mindforge-cc 8.1.0 → 8.2.0

package/.mindforge/mirrors/mirror-sre-776a1cf9/.mindforge/intelligence/skill-gap-analyser.md

@@ -0,0 +1,40 @@
+# MindForge Intelligence — Skill Gap Analyser
+
+## Purpose
+Identify required/recommended skills for a phase, compare with installed
+skills, and report gaps before planning.
+
+## Run points
+1. `/mindforge:plan-phase` start
+2. `/mindforge:discuss-phase`
+3. `/mindforge:health`
+
+## Analysis flow
+1. Extract work signals from roadmap/context/requirements/architecture.
+2. Map work categories to required/recommended skills.
+3. Validate skills in manifest + filesystem + status.
+4. Report missing/deprecated required skills and mitigation options.
+
+## Category -> skill mapping
+- UI/UX -> accessibility (required), performance (recommended)
+- Database changes -> database-patterns (required), data-privacy (if PII)
+- API endpoints -> api-design (required), security-review (if auth)
+- Auth/payments/PII -> security-review (required), data-privacy (required for PII)
+- Performance work -> performance (required)
+- New feature -> testing-standards (required), documentation (recommended)
+
+## Gap definition
+A required skill is a gap if:
+- not present in MANIFEST
+- manifest path missing
+- marked deprecated with no replacement
+
+## Output
+Produce `Skill Gap Analysis` section with:
+- required skills table
+- recommended skills table
+- explicit gap count and options
+
+## Health mode
+In `/mindforge:health`, include org-level recommendation list using recent
+phase patterns and missing capability frequencies.

package/.mindforge/mirrors/mirror-sre-776a1cf9/.mindforge/intelligence/smart-compaction.md

@@ -0,0 +1,71 @@
+# MindForge Intelligence — Smart Context Compaction
+
+## Purpose
+Replace truncation with structured extraction so future sessions preserve reasoning,
+constraints, and in-progress state.
+
+## Triggering and level selection
+Day 2 threshold (`70%`) remains valid and now invokes Day 5 Level 1.
+Level selection uses both context usage and session depth.
+
+- Light session (`<5` tasks): use Level 1 even when context is high.
+- Moderate session (`5-10` tasks): use Level 2 at `80%+`.
+- Deep session (`10+` tasks): use Level 2 at `75%+`.
+- Level 3 at `90%+` or urgent near-limit state.
+
+## Levels
+### Level 1 — Lightweight
+Quick checkpoint: update `STATE.md` and `HANDOFF.json` with next task.
+
+### Level 2 — Structured extraction
+Capture:
+- decisions made
+- discoveries
+- implicit knowledge
+- quality signals
+- precise in-progress state
+
+### Level 3 — Emergency extraction
+Capture minimum viable continuation context before hard limit.
+
+## Level 2 extraction protocol
+### Block A — Decisions made
+For each decision: what, why, alternatives ruled out, impact area, ADR need.
+
+### Block B — Discoveries
+For each finding: fact, location, significance, assumption impact.
+
+### Block C — Current task state
+Record plan id, completed steps, remaining steps, current file, and inconsistency state.
+
+### Block D — Implicit knowledge
+Capture quirks, environment facts, and workarounds.
+
+A knowledge item qualifies as a **quirk** if either:
+1. It contradicts documentation or expected tool behavior.
+2. A new agent reading only plans/personas would not infer it.
+
+### Block E — Quality signals
+Capture failed assumptions, friction patterns, quality gate root causes, and next-plan suggestions.
+
+## Level 2 HANDOFF schema additions
+- `decisions_made[]`
+- `discoveries[]`
+- `implicit_knowledge[]`
+- `quality_signals[]`
+- `in_progress.current_state`
+
+## Session restart protocol
+When loading Level 2 handoff:
+1. Print compact summary counts.
+2. Preload `implicit_knowledge` before plan execution.
+3. If `in_progress.current_state` is inconsistent, block execution and ask to resolve first.
+4. If user asks to inspect first, print first 50 lines of the inconsistent file inline.
+
+## Compaction quality metric
+Write one entry to `.mindforge/metrics/compaction-quality.jsonl`:
+- `decisions_captured`
+- `discoveries_captured`
+- `implicit_knowledge_items`
+- `quality_signals_captured`
+- `next_session_continuation_success` (auto-inferred from next session, nullable initially)

package/.mindforge/mirrors/mirror-sre-776a1cf9/.mindforge/memory/MEMORY-SCHEMA.md

@@ -0,0 +1,155 @@
+# MindForge v2 — Knowledge Graph Schema
+
+## Overview
+
+The MindForge knowledge graph is built from five JSONL files — all append-only,
+all at `.mindforge/memory/`. Every entry has a consistent base schema plus
+type-specific fields.
+
+## Base schema (all entry types)
+
+```json
+{
+  "id":               "kb-uuid-v4",
+  "timestamp":        "ISO-8601",
+  "type":             "architectural_decision|code_pattern|bug_pattern|team_preference|domain_knowledge",
+  "topic":            "Short title (≤ 80 chars)",
+  "content":          "Full knowledge content (no limit)",
+  "source":           "Where this was captured: 'Phase 3 retro', 'Debug session 2026-01', 'Manual entry'",
+  "project":          "project-name from PROJECT.md",
+  "confidence":       0.0,
+  "tags":             ["tag1", "tag2"],
+  "linked_adrs":      ["ADR-007"],
+  "times_referenced": 0,
+  "last_referenced":  null,
+  "deprecated":       false,
+  "deprecated_by":    null
+}
+```
+
+## Type-specific schemas
+
+### `architectural_decision`
+
+Extended fields:
+
+```json
+{
+  "decision":     "Use argon2id for password hashing",
+  "rationale":    "bcrypt is showing its age; argon2id is the modern standard",
+  "alternatives": ["bcrypt", "scrypt", "pbkdf2"],
+  "phase_number": 3,
+  "adr_reference": "ADR-007"
+}
+```
+
+Capture trigger: Phase completion (ADR files extracted), `/mindforge:discuss-phase`
+
+### `code_pattern`
+
+Extended fields:
+
+```json
+{
+  "pattern_type": "good|anti-pattern",
+  "language":     "typescript",
+  "example_good": "// Code showing the correct pattern",
+  "example_bad":  "// Code showing the incorrect pattern (for anti-patterns)",
+  "applies_to":   ["auth", "database", "api"]
+}
+```
+
+Capture trigger: Smart compaction Block D, debug session root cause
+
+### `bug_pattern`
+
+Extended fields:
+
+```json
+{
+  "bug_category":   "auth|database|api|ui|performance|security",
+  "symptom":        "Login works in tests but fails in production with cookie errors",
+  "root_cause":     "httpOnly cookies require HTTPS; dev server was HTTP",
+  "fix":            "Enable HTTPS in dev or use secure: false in development only",
+  "prevention":     "Always configure cookie settings per-environment",
+  "severity_when_missed": "HIGH",
+  "recurrence_count": 1
+}
+```
+
+Capture trigger: Debug session completion, security findings, retrospective "what went wrong"
+
+### `team_preference`
+
+Extended fields:
+
+```json
+{
+  "preference_type": "style|tool|process|convention|architecture",
+  "applies_when":    "When implementing auth",
+  "preference":      "Always use argon2id, never bcrypt. See ADR-007.",
+  "strength":        "strong|moderate|weak",
+  "override_allowed": true
+}
+```
+
+Capture trigger: Retrospective "what should we keep doing", steering instructions,
+MINDFORGE.md changes from retrospective
+
+### `domain_knowledge`
+
+Extended fields:
+
+```json
+{
+  "domain":     "JWT security",
+  "tech_stack": ["node.js", "typescript"],
+  "knowledge":  "JWT access tokens should be short-lived (15 min). Refresh tokens in httpOnly cookies, not localStorage. Never decode JWT without verifying signature.",
+  "source_url": "https://auth0.com/docs/secure/tokens/json-web-tokens",
+  "verified_at": "2026-01-15"
+}
+```
+
+Capture trigger: Research engine output, `/mindforge:remember --add`
+
+## Confidence scoring
+
+Confidence (0.0–1.0) represents how strongly this knowledge should be applied:
+
+| Confidence | Meaning | How set |
+| :--- | :--- | :--- |
+| 0.9–1.0 | Strong team decision — always apply | Manual entry, strong steering, ADR |
+| 0.7–0.9 | Clear pattern — apply unless there's a reason not to | Retrospective, debug session |
+| 0.5–0.7 | Observed pattern — consider but verify | Smart compaction Block D |
+| 0.3–0.5 | Weak signal — informational only | Single observation |
+| < 0.3 | Hypothesis — don't apply proactively | Auto-captured, unverified |
+
+## Confidence reinforcement
+
+Each time an entry is referenced (retrieved and the advice was followed):
+
+```bash
+confidence = min(1.0, confidence + 0.05)
+times_referenced += 1
+```
+
+Each time an entry is contradicted (a different decision was made):
+
+```bash
+confidence = max(0.0, confidence - 0.10)
+```
+
+## Deprecation
+
+When a knowledge entry becomes outdated (e.g., team switches from argon2id to Passkeys):
+
+```json
+{
+  "deprecated": true,
+  "deprecated_by": "kb-uuid-of-newer-entry",
+  "deprecated_reason": "Superseded by WebAuthn/Passkeys approach (ADR-031)"
+}
+```
+
+Deprecated entries are never deleted — they are excluded from active retrieval
+but available for historical queries.

package/.mindforge/mirrors/mirror-sre-776a1cf9/.mindforge/memory/engine/capture-protocol.md

@@ -0,0 +1,36 @@
+# MindForge v2 — Knowledge Capture Protocol
+
+## Purpose
+
+The Capture Engine automatically extracts valuable engineering knowledge from session artifacts to ensure that project-specific wisdom is preserved beyond a single handoff.
+
+## Capture Triggers
+
+MindForge monitors the following lifecycle events for knowledge extraction:
+
+1. **Phase Completion**: Extracting architectural decisions from `PROJECT.md` and phase retrospectives.
+2. **Hardened Debugging**: Capturing "root cause -> fix" patterns from completed debug reports.
+3. **Manual Command**: Using `/mindforge:remember --add` to explicitly record team preferences or domain knowledge.
+4. **ADR Generation**: Parsing newly created Architecture Decision Records.
+
+## Extraction Logic
+
+The extraction process (via `knowledge-capture.js`) involves:
+
+- **Structural Parsing**: Identifying standardized headings (e.g., "Decision", "Root Cause", "Pattern").
+- **Deduplication**: Checking for existing entries with similar content to prevent knowledge bloat.
+- **Initial Confidence**: Assigning a default confidence score (usually 0.5) to newly captured entries.
+- **Auto-Tagging**: Inferring tags based on the file path and content context (e.g., `#nodejs`, `#security`).
+
+## Inhibitors (What NOT to capture)
+
+To prevent the knowledge base from becoming "noisy", the system ignores:
+
+- Transient session state (e.g., local file counts).
+- Intermediate debugging steps that didn't lead to a fix.
+- Content marked with `#no-capture` or inside blocks labeled "Confidential".
+
+## Retention Policy
+
+- **Append-only**: Entries are never physically deleted.
+- **Deprecation**: Outdated entries are marked as `deprecated: true` and replaced by newer entries with the same `source_id`.

package/.mindforge/mirrors/mirror-sre-776a1cf9/.mindforge/memory/engine/global-sync-spec.md

@@ -0,0 +1,42 @@
+# MindForge v2 — Global Knowledge Sync Specification
+
+## Purpose
+
+Allow knowledge to flow between projects via a machine-level global store.
+Team members can promote project-specific insights to their global knowledge base,
+which loads into every future project's session start.
+
+## Storage locations
+
+- **Project-local:** `.mindforge/memory/knowledge-base.jsonl` (per repo)
+- **Global (machine):** `~/.mindforge/global-knowledge-base.jsonl` (per developer)
+
+## Sync rules
+
+- Promotion is MANUAL — nothing is auto-promoted to global. Use `/mindforge:remember --promote [ID]`.
+- Loading is AUTOMATIC — global entries load at every session start if relevant.
+- Project entries take precedence over global entries (same ID = local wins).
+- Global entries receive a metadata flag `global: true` but are ranked alongside project entries based on relevance.
+
+## What should be promoted to global
+
+Good candidates for global promotion:
+
+- Language-agnostic security practices ("always validate input before database queries")
+- Technology-specific best practices learned through experience ("argon2id over bcrypt")
+- Universal debugging patterns ("check the timezone mismatch before blaming async code")
+- Cross-project architectural preferences ("Repository pattern over active record")
+
+## Global entry metadata
+
+When promoted to global, entries gain:
+
+```json
+{
+  "global": true,
+  "promoted_at": "ISO-8601",
+  "promoted_from_project": "saas-app",
+  "promoted_by": "git-config-user-email",
+  "global_applicability": "all|typescript|nodejs|react|[specific]"
+}
+```

package/.mindforge/mirrors/mirror-sre-776a1cf9/.mindforge/memory/engine/retrieval-spec.md

@@ -0,0 +1,44 @@
+# MindForge v2 — Knowledge Retrieval Specification
+
+## Purpose
+
+The Retrieval Engine is responsible for identifying and loading the most relevant pieces of long-term memory into the current session context to improve agent performance and consistency.
+
+## Components
+
+- **Indexer (`knowledge-indexer.js`)**: Handles the scoring and ranking of knowledge entries.
+- **Session Loader (`session-memory-loader.js`)**: Orchestrates the loading of relevant memories at the start of a session.
+
+## Scoring Logic (TF-IDF + Heuristics)
+
+Relevance is calculated using a multi-factor scoring model:
+
+1. **Text Search (TF-IDF)**:
+    - Tokenizes the query and knowledge text.
+    - Calculates Term Frequency (TF) and Inverse Document Frequency (IDF).
+    - Entries must have a positive TF-IDF score to be considered relevant if keywords are present.
+
+2. **Confidence & Recency**:
+    - **Confidence Score**: Boosts entries with higher confidence levels (reinforced over time).
+    - **Recency Decay**: Prefer more recent knowledge entries, especially for bug patterns.
+
+3. **Tag Overlap**:
+    - Queries can include required tags (e.g., `#typescript`).
+    - Entries with matching tags receive a significant boost.
+
+4. **Inhibitors**:
+    - Deprecated entries are never retrieved.
+    - Low-confidence "noise" entries are filtered out.
+
+## Context Injection
+
+During session start, the loader:
+1. Performs a broad search based on the current project's primary technology stack and recent activities.
+2. Selects the top `N` (default 5-10) most relevant entries.
+3. Injects them into the agent's secondary context (e.g., as a dedicated "Memory context" block).
+
+## SDK Interface
+
+```typescript
+const relevantKnowledge = await memory.query("query string", ["#tag1"], 5);
+```

package/.mindforge/mirrors/mirror-sre-776a1cf9/.mindforge/metrics/METRICS-SCHEMA.md

@@ -0,0 +1,42 @@
+# MindForge Metrics — Schema Reference
+
+## Files
+- `session-quality.jsonl`
+- `phase-metrics.jsonl`
+- `skill-usage.jsonl`
+- `compaction-quality.jsonl`
+
+All files are append-only JSONL.
+
+## session-quality.jsonl fields
+- session_id, date, phase, developer_id
+- tasks_attempted, tasks_completed, tasks_failed
+- verify_pass_rate, quality_gates_failed
+- security_findings, context_compactions
+- skills_loaded_count, antipatterns_detected
+- session_quality_score, session_quality_score_raw
+
+Score formula:
+- base 100
+- minus task/gate/security/antipattern penalties
+- bonus for zero gate failures and zero security findings
+- keep both raw and clamped (`0-100`) score for trend analysis
+
+## phase-metrics.jsonl fields
+- phase lifecycle and delivery coverage
+- difficulty score
+- security and UAT outcomes
+- phase_quality_score
+
+## skill-usage.jsonl fields
+- date, phase, plan
+- skill_name, skill_version
+- trigger_type (`text_match|file_path_match|file_name_match`)
+- trigger_keyword
+- task_outcome, verify_passed_first_try
+
+## compaction-quality.jsonl fields
+- date, compaction_level, context_pct_at_compaction
+- decisions_captured, discoveries_captured
+- implicit_knowledge_items, quality_signals_captured
+- next_session_continuation_success (auto-inferred where possible)

package/.mindforge/mirrors/mirror-sre-776a1cf9/.mindforge/metrics/quality-tracker.md

@@ -0,0 +1,32 @@
+# MindForge Metrics — Quality Tracker
+
+## Trend windows
+- short: last 5 sessions
+- medium: last 20 sessions
+- long: all sessions
+
+## Key targets
+- Verify pass rate > 85% (warning < 75%)
+- Task failure rate < 10% (warning > 20%)
+- Compactions < 1 per 8 tasks (warning > 2 per session)
+
+## Early warnings
+- quality score drop >= 10 points over 3 sessions
+- any CRITICAL security finding after clean streak
+- compaction overhead > 2/session
+- stale skills not triggered in 10 sessions
+
+## Quality metrics -> agent behaviour adjustment
+Automatic adjustments at session start:
+
+- If verify pass rate < 75% over last 3 sessions:
+  tighten verify-step determinism checks before execution.
+- If last-session task failure rate > 20%:
+  halve estimated scope and suggest splitting.
+- If compaction frequency > 2/session:
+  proactively summarize every 4 tasks.
+- If security findings trend worsens over 3 phases:
+  force-load `security-review` for all tasks in current phase.
+
+Always report adjustment to user:
+`Quality signal detected: [signal]. Adjusting behaviour: [adjustment].`

package/.mindforge/mirrors/mirror-sre-776a1cf9/.mindforge/models/model-registry.md

@@ -0,0 +1,48 @@
+# MindForge v2 — Model Registry
+# Version: v2.0.0-alpha.3
+
+MindForge v2 uses a tiered multi-model architecture. Models are selected based on the task persona and difficulty tier.
+
+## 1. Primary Reasoning Tier (Tier 1)
+Used for complex architecture, planning, and high-stakes reasoning.
+
+- **Anthropic Claude 3.5 Opus** (`claude-3-opus-20240229`)
+  - Strengths: Deepest reasoning, highly reliable for planning.
+  - Persona mapping: `architect`, `planner`.
+  - Setting: `PLANNER_MODEL`.
+
+- **OpenAI GPT-4o** (`gpt-4o`)
+  - Strengths: Fast reasoning, excellent tool use, adversarial insight.
+  - Persona mapping: `reviewer`, `security-reviewer`.
+  - Setting: `REVIEWER_MODEL`.
+
+## 2. Execution Tier (Tier 2)
+Used for the bulk of code implementation and transformation tasks.
+
+- **Anthropic Claude 3.5 Sonnet** (`claude-3-5-sonnet-20240620`)
+  - Strengths: Balanced speed/quality, best-in-class coding.
+  - Persona mapping: `developer`, `qa-engineer`.
+  - Setting: `EXECUTOR_MODEL`.
+
+- **Google Gemini 1.5 Pro** (`gemini-1.5-pro`)
+  - Strengths: 1M context window, deep codebase research.
+  - Persona mapping: `research-agent`.
+  - Setting: `RESEARCH_MODEL`.
+
+## 3. High-Security/Privacy Tier (Tier 3)
+Used for tasks involving authentication, PII, or financial operations.
+
+- **Anthropic Claude 3.5 Opus** (`claude-3-opus-20240229`)
+  - Setting: `SECURITY_MODEL`.
+
+## 4. Utility/Quick Tier (Tier 4)
+Used for low-ambiguity tasks and micro-transformations.
+
+- **Anthropic Claude 3.5 Haiku** (`claude-3-5-haiku-20241022`)
+  - Setting: `QUICK_MODEL`.
+
+## 5. Fallback Chain
+If a primary model is unavailable (API key missing or 429), follow these defaults:
+1. `gpt-4o` → `claude-3-5-sonnet`
+2. `claude-3-opus` → `gpt-4o`
+3. `gemini-1.5-pro` → `claude-3-5-sonnet` (limited context)

package/.mindforge/mirrors/mirror-sre-776a1cf9/.mindforge/models/model-router.md

@@ -0,0 +1,30 @@
+# MindForge v2 — Model Router
+# Version: v2.0.0-alpha.3
+
+## Routing Logic
+
+The MindForge router resolves which model to use at task dispatch time using this algorithm:
+
+1. **Check Persona:** Map the task `<persona>` to a setting key.
+   - `developer` → `EXECUTOR_MODEL`
+   - `architect` → `PLANNER_MODEL`
+   - `security-reviewer` → `SECURITY_MODEL`
+   - `qa-engineer` → `QA_MODEL`
+   - `research-agent` → `RESEARCH_MODEL`
+   - `debug-specialist` → `DEBUG_MODEL`
+   - `unknown` → `EXECUTOR_MODEL`
+
+2. **Check Tier Override:**
+   - If `tier == 3` (Security/Privacy), force `SECURITY_MODEL` regardless of persona.
+
+3. **Check Budget Bias:**
+   - If `difficulty < 2.0` and `MODEL_PREFER_CHEAP_BELOW_DIFFICULTY` is set:
+     - Use `QUICK_MODEL` instead of the resolved model.
+
+4. **Resolve Model ID:**
+   - Read the setting key from `MINDFORGE.md`.
+   - Use baked-in defaults if not present.
+
+5. **Verify Availability:**
+   - Check if matching `[PROVIDER]_API_KEY` is present in environment.
+   - If missing: Apply fallback chain from `model-registry.md`.

package/.mindforge/mirrors/mirror-sre-776a1cf9/.mindforge/monorepo/cross-package-planner.md

@@ -0,0 +1,114 @@
+# MindForge — Cross-Package Planner
+
+## Purpose
+When a phase spans multiple packages in a monorepo, coordinate execution
+so that shared dependencies are built and tested before packages that
+depend on them.
+
+## Execution order algorithm
+
+```
+Given: a set of packages involved in the current phase
+       and their inter-package dependencies
+
+Algorithm: topological sort of package dependency graph
+
+Input: WORKSPACE-MANIFEST.json dependency_order
+Output: ordered list of packages to process
+
+Example:
+  Phase touches: @myapp/api, @myapp/shared, @myapp/web
+  Dependencies: api→shared, web→shared, web→api
+  Topological order: shared → api → web
+  
+  Execution: 
+    Wave 1: process @myapp/shared (no dependencies on other changed packages)
+    Wave 2: process @myapp/api (depends on Wave 1: shared)
+    Wave 3: process @myapp/web (depends on Wave 2: api + Wave 1: shared)
+```
+
+## Per-package PLAN file routing
+
+When PLAN files are created for a monorepo phase, each plan specifies its target package:
+
+```xml
+<task type="auto">
+  <n>Add auth middleware to API</n>
+  <package>@myapp/api</package>
+  <working-dir>apps/api</working-dir>
+  <files>
+    apps/api/src/middleware/auth.ts
+    apps/api/src/middleware/auth.test.ts
+  </files>
+  <action>...</action>
+  <verify>cd apps/api && npm test -- --testPathPattern=auth.middleware</verify>
+  <done>Auth middleware tests passing in apps/api</done>
+</task>
+```
+
+## Cross-package test execution
+
+After all packages in the phase are processed:
+```bash
+# Run tests across all affected packages
+AFFECTED_PACKAGES=$(cat .planning/WORKSPACE-MANIFEST.json | \
+  node -e "const m=JSON.parse(require('fs').readFileSync('/dev/stdin','utf8')); \
+  console.log(m.packages.map(p=>p.path).join(' '))")
+
+for PKG_PATH in ${AFFECTED_PACKAGES}; do
+  echo "Testing ${PKG_PATH}..."
+  cd "${PKG_PATH}" && npm test && cd -
+done
+
+# Run integration tests from root (if configured)
+[ -f "package.json" ] && npm run test:integration 2>/dev/null || true
+```
+
+## Affected package detection (revised)
+
+```bash
+# CORRECTED: Match against declared package paths, not assume 2-level depth
+detect_affected_packages() {
+  local MANIFEST=".planning/WORKSPACE-MANIFEST.json"
+
+  if [ ! -f "${MANIFEST}" ]; then
+    echo "Run /mindforge:workspace detect first"
+    return 1
+  fi
+
+  # Get list of changed files
+  CHANGED_FILES=$(git diff --name-only HEAD~1 2>/dev/null || git diff --cached --name-only)
+
+  # For each package in manifest, check if any changed file is within that package
+  node -e "
+    const fs = require('fs');
+    const manifest = JSON.parse(fs.readFileSync('${MANIFEST}', 'utf8'));
+    const changedFiles = \`${CHANGED_FILES}\`.split('\n').filter(Boolean);
+
+    const affected = new Set();
+    manifest.packages.forEach(pkg => {
+      // Check if any changed file is within this package's path
+      const pkgPath = pkg.path.replace(/\/$/, ''); // remove trailing slash
+      changedFiles.forEach(file => {
+        if (file.startsWith(pkgPath + '/') || file === pkgPath) {
+          affected.add(pkg.name);
+        }
+      });
+    });
+
+    // Also add packages that depend on affected packages
+    manifest.packages.forEach(pkg => {
+      if (pkg.dependencies && pkg.dependencies.some(dep => affected.has(dep))) {
+        affected.add(pkg.name);
+      }
+    });
+
+    console.log([...affected].join('\n'));
+  "
+}
+```
+
+This correctly handles:
+- Packages at any nesting depth (`libs/shared/utils/` → 3 levels deep)
+- Packages whose path is a prefix of another (avoid false matches)
+- Transitive dependencies (packages that depend on changed packages)