Skip to content

TODO.md

Latest commit

 

History

History
166 lines (111 loc) · 5.53 KB

TODO.md

File metadata and controls

166 lines (111 loc) · 5.53 KB
Raw

aide TODO

Active Work: SDLC-Enhanced Swarm

Overview

Enhance /aide:swarm to use structured SDLC stages for code changes, leveraging Claude Code's native task system.

Project Stack

  • TypeScript (Node.js 20+): Hooks, plugins - vitest, eslint, prettier
  • Go: aide CLI binary

Completed

1. Skill Updates ✅

  • 1.1 Update plan → design skill (output-focused, not interview)
  • 1.2 Create implement skill (TDD implementation for DEV stage)
  • 1.3 Create verify skill (QA validation for VERIFY stage)
  • 1.4 Create docs skill (documentation updates for DOCS stage)
  • 1.5 Delete autopilot skill (replaced by swarm 1 with SDLC)

2. Swarm Skill Enhancement ✅

  • 2.1 Migrate to native tasks (TaskCreate/Update/List instead of aide task)
  • 2.2 Add SDLC mode (story decomposition with stage dependencies)
  • 2.3 Update agent instructions (SDLC pipeline management)

4.4 Enhanced Subagent Context Injection ✅

  • Project memories (project:<name>) now injected to both session-start and subagent-start

3.2 TaskCompleted Hook ✅

  • SDLC stage validation on task completion
  • Parses [story-id][STAGE] pattern from task subject
  • Stage-specific validation (DEV: tests pass, VERIFY: full QA)
  • Works WITHOUT CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS

4.1 Auto Session Summary ✅

  • Captures session summary on Stop hook
  • Extracts: files modified, tools used, user tasks
  • Stores with category=session, tags=session-summary,session:<id>

4.2 Decision Capture ✅

  • <aide-decision topic="..."> tag parsing in PostToolUse
  • Extracts topic, decision, rationale
  • Stores via aide decision set

4.3 /aide:decide Skill ✅

  • Formal decision-making interview workflow
  • Triggers: "decide", "help me decide", "help me choose", "how should we", etc.
  • 6-phase flow: IDENTIFY → EXPLORE → ANALYZE → RECOMMEND → CONFIRM → RECORD

Pending Implementation

3.1 TeammateIdle Hook (Future)

Fires when agent goes idle - validate work complete.

File: src/hooks/teammate-idle.ts (new)

Checks:

  • All claimed tasks complete
  • Tests pass in worktree
  • No uncommitted changes
  • No lint/type errors

Note: Requires CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS enabled.

4.5 Session Memory Injection (Future)

Add current session memories to subagent context.

File: src/hooks/subagent-tracker.ts

Add:

  • Current session memories (session:<current-id>)
  • Task/story-specific memories (task:<id> or story:<id> tags)

5. Qualified Symbol Names in Code Index (Future)

Add qualifier/scope information to the code index so call graph traversal can distinguish auth.GetUser from cache.GetUser.

Steps:

  1. Add @qualifier capture to tree-sitter ref queries in pack.json (all languages, data-only)
  2. Extend parser.go extractReferences to store qualifier on Reference struct
  3. Implement import resolution per language (Go + Rust first, then Python, then TS/JS)
  4. Schema migration: add Qualifier to Reference, QualifiedName to Symbol, secondary index

Enables: Reliable survey_graph BFS, blast-radius-from-diff, execution flow tracing, accurate dead code detection, scoped code_references results.

Effort: ~1 day for steps 1-2 (all languages), ~1 day each for Go/Rust resolution, ~2-3 days for TS/JS resolution. See design doc for full per-language feasibility analysis.

6. LLM-Based Classification of Unknown Files (Future)

Surface unclassified files (not matching any grammar pack) via MCP tools so the calling LLM can classify them. Since aide is an MCP server (not an LLM client), classification must be initiated by the LLM via survey skill tools.

Approach:

  • The topology analyzer already groups files by extension during scanning
  • KindUnclassified constant added to pkg/survey/types.go
  • Add a processUnclassified() pass to topology that emits KindUnclassified entries grouped by extension (e.g., ".tf" -> 12 files, ".proto" -> 3 files)
  • Expose via survey_list / survey_search MCP tools (already works by kind filter)
  • Survey skill can prompt the LLM: "These file extensions are unclassified: .tf, .proto. What languages/technologies do they represent?"
  • LLM responses can be stored as decisions or memories for future sessions

SDLC Stage → Skill Mapping

Stage Skill Purpose
DESIGN design (updated from plan) Output technical spec
TEST test (existing) Write failing tests
DEV implement (new) Make tests pass
VERIFY verify (new) Full QA validation
FIX build-fix (existing) Fix failures
DOCS docs (new) Update documentation

Reference

Claude Code Agent Teams

See: https://code.claude.com/docs/en/agent-teams

Key integrations:

  • Native TaskCreate/TaskUpdate/TaskList - shared across agents
  • TeammateIdle hook - quality gate when agent finishes
  • TaskCompleted hook - quality gate when task completes

Architecture

ORCHESTRATOR (swarm)
    │
    ├── Story A Agent (worktree-a)
    │   └── [DESIGN] → [TEST] → [DEV] → [VERIFY] → [DOCS]
    │
    ├── Story B Agent (worktree-b)
    │   └── [DESIGN] → [TEST] → [DEV] → [VERIFY] → [DOCS]
    │
    └── SHARED NATIVE TASKS (Claude Code)
        All agents see all tasks, dependencies auto-managed