compound-agent 1.2.3 → 1.2.5

package/CHANGELOG.md

@@ -9,6 +9,20 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [1.2.4] - 2026-02-15
+
+### Changed
+
+- **lfg.md delegates phases to slash commands**: Instead of inlining all 5 phase workflows (~80 lines), lfg.md now invokes `/compound:{brainstorm,plan,work,review,compound}`. This prevents phase instructions from being compacted away by late phases.
+- **lfg.md slimmed to thin orchestrator**: Reduced from ~2200 to ~1324 characters. Removed inlined Purpose, Stop Conditions, and Memory Integration sections.
+- **Phase gates relocated to individual commands**: PHASE GATE 3 moved to work.md, PHASE GATE 4 to review.md, FINAL GATE to compound.md — gates now survive compaction.
+- **YAML frontmatter on all 13 command templates**: Each template now has name, description, and argument-hint metadata. lfg.md additionally uses `disable-model-invocation: true`.
+- **Anti-MEMORY.md guardrails**: compound.md and review.md now explicitly warn against using MEMORY.md for lesson storage, directing Claude to use `memory_capture` MCP tool instead.
+
+### Fixed
+
+- **Phase 5 context drift**: Claude no longer falls back to MEMORY.md during compound phase because each phase loads fresh instructions from its dedicated slash command.
+
 ## [1.2.1] - 2026-02-15
 
 ### Added
@@ -456,7 +470,8 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   - Vitest test suite
   - tsup build configuration
 
-[Unreleased]: https://github.com/Nathandela/learning_agent/compare/v1.2.1...HEAD
+[Unreleased]: https://github.com/Nathandela/learning_agent/compare/v1.2.4...HEAD
+[1.2.4]: https://github.com/Nathandela/learning_agent/compare/v1.2.1...v1.2.4
 [1.2.1]: https://github.com/Nathandela/learning_agent/compare/v1.2.0...v1.2.1
 [1.2.0]: https://github.com/Nathandela/learning_agent/compare/v1.1.0...v1.2.0
 [1.1.0]: https://github.com/Nathandela/learning_agent/compare/v1.0.0...v1.1.0

package/dist/cli.js

@@ -2881,7 +2881,12 @@ var AGENT_TEMPLATES = {
 
 // src/setup/templates/commands.ts
 var WORKFLOW_COMMANDS = {
-  "brainstorm.md": `$ARGUMENTS
+  "brainstorm.md": `---
+name: compound:brainstorm
+description: Explore requirements through collaborative dialogue before committing to a plan
+argument-hint: "<goal or topic to brainstorm>"
+---
+$ARGUMENTS
 
 # Brainstorm
 
@@ -2939,7 +2944,12 @@ Explore requirements through collaborative dialogue before committing to a plan.
 - Create a beads epic from brainstorm conclusions with \`bd create --type=feature\`.
 - If the brainstorm identifies sub-tasks, suggest creating them with \`bd create\`.
 `,
-  "plan.md": `$ARGUMENTS
+  "plan.md": `---
+name: compound:plan
+description: Create a structured implementation plan with concrete tasks and dependencies
+argument-hint: "<goal or epic to plan>"
+---
+$ARGUMENTS
 
 # Plan
 
@@ -2997,7 +3007,12 @@ If either is missing, CREATE THEM NOW. The plan is NOT complete without these ga
 - Map dependencies with bd dep add (dependent) (blocker).
 - Each task should include acceptance criteria in its description.
 `,
-  "work.md": `$ARGUMENTS
+  "work.md": `---
+name: compound:work
+description: Execute implementation by delegating to an agent team
+argument-hint: "<task ID or description>"
+---
+$ARGUMENTS
 
 # Work
 
@@ -3054,8 +3069,19 @@ for complex changes. For all changes, /implementation-reviewer is the minimum re
 - Start with \`bd ready\` to pick work.
 - Update status with bd update (id) --status=in_progress.
 - Close with bd close (id) when all tests pass.
+
+## PHASE GATE 3 -- MANDATORY
+Before starting Review, verify ALL work tasks are closed:
+- Run bd list with status in_progress -- must return empty
+- Run bd list with status open -- only Review and Compound tasks should remain
+If any work tasks remain open, DO NOT proceed. Complete them first.
 `,
-  "review.md": `$ARGUMENTS
+  "review.md": `---
+name: compound:review
+description: Multi-agent code review with severity classification and mandatory gate
+argument-hint: "<scope or git diff range>"
+---
+$ARGUMENTS
 
 # Review
 
@@ -3098,6 +3124,7 @@ Multi-agent code review with severity classification and a mandatory \`/implemen
 - **pattern-matcher** auto-reinforces recurring findings via \`memory_capture\`.
 - **cct-reviewer** reads \`.claude/lessons/cct-patterns.jsonl\` for known Claude mistakes.
 - After review, call \`memory_capture\` with \`type=solution\` to store the review report.
+- **CRITICAL**: Use \`memory_capture\` MCP tool for ALL lesson storage -- NOT MEMORY.md.
 
 ## Docs Integration
 - **docs-reviewer** checks code changes align with \`docs/\` and existing ADRs.
@@ -3106,14 +3133,27 @@ Multi-agent code review with severity classification and a mandatory \`/implemen
 ## Beads Integration
 - Create \`bd\` issues for P1 and P2 findings with \`bd create\`.
 - Close related issues with \`bd close\` when findings are resolved.
+
+## PHASE GATE 4 -- MANDATORY
+Before starting Compound, verify review is complete:
+- /implementation-reviewer must have returned APPROVED
+- All P1 findings must be resolved
 `,
-  "compound.md": `$ARGUMENTS
+  "compound.md": `---
+name: compound:compound
+description: Capture high-quality lessons from completed work into the memory system
+argument-hint: "<topic or epic context>"
+---
+$ARGUMENTS
 
 # Compound
 
 ## Purpose
 Multi-agent analysis to capture high-quality lessons from completed work into the memory system and update project documentation.
 
+**CRITICAL**: Store all lessons via \`memory_capture\` MCP tool -- NOT via MEMORY.md, NOT via markdown files.
+Lessons go to \`.claude/lessons/index.jsonl\` through the MCP tool. MEMORY.md is a different system and MUST NOT be used for compounding.
+
 ## Workflow
 1. Parse what was done from the arguments above or recent git history (\`git diff\`, \`git log\`).
 2. Call \`memory_search\` with the topic to check what is already known (avoid duplicates).
@@ -3162,118 +3202,53 @@ Multi-agent analysis to capture high-quality lessons from completed work into th
 ## Beads Integration
 - Check \`bd ready\` for related open issues.
 - Close resolved issues with \`bd close\`.
-`,
-  "lfg.md": `$ARGUMENTS
-
-# LFG (Full Cycle)
-
-## Purpose
-Chain all phases: brainstorm, plan, work, review, compound. End-to-end delivery.
-
-## Workflow
-1. **Brainstorm phase**: Explore the goal stated above.
-   - MEMORY CHECK: Call \`memory_search\` with the current goal/task. Display results to user. If relevant items found, state which ones apply and why. If none found, state "No relevant lessons found."
-   - Call \`memory_search\` with the goal.
-   - \`TeamCreate\` team "brainstorm-<slug>", spawn docs-explorer + code-explorer as parallel teammates.
-   - Ask clarifying questions via \`AskUserQuestion\`, explore alternatives.
-   - Auto-create ADRs for significant decisions in \`docs/decisions/\`.
-   - Create a beads epic from conclusions with \`bd create --type=feature\`.
-   - Shut down brainstorm team before next phase.
-   - Update epic phase state: bd update (epic-id) --notes="Phase: brainstorm COMPLETE, Next: plan"
-
-2. **Plan phase**: Structure the work.
-   - MEMORY CHECK: Call \`memory_search\` with the current goal/task. Display results to user. If relevant items found, state which ones apply and why. If none found, state "No relevant lessons found."
-   - Check for brainstorm epic via \`bd list\`.
-   - \`TeamCreate\` team "plan-<slug>", spawn docs-analyst + repo-analyst + memory-analyst as parallel teammates.
-   - Break into tasks with dependencies and acceptance criteria.
-   - Create beads issues with \`bd create\` and map dependencies with \`bd dep add\`.
-   - Create review and compound blocking tasks (\`bd create\` + \`bd dep add\`) so they survive compaction and surface via \`bd ready\` after work completes.
-   - Shut down plan team before next phase.
-   - Update epic phase state: bd update (epic-id) --notes="Phase: plan COMPLETE, Next: work"
-
-3. **Work phase**: Implement with adaptive TDD.
-   - MEMORY CHECK: Call \`memory_search\` with the current goal/task. Display results to user. If relevant items found, state which ones apply and why. If none found, state "No relevant lessons found."
-   - Assess complexity (trivial/simple/complex) to choose strategy.
-   - Trivial: single subagent, no team. Simple/complex: \`TeamCreate\` team "work-<task-id>".
-   - Spawn test-analyst first, then test-writer + implementer as teammates.
-   - Call \`memory_search\` per subtask; \`memory_capture\` after corrections.
-   - Commit incrementally. Close tasks as they complete.
-   - Run verification gate before marking complete. Shut down work team.
-   - Update epic phase state: bd update (epic-id) --notes="Phase: work COMPLETE, Next: review"
-
-## PHASE GATE 3->4 -- MANDATORY
-Before starting Review, verify ALL work tasks are closed:
-- Run \`bd list --status=in_progress\` \u2014 must return empty
-- Run \`bd list --status=open\` \u2014 only Review and Compound tasks should remain
-If any work tasks remain open, DO NOT proceed. Complete them first.
-Update epic phase: bd update (epic-id) --notes="Phase: work COMPLETE, Next: review"
-
-4. **Review phase**: 11-agent review with severity classification.
-   - MEMORY CHECK: Call \`memory_search\` with the current goal/task. Display results to user. If relevant items found, state which ones apply and why. If none found, state "No relevant lessons found."
-   - Run quality gates first: pnpm test, then pnpm lint.
-   - \`TeamCreate\` team "review-<slug>", spawn all 11 reviewers as parallel teammates.
-   - Classify findings as P1 (critical/blocking), P2 (important), P3 (minor).
-   - P1 findings must be fixed before proceeding \u2014 they block completion.
-   - Submit to \`/implementation-reviewer\` as the mandatory gate. Shut down review team.
-   - Update epic phase state: bd update (epic-id) --notes="Phase: review COMPLETE, Next: compound"
-
-## PHASE GATE 4->5 -- MANDATORY
-Before starting Compound, verify review is complete:
-- /implementation-reviewer must have returned APPROVED
-- All P1 findings must be resolved
-Update epic phase: bd update (epic-id) --notes="Phase: review COMPLETE, Next: compound"
-
-5. **Compound phase**: Capture learnings.
-   - MEMORY CHECK: Call \`memory_search\` with the current goal/task. Display results to user. If relevant items found, state which ones apply and why. If none found, state "No relevant lessons found."
-   - \`TeamCreate\` team "compound-<slug>", spawn 6 analysis agents as parallel teammates.
-   - Search first with \`memory_search\` to avoid duplicates. Apply quality filters (novelty + specificity).
-   - Store novel insights via \`memory_capture\` with supersedes/related links.
-   - Update outdated docs and deprecate superseded ADRs.
-   - Use \`AskUserQuestion\` to confirm high-severity items. Shut down compound team.
-   - Update epic phase state: bd update (epic-id) --notes="Phase: compound COMPLETE, Next: close"
 
 ## FINAL GATE -- EPIC CLOSURE
 Before closing the epic:
-- Run ca verify-gates (epic-id) \u2014 must return PASS for both gates
-- Run \`pnpm test\` \u2014 must pass
-- Run \`pnpm lint\` \u2014 must pass
+- Run ca verify-gates (epic-id) -- must return PASS for both gates
+- Run pnpm test -- must pass
+- Run pnpm lint -- must pass
 If verify-gates fails, the missing phase was SKIPPED. Go back and complete it.
 CRITICAL: 3/5 phases is NOT success. All 5 phases are required.
+`,
+  "lfg.md": `---
+name: compound:lfg
+description: Full workflow cycle chaining brainstorm, plan, work, review, and compound phases
+argument-hint: "<goal>"
+disable-model-invocation: true
+---
+$ARGUMENTS
 
-## Agent Team Pattern
-Each phase creates its own AgentTeam via \`TeamCreate\`, spawns teammates via \`Task\` tool with \`team_name\`, coordinates via \`SendMessage\`, and shuts down with \`shutdown_request\` before the next phase starts. Use subagents (Task without team_name) only for quick lookups like \`memory_search\` or \`bd\` commands.
-
-## Phase Control
-- **Skip phases**: If the arguments above contain "from <phase>" (e.g., "from plan"), skip all phases before the named one.
-- **Progress**: Announce the current phase before starting it (e.g., "[Phase 2/5] Plan").
-- **Retry**: If a phase fails, report the failure and ask the user whether to retry, skip, or abort.
-- **Resume**: After interruption, run bd show (epic-id) and read the notes field for current phase state. Resume from that phase. If no phase state, check \`bd list --status=in_progress\` to infer.
+# LFG (Full Cycle)
 
-## Stop Conditions
-- Stop if brainstorm reveals the goal is unclear (ask user).
-- Stop if any test phase produces failures that cannot be resolved.
-- Stop if review finds critical security issues.
+## Workflow
+1. **Brainstorm**: /compound:brainstorm with the goal. Update: bd update (epic-id) --notes="Phase: brainstorm COMPLETE, Next: plan"
+2. **Plan**: /compound:plan with conclusions. Update: bd update (epic-id) --notes="Phase: plan COMPLETE, Next: work"
+3. **Work**: /compound:work (finds tasks via bd ready). Update: bd update (epic-id) --notes="Phase: work COMPLETE, Next: review"
+4. **Review**: /compound:review on changed code. Update: bd update (epic-id) --notes="Phase: review COMPLETE, Next: compound"
+5. **Compound**: /compound:compound to capture learnings (via memory_capture, NOT MEMORY.md). Update: bd update (epic-id) --notes="Phase: compound COMPLETE, Next: close"
 
-## Memory Integration
-- \`memory_search\` is called in brainstorm, work, and compound phases.
-- \`memory_capture\` is called in work and compound phases.
+## Phase Control
+- Skip: "from <phase>" in arguments skips earlier phases.
+- Resume: bd show (epic-id), read notes field for phase state.
+- Progress: announce "[Phase N/5] Name" before each phase.
 
 ## SESSION CLOSE -- INVIOLABLE
-Before saying "done" or "complete", ALL of these must pass:
-1. ca verify-gates (epic-id) -- All workflow gates satisfied
-2. pnpm test -- All tests green
-3. pnpm lint -- Zero violations
-4. git status -- Review changes
-5. git add (specific files) -- Stage (never git add .)
-6. bd sync -- Sync beads
-7. git commit -- Commit with descriptive message
-8. git push -- Push to remote
-If ANY step fails, fix it. Work is NOT done until git push succeeds.
+1. ca verify-gates (epic-id)
+2. pnpm test -- all green
+3. pnpm lint -- zero violations
+4. git status, git add (specific files), bd sync, git commit, git push
+Work is NOT done until git push succeeds.
 `,
   // =========================================================================
   // Utility commands (CLI wrappers)
   // =========================================================================
-  "learn.md": `Capture a lesson from this session.
+  "learn.md": `---
+name: compound:learn
+description: Capture a lesson from this session into the memory system
+argument-hint: "<insight to remember>"
+---
+Capture a lesson from this session.
 
 Usage: /compound learn <insight>
 
@@ -3285,7 +3260,12 @@ Examples:
 npx ca learn "$ARGUMENTS"
 \`\`\`
 `,
-  "search.md": `Search lessons for relevant context.
+  "search.md": `---
+name: compound:search
+description: Search stored lessons for relevant context
+argument-hint: "<search query>"
+---
+Search lessons for relevant context.
 
 Usage: /compound search <query>
 
@@ -3299,19 +3279,32 @@ npx ca search "$ARGUMENTS"
 
 Note: You can also use the \`memory_search\` MCP tool directly.
 `,
-  "list.md": `Show all stored lessons.
+  "list.md": `---
+name: compound:list
+description: Show all stored lessons
+---
+Show all stored lessons.
 
 \`\`\`bash
 npx ca list
 \`\`\`
 `,
-  "prime.md": `Load compound-agent workflow context after compaction or context loss.
+  "prime.md": `---
+name: compound:prime
+description: Load compound-agent workflow context after compaction or context loss
+---
+Load compound-agent workflow context after compaction or context loss.
 
 \`\`\`bash
 npx ca prime
 \`\`\`
 `,
-  "show.md": `Show details of a specific lesson.
+  "show.md": `---
+name: compound:show
+description: Show details of a specific lesson
+argument-hint: "<lesson-id>"
+---
+Show details of a specific lesson.
 
 Usage: /compound show <lesson-id>
 
@@ -3319,7 +3312,12 @@ Usage: /compound show <lesson-id>
 npx ca show "$ARGUMENTS"
 \`\`\`
 `,
-  "wrong.md": `Mark a lesson as incorrect or invalid.
+  "wrong.md": `---
+name: compound:wrong
+description: Mark a lesson as incorrect or invalid
+argument-hint: "<lesson-id>"
+---
+Mark a lesson as incorrect or invalid.
 
 Usage: /compound wrong <lesson-id>
 
@@ -3327,7 +3325,11 @@ Usage: /compound wrong <lesson-id>
 npx ca wrong "$ARGUMENTS"
 \`\`\`
 `,
-  "stats.md": `Show compound-agent database statistics and health.
+  "stats.md": `---
+name: compound:stats
+description: Show compound-agent database statistics and health
+---
+Show compound-agent database statistics and health.
 
 \`\`\`bash
 npx ca stats
@@ -3436,6 +3438,11 @@ Create a concrete implementation plan by decomposing work into small, testable t
 - Existing docs and ADRs were checked for constraints
 - Ambiguities resolved via \`AskUserQuestion\` before decomposing
 - Complexity estimates are realistic (no "should be quick")
+
+## POST-PLAN VERIFICATION -- MANDATORY
+After creating all tasks, verify review and compound tasks exist:
+- Run \`bd list --status=open\` and check for a "Review:" task and a "Compound:" task
+- If either is missing, CREATE THEM NOW. The plan is NOT complete without these gates.
 `,
   work: `---
 name: Work
@@ -3473,9 +3480,14 @@ Spawn teammates via \`Task\` with \`team_name\`, coordinate via \`SendMessage\`,
 - Each agent receives memory items tailored to their assigned task, not a shared blob
 - Call \`memory_capture\` after corrections or novel discoveries
 
-## Verification Gate
-Before marking work complete, run the 8-step TDD verification pipeline:
-1. /invariant-designer \u2192 2. /cct-subagent \u2192 3. /test-first-enforcer \u2192 4. /property-test-generator \u2192 5. /anti-cargo-cult-reviewer \u2192 6. /module-boundary-reviewer \u2192 7. /drift-detector \u2192 8. /implementation-reviewer
+## MANDATORY VERIFICATION -- DO NOT CLOSE TASK WITHOUT THIS
+Before \`bd close\`, you MUST:
+1. Run \`pnpm test\` then \`pnpm lint\` (quality gates)
+2. Run \`/implementation-reviewer\` on changed code -- wait for APPROVED
+If REJECTED: fix ALL issues, re-run tests, resubmit. INVIOLABLE per CLAUDE.md.
+
+The full 8-step pipeline (invariant-designer through implementation-reviewer) is recommended
+for complex changes. For all changes, \`/implementation-reviewer\` is the minimum required gate.
 
 ## Beads Lifecycle
 - \`bd ready\` to find available tasks
@@ -3499,6 +3511,12 @@ Before marking work complete, run the 8-step TDD verification pipeline:
 - Incremental commits made as tests pass
 - All tests pass after refactoring
 - Task lifecycle tracked via beads (\`bd\`)
+
+## PHASE GATE 3 -- MANDATORY
+Before starting Review, verify ALL work tasks are closed:
+- \`bd list --status=in_progress\` must return empty
+- \`bd list --status=open\` should only have Review and Compound tasks remaining
+If any work tasks remain open, DO NOT proceed. Complete them first.
 `,
   review: `---
 name: Review
@@ -3513,18 +3531,19 @@ Perform thorough code review by spawning specialized reviewers in parallel, cons
 ## Methodology
 1. Run quality gates first: \`pnpm test && pnpm lint\`
 2. Search memory with \`memory_search\` for known patterns and recurring issues
-3. Create an AgentTeam (\`TeamCreate\`) and spawn all 11 reviewers as parallel teammates via \`Task\` tool with \`team_name\`:
-   - **Core**: security, architecture, performance, test-coverage, simplicity
-   - **Quality**: docs-reviewer, consistency-reviewer, error-handling-reviewer
-   - **Intelligence**: edge-case-reviewer, pattern-matcher (memory-backed), cct-reviewer (CCT patterns)
-4. Reviewers communicate findings to each other via \`SendMessage\`
-5. Collect, consolidate, and deduplicate all findings
-6. Classify by severity: P1 (critical/blocking), P2 (important), P3 (minor)
-7. Use \`AskUserQuestion\` when severity is ambiguous or fix has multiple valid options
-8. Create beads issues for P1 findings: \`bd create --title="P1: ..."\`
-9. Fix all P1 findings before proceeding
-10. Run \`/implementation-reviewer\` as mandatory gate
-11. Capture novel findings with \`memory_capture\`; pattern-matcher auto-reinforces recurring issues
+3. Select reviewer tier based on diff size:
+   - **Small** (<100 lines): 4 core -- security, test-coverage, simplicity, cct-reviewer
+   - **Medium** (100-500): add architecture, performance, edge-case (7 total)
+   - **Large** (500+): all 11 reviewers including docs, consistency, error-handling, pattern-matcher
+4. Create an AgentTeam (\`TeamCreate\`) and spawn selected reviewers as parallel teammates via \`Task\` tool with \`team_name\`
+5. Reviewers communicate findings to each other via \`SendMessage\`
+6. Collect, consolidate, and deduplicate all findings
+7. Classify by severity: P1 (critical/blocking), P2 (important), P3 (minor)
+8. Use \`AskUserQuestion\` when severity is ambiguous or fix has multiple valid options
+9. Create beads issues for P1 findings: \`bd create --title="P1: ..."\`
+10. Fix all P1 findings before proceeding
+11. Run \`/implementation-reviewer\` as mandatory gate
+12. Capture novel findings with \`memory_capture\`; pattern-matcher auto-reinforces recurring issues
 
 ## Memory Integration
 - Call \`memory_search\` before review for known recurring issues
@@ -3554,6 +3573,13 @@ Perform thorough code review by spawning specialized reviewers in parallel, cons
 - docs-reviewer confirmed docs/ADR alignment
 - All P1 findings fixed before \`/implementation-reviewer\` approval
 - \`/implementation-reviewer\` approved as mandatory gate
+
+## PHASE GATE 4 -- MANDATORY
+Before starting Compound, verify review is complete:
+- \`/implementation-reviewer\` must have returned APPROVED
+- All P1 findings must be resolved
+
+**CRITICAL**: Use \`memory_capture\` MCP tool for ALL lesson storage -- NOT MEMORY.md.
 `,
   compound: `---
 name: Compound
@@ -3565,6 +3591,9 @@ description: Reflect on the cycle and capture high-quality lessons for future se
 ## Overview
 Extract and store lessons learned during the cycle, and update project documentation. This is what makes the system compound -- each session leaves the next one better equipped.
 
+**CRITICAL**: Store all lessons via \`memory_capture\` MCP tool -- NOT via MEMORY.md, NOT via markdown files.
+Lessons go to \`.claude/lessons/index.jsonl\` through the MCP tool. MEMORY.md is a different system and MUST NOT be used for compounding.
+
 ## Methodology
 1. Review what happened during this cycle (git diff, test results, plan context)
 2. Create an AgentTeam (\`TeamCreate\`) and spawn the 6 analysis agents as parallel teammates via \`Task\` tool with \`team_name\`:
@@ -3607,6 +3636,13 @@ Extract and store lessons learned during the cycle, and update project documenta
 - User confirmed high-severity items
 - Beads checked for related issues (\`bd\`)
 - Each item gives clear, concrete guidance for future sessions
+
+## FINAL GATE -- EPIC CLOSURE
+Before closing the epic:
+- Run \`ca verify-gates <epic-id>\` -- must return PASS for both gates
+- Run \`pnpm test\` and \`pnpm lint\` -- must pass
+If verify-gates fails, the missing phase was SKIPPED. Go back and complete it.
+CRITICAL: 3/5 phases is NOT success. All 5 phases are required.
 `
 };