusted verdict directly, no user prompt |
+| Interactive + confidence_score >= 80 | Display recommendation with confidence, prompt user |
+| Interactive + confidence_score < 80 | Display recommendation **with confidence warning**, prompt user |
 
 Interactive prompt (via `request_user_input`):
 ```json
@@ -577,8 +593,13 @@ functions.update_plan({
   explanation: "Ralph lifecycle complete",
   plan: steps.map(step => ({ step: stepLabel(step), status: "completed" }))
 })
+
+// Release goal constraint — only on true completion
+functions.update_goal({ status: "complete" })
 ```
 
+**Note**: Pause/escalate paths (`post-debug-escalate` STOP, session pause) do NOT call `update_goal` — goal stays running for resume.
+
 Display:
 ```
 ============================================================
@@ -665,9 +686,11 @@ Rules:
 - [ ] Conditional steps evaluated at decision time (coverage threshold)
 - [ ] buildSkillCall() completes arg enrichment + auto flag, CSV contains full commands
 - [ ] Quality-gate decisions delegate-evaluated via `maestro delegate --role analyze`
-- [ ] Delegate verdict parsed: STATUS / REASON / GAP_SUMMARY / CONFIDENCE
-- [ ] `-y` mode: auto-follow delegate verdict, no STOP (except post-debug-escalate)
-- [ ] Interactive mode: display recommendation + request_user_input with override
+- [ ] Delegate verdict parsed: STATUS / REASON / GAP_SUMMARY / CONFIDENCE / CONFIDENCE_SCORE / WEAKEST_DIMENSION
+- [ ] Confidence-based verdict adjustment applied (< 60% bias fix, > 95% bias proceed)
+- [ ] Artifact confidence sections read when available as additional signal
+- [ ] `-y` mode: auto-follow adjusted verdict, no STOP (except post-debug-escalate)
+- [ ] Interactive mode: display recommendation with confidence score + request_user_input with override
 - [ ] Delegate failure fallback: treat as "fix" verdict
 - [ ] passed_gates[] tracks passed quality gates, skips re-runs in retry loops
 - [ ] passed_gates cleared when code changes (fix-loop inserts execute step)

package/.codex/skills/quality-auto-test/SKILL.md

@@ -385,6 +385,9 @@ OUTER LOOP (max_iter iterations):
     Analyze: pass rate delta, failure clusters, strategy effectiveness
     Append to reflection-log.md
 
+    **Test confidence scoring** (at each REFLECT step):
+       Dimensions (5): scenario_coverage, test_quality, diagnostic_accuracy, strategy_effectiveness, infrastructure_fitness. Factors (weights): completeness(.30), pass_rate_trend(.25), classification_accuracy(.20), coverage_breadth(.15), consistency(.10). Enhanced convergence: BOTH pass_rate ≥ threshold AND confidence ≥ 60%. Add confidence to `report.json`.
+
   ADJUST (Adaptive Strategy):
     IF startStrategy provided AND iteration == 1: use startStrategy as initial
     OTHERWISE auto-select:
@@ -540,6 +543,9 @@ CSV Session: .tests/auto-test/.csv-session/
 - [ ] discoveries.ndjson append-only throughout
 - [ ] Cross-layer context propagation via prev_context column
 - [ ] Iteration engine ran (inner: test_defect fix, outer: strategy adjust)
+- [ ] Test confidence scored per iteration with 5-dimension factor model
+- [ ] Convergence check includes confidence >= 60% alongside pass_rate
+- [ ] Confidence section added to report.json
 - [ ] state.json, report.json, reflection-log.md written
 - [ ] If spec: traceability.md produced
 - [ ] If failures: issues auto-created in issues.jsonl

package/.codex/skills/quality-debug/SKILL.md

@@ -257,6 +257,10 @@ spawn_agents_on_csv({
 
 2. **Generate context.md**: Debug report with summary (mode, hypothesis/confirmed/fixed/verified counts), per-hypothesis results (hypothesis, evidence for/against, findings, status), per-fix results (fix applied, verified, findings), aggregated root causes, and next steps.
 
+2b. **Debug confidence scoring**:
+
+   Dimensions (4): hypothesis_quality, evidence_completeness, root_cause_isolation, fix_confidence. Factors (weights): evidence_depth(.30), evidence_strength(.25), coverage_breadth(.20), reproduction(.15), consistency(.10). Map to legacy: <40% = low, 40-70% = medium, >70% = high. Append confidence assessment to context.md.
+
 3. **UAT update** (if --from-uat): Update `uat.md` gaps with `root_cause`, `fix_direction`, `affected_files` for confirmed hypotheses.
 
 4. **Issue update**: If `issues.jsonl` exists, update matching issues with status `diagnosed`, add `context.suggested_fix` and `context.notes`.
@@ -296,7 +300,7 @@ spawn_agents_on_csv({
 
 | Type | Dedup Key | Data Schema | Description |
 |------|-----------|-------------|-------------|
-| `root_cause` | `data.location` | `{location, cause, severity, confidence}` | Confirmed root cause |
+| `root_cause` | `data.location` | `{location, cause, severity, confidence_score, confidence_factors}` | Confirmed root cause |
 | `hypothesis_evidence` | `data.hypothesis+data.location` | `{hypothesis, location, type, conclusion}` | Evidence for/against hypothesis |
 | `affected_component` | `data.component` | `{component, files[], impact}` | Component affected by bug |
 | `reproduction_path` | `data.trigger` | `{trigger, steps[], frequency}` | Bug reproduction path |
@@ -333,6 +337,8 @@ echo '{"ts":"<ISO>","worker":"{id}","type":"root_cause","data":{"location":"src/
 - [ ] Refuted/inconclusive hypotheses correctly skip wave 2 fix tasks
 - [ ] Wave 2 fixes attempted only for confirmed hypotheses
 - [ ] context.md produced with diagnosis summary
+- [ ] Multi-factor confidence scored per hypothesis replacing simple high/medium/low
+- [ ] Confidence assessment appended to context.md
 - [ ] UAT gaps updated (if --from-uat)
 - [ ] Issues updated with diagnosis results
 - [ ] discoveries.ndjson append-only throughout

package/.codex/skills/quality-test/SKILL.md

@@ -434,6 +434,12 @@ Options:
 If re-verify passes: update uat.md gaps as resolved, report success.
 If gaps remain after 2 iterations: report remaining, suggest manual intervention.
 
+### Step 12.5: UAT Confidence Scoring
+
+Dimensions (4): scenario_coverage, diagnostic_depth, observation_quality, closure_completeness. Factors (weights): requirements_mapped(.30), observation_specificity(.25), user_validation(.20), diagnostic_depth(.15), consistency(.10). Append confidence summary to `uat.md`.
+
+**Readiness gate** (before final report): Block if scenario_coverage < 40% or any blocker-severity gap without diagnosis.
+
 ### Step 13: Report
 
 ```
@@ -490,6 +496,9 @@ Files:
 - [ ] test-results.json and coverage-report.json written
 - [ ] index.json uat fields updated
 - [ ] Artifact registered in state.json
+- [ ] UAT confidence scored with 4-dimension factor model
+- [ ] Readiness gate checked before final report
+- [ ] Confidence summary appended to uat.md
 - [ ] If issues: diagnosis.csv built, spawn_agents_on_csv executed per gap cluster
 - [ ] Gaps updated with root_cause, fix_direction, affected_files
 - [ ] Gap-fix loop triggered if --auto-fix (max 2 iterations)

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "maestro-flow",
-  "version": "0.3.39",
+  "version": "0.3.41",
   "description": "Workflow orchestration CLI with MCP endpoint support and extensible architecture",
   "type": "module",
   "imports": {

package/workflows/analyze.md

@@ -252,6 +252,10 @@ Re-read original User Intent from discussion.md header. Check each intent item a
 
 Append initial Intent Coverage Check to discussion.md.
 
+**Step 4.6: Baseline Confidence Scoring**
+
+Dimensions = the 6 analysis dimensions. Factors (weights): findings_depth(.30), evidence_strength(.25), coverage_breadth(.20), user_validation(.15), consistency(.10). Score each factor per dimension from Round 1 results. Append baseline confidence table to discussion.md. Thresholds: <60% 继续深入 | 60-80% 可选 | 80-95% 接近收敛 | >95% 建议收敛.
+
 ### Step 5: Interactive Discussion Loop
 
 Max 5 rounds. Each round follows this sequence:
@@ -301,7 +305,19 @@ Re-read original User Intent. Check each item:
 
 If ❌ or ⚠️ items exist → proactively surface to user at start of next round.
 
-**Auto mode (-y)**: auto-deepen for up to 3 rounds, then synthesize.
+**5.8: Re-score Confidence** (every round):
+Re-evaluate factors per dimension. Show delta: `Confidence: {prev}% → {current}% ({±N%}), {weakest_dim} 仍需深入`
+
+**5.9: Quality Mechanisms**:
+- **Pressure Pass** (mandatory ≥1 before Step 6): highest-confidence finding → pressure ladder (evidence demand → assumption probe → boundary/tradeoff → root cause check). Record under `#### 压力测试`.
+- **Devil's Advocate**: dimension > 0.7 → challenge "如果 [finding] 不成立？" (once per dimension)
+- **Scope Minimizer**: findings > 5 + scope expanding → "最小可行结论集？"
+- **Stall Detection**: delta < 5% for 2 consecutive rounds → "分析可能停滞，建议切换方向或收敛"
+
+**5.10: Pre-Synthesis Readiness Gate** (on "分析完成"):
+Block if: ❌ items without deferral | any dimension < 40% | no pressure pass | unresolved contradictions. If blocked → AskUserQuestion: 补充后继续 or 忽略风险并继续 (record `residual_risks[]`).
+
+**Auto mode (-y)**: auto-deepen ≤3 rounds, readiness gate auto-overrides with residual risk recording.
 
 ### Step 6: Six-Dimension Scoring
 
@@ -318,9 +334,11 @@ Using all exploration findings, discussion insights, and user feedback, score ac
 
 Each dimension scored with specific evidence (code refs, data points from exploration).
 
+Each 1-5 score justified by confidence factors from Step 4.6/5.8. Include per-dimension confidence % alongside the score.
+
 Build probability-impact risk matrix from identified risks.
 
-Formulate Go/No-Go/Conditional recommendation with confidence level.
+Formulate Go/No-Go/Conditional recommendation with overall confidence %. Write confidence summary (per-dimension scores + overall + pressure pass result + residual risks) to analysis.md.
 
 ### Step 7: Synthesis & Conclusion
 
@@ -673,6 +691,10 @@ Replaceable blocks (overwritten each round):
 - At least 2 alternatives compared with tradeoffs
 - Go/No-Go/Conditional recommendation with confidence level
 - Code references included where relevant (file paths, line numbers)
+- Confidence tracking initialized and re-scored each round
+- Readiness gate checked before synthesis (Step 5.10)
+- Pressure pass completed ≥ 1 time before Step 6
+- Confidence summary with factor decomposition in analysis.md
 
 **Both modes (full + quick):**
 - context.md written with all decisions classified as Locked/Free/Deferred

package/workflows/auto-test.md

@@ -498,6 +498,18 @@ END OUTER
 
 ---
 
+### Step 7.5: Test Confidence Scoring
+
+Scored after each REFLECT step. Dimensions (5): scenario_coverage, test_quality, diagnostic_accuracy, strategy_effectiveness, infrastructure_fitness. Factors (weights): completeness(.30), pass_rate_trend(.25), classification_accuracy(.20), coverage_breadth(.15), consistency(.10). Append confidence table to reflection-log.md.
+
+**Enhanced Convergence**: pass_rate ≥ 95% AND confidence ≥ 60% → converged. pass_rate ≥ 95% BUT confidence < 60% → continue (tests may be weak). max_iter reached or all failures = code_defect → Step 8.
+
+**Quality mechanisms**: Pressure Pass (before Step 8) — select 2-3 passing tests from highest-pass-rate layer, verify they exercise real behavior (not mock-only, non-trivial assertions). Devil's Advocate — pass_rate > 80% → challenge assertion specificity, error path coverage, mock over-reliance. Stall Detection — delta < 5% for 2 iterations + pass_rate flat → force Reflective strategy.
+
+**Readiness Gate** (before Step 8, skip if max_iter=1): scenario_coverage < 40% | no pressure pass | diagnostic_accuracy < 40% | unclassified failures. If blocked → force one additional iteration. Add confidence section to report.json.
+
+---
+
 ### Step 8: Complete & Write Artifacts
 
 1. Update session state:

package/workflows/brainstorm.md

@@ -331,7 +331,7 @@ Each agent receives:
 
 ### Step 5: Synthesis Integration (Auto Mode)
 
-Six sub-phases producing feature specs from cross-role analysis:
+Seven sub-phases producing feature specs from cross-role analysis:
 
 **Sub-phase 1: Discovery**
 - Detect session, validate analysis files exist
@@ -349,6 +349,16 @@ Six sub-phases producing feature specs from cross-role analysis:
 - Output: `enhancement_recommendations` (EP-001, EP-002, ...) + `feature_conflict_map` (per-feature consensus/conflicts/cross_refs)
 - Conflict resolution quality: actionable, justified ("because...tradeoff:..."), scoped, confidence-tagged ([RESOLVED]|[SUGGESTED]|[UNRESOLVED])
 
+**Sub-phase 3B: Brainstorm Confidence Scoring**
+
+Compute after cross-role analysis, before user interaction. Dimensions (5): role_coverage, cross_role_consistency, feature_completeness, spec_quality, design_feasibility. Factors (weights): analysis_depth(.30), evidence_strength(.25), coverage_breadth(.20), user_validation(.15), consistency(.10).
+
+Scoring points: per-role completion → `role_coverage`; after cross-role → `cross_role_consistency`; after user interaction → `user_validation`.
+
+**Quality mechanisms**: Pressure Pass (before spec gen) — verify claims backed by ≥2 role analyses. Devil's Advocate — challenge when dimension > 0.7. Conflict Detection (replaces stall) — >3 `[UNRESOLVED]` conflicts → challenge.
+
+**Readiness Gate** (before Sub-phase 4): Block if role_coverage < 100% | cross_role_consistency < 40% | feature_completeness below intent count | no pressure pass. If blocked → AskUserQuestion: fill gaps or proceed with residual risks. Append confidence summary to synthesis-changelog.md.
+
 **Sub-phase 4: User Interaction**
 - Enhancement selection: AskUserQuestion (multiSelect=true, batched by 4)
 - Clarification questions: 9-category taxonomy scan, AskUserQuestion (single-select, multi-round)

package/workflows/debug.md

@@ -137,7 +137,7 @@ For each cluster, spawn concurrently as general-purpose agent (`run_in_backgroun
 
 - **Input**: cluster name, phase, all gaps (test_id, truth, reason, severity). Mode: `symptoms_prefilled`.
 - **Process**: read source files, form 2-3 hypotheses per gap ranked by likelihood, search code for evidence, log each as NDJSON line, confirm/refute.
-- **Output per gap**: `root_cause`, `fix_direction`, `affected_files` (file:line), `confidence` (high/medium/low), `evidence` summary.
+- **Output per gap**: `root_cause`, `fix_direction`, `affected_files` (file:line), `confidence` (multi-factor; also include legacy `confidence_level`: high/medium/low for backward compat), `evidence` summary.
 - **Files**: `{debug_dir}/evidence-{cluster_slug}.ndjson`, `{debug_dir}/understanding-{cluster_slug}.md`
 
 All agents run concurrently. Collect all results.
@@ -198,7 +198,7 @@ For each agent result, extract:
 - root_cause per gap
 - fix_direction per gap
 - affected_files per gap
-- confidence level
+- confidence (multi-factor, see 7.1)
 - evidence summary
 
 Build unified diagnosis:
@@ -215,14 +215,23 @@ Build unified diagnosis:
           "root_cause": "...",
           "fix_direction": "...",
           "affected_files": ["src/components/Comments.tsx:42"],
-          "confidence": "high"
+          "confidence": { "overall": 0.78, "dimensions": {} }
         }
       ]
     }
-  ]
+  ],
+  "confidence": {}
 }
 ```
 
+### Step 7.0: Debug Confidence Scoring
+
+Dimensions (4): hypothesis_quality, evidence_completeness, root_cause_isolation, fix_confidence. Factors (weights): evidence_depth(.30), evidence_strength(.25), coverage_breadth(.20), reproduction(.15), consistency(.10). Map to legacy levels: <40% = low, 40-70% = medium, >70% = high.
+
+Quality mechanisms: Pressure Pass (before Step 9) — cross-check confirmed vs refuted hypotheses. Devil's Advocate — root_cause_isolation > 0.7 → "根因在更深层？". Stall Detection — no new evidence + delta < 5% for 2 continuations → "调查可能停滞".
+
+Readiness Gate (blocks Step 9): evidence_completeness ≥ 40% | pressure pass done | no contradicting evidence | fix_direction has specific files. If blocked → AskUserQuestion: 补充调查 or 忽略风险并确认. Append confidence table to understanding.md.
+
 ### Step 7.1: Update Issues with Diagnosis
 
 For each diagnosed gap with an `issue_id`, update the corresponding issue in `.workflow/issues/issues.jsonl`:

package/workflows/plan.md

@@ -251,12 +251,22 @@ Bidirectional linking: update matching issues in `.workflow/issues/issues.jsonl`
    - Critical issues → re-spawn planner with issues, revise, re-check
    - Warnings only → log and proceed
 
-3. **Update index.json**
-   - Set `index.json.plan` = `{ task_ids, task_count, complexity, waves, executor_assignments: {} }`
+3. **Plan Confidence Scoring**
+
+   Dimensions (5): requirements_coverage, task_quality, dependency_correctness, estimation_accuracy, collision_safety. Factors (weights): completeness(.30), specificity(.25), structural_validity(.20), user_validation(.15), consistency(.10). Re-score after each revision round.
+
+   Quality mechanisms: Pressure Pass (mandatory before P4.5) — verify highest-complexity task's read_first/convergence.criteria/action. Devil's Advocate — requirements_coverage > 0.7 → "隐含需求？". Scope Minimizer — task_count exceeds guard → "最小可行任务集？". Stall Detection — delta < 5% → suggest broader revision.
+
+4. **Plan Readiness Gate** (blocks P4.5)
+
+   Block if: requirements_coverage < 40% | task missing read_first/convergence.criteria | no pressure pass | circular deps. If blocked → AskUserQuestion: 修订计划 or 忽略风险并继续 (record residual_risks). Add confidence section to plan.json.
+
+5. **Update index.json**
+   - Set `index.json.plan` = `{ task_ids, task_count, complexity, waves, executor_assignments: {}, confidence: overall_score }`
    - Set `status: "planning"`, `updated_at: now()`
 
 ### Output
-- Updated plan.json (if revised)
+- Updated plan.json (if revised) with confidence section
 - Updated .task/ files (if revised)
 - Updated index.json with plan fields
 
@@ -286,7 +296,7 @@ Bidirectional linking: update matching issues in `.workflow/issues/issues.jsonl`
 
 ### Steps
 
-1. **Display plan summary** — summary, approach, task count, wave structure, complexity, key dependencies
+1. **Display plan summary** — summary, approach, task count, wave structure, complexity, key dependencies, **plan confidence** (overall %, weakest dimension, pressure pass result)
 
 2. **Present options via AskUserQuestion** (skip if `config.gates.confirm_plan == false`, auto-proceed)
    - Execute now → build executionContext, hand off to /workflow:execute

package/workflows/test.md

@@ -416,6 +416,16 @@ If re-verify still has gaps: report remaining gaps, suggest manual intervention.
 
 ---
 
+### Step 12.5: UAT Confidence Scoring
+
+Dimensions (4): scenario_coverage, diagnostic_depth, observation_quality, closure_completeness. Factors (weights): requirements_mapped(.30), observation_specificity(.25), user_validation(.20), diagnostic_depth(.15), consistency(.10). Score at: init (Step 5), per user response (Step 8), after gap-fix loop (Step 12).
+
+Quality mechanisms: Pressure Pass — >80% pass → ask user to try edge case. Devil's Advocate — >70% first-try pass → challenge scenario difficulty. Stall Detection — 2 gap-fix iterations without improvement → stop.
+
+Readiness Gate (blocks Step 13): scenario_coverage < 40% | blocker gap without diagnosis | no pressure pass (if >80%) | unresolved gaps without acknowledgment. Append confidence summary to uat.md.
+
+---
+
 ### Step 13: Report
 
 ```