Sonnet 4.6 is too slow for small stories — agents burn through turns
without completing. Reverting coders, QA, and mergemaster to Sonnet 4.5.
Supervisor and coder-opus remain on Opus.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
model = "sonnet-4.6" is not a valid Claude CLI model identifier,
causing all coder/qa/mergemaster agents to get 404 errors from the
API and exit immediately with no work done. Change to
"claude-sonnet-4-6". Also disable automatic worktree cleanup on
archive so agent work can be inspected post-mortem.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Configures sparse checkout on new and existing worktrees to exclude
the pipeline state directory. This prevents feature branches from
containing .story_kit/work/ file moves that cause rename/delete merge
conflicts when merging back to master.
Also removes "pick up the story from .story_kit/work/" instruction
from agent prompts since the story content is already in the prompt.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Adds an opus-tier coder agent for complex tasks requiring deep
architectural understanding (80 max turns, $20 budget).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
When an agent process exits normally, the server now automatically runs
acceptance gates (uncommitted changes check + cargo clippy + tests) and
advances the pipeline based on results. This replaces the previous model
where agents had to explicitly call report_completion as an MCP tool.
Changes:
- Add run_server_owned_completion() free function in agents.rs that runs
gates on process exit, stores a CompletionReport, and advances pipeline
- Wire it into start_agent's spawned task (replaces simple status setting)
- Remove report_completion from MCP tools list and handler (mcp.rs)
- Update default_agent_prompt() to not reference report_completion
- Update all agent prompts in project.toml (supervisor, coders, qa,
mergemaster) to reflect server-owned completion
- Add guard: skip gates if completion was already recorded (legacy path)
- Add 4 new tests for server-owned completion behavior
- Update tools_list test (26 tools, report_completion excluded)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add `qa` agent entry to `.story_kit/project.toml` with a detailed
prompt covering code quality scan, test verification, manual testing
support, and structured report generation
- Add `move_story_to_qa` function in `agents.rs` that moves a work item
from `work/2_current/` to `work/3_qa/` and auto-commits (idempotent)
- Add `request_qa` MCP tool in `mcp.rs` that moves the story to
`work/3_qa/` and starts the QA agent on the existing worktree
- Add unit tests for `move_story_to_qa` (moves, idempotent, error cases)
- Update `tools_list_returns_all_tools` test to expect 27 tools
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- report_completion MCP tool for agents to signal done
- Rejects if worktree has uncommitted changes
- Runs acceptance gates (clippy, tests) automatically
- Stores completion status on agent record
- 10 new tests
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add base_branch detection to WorktreeInfo (from project root HEAD)
- Expose base_branch in AgentInfo API response
- Add {{base_branch}} template variable to agent config rendering
- Show git difftool command with copy-to-clipboard in AgentPanel UI
- Add diff command instruction to coder agent prompts
- Add AgentPanel tests for diff command rendering and clipboard
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Supervisor reads CLAUDE.md and README before coordinating
- Coders commit their work before finishing
- All agents stop at "gates green" - never accept or merge
- Human always does final acceptance and merge to master
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Fix CLAUDECODE env var blocking nested Claude Code sessions
- Add drain-based event_log for reliable get_agent_output polling
- Add non-SSE get_agent_output fallback (critical for MCP tool calls)
- Preserve worktrees on agent stop instead of destroying work
- Reap zombie processes with child.wait() after kill
- Increase broadcast buffer from 256 to 1024
- Engineer supervisor and coder prompts in project.toml
- Point .mcp.json to test port 3002
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace single [agent] config with multi-agent [[agent]] roster system.
Each agent has name, role, model, allowed_tools, max_turns, max_budget_usd,
and system_prompt fields that map to Claude CLI flags at spawn time.
- AgentConfig expanded with structured fields, validated at startup (panics
on duplicate names, empty names, non-positive budgets/turns)
- Backwards-compatible: legacy [agent] format auto-wraps with deprecation warning
- AgentPool uses composite "story_id:agent_name" keys for concurrent agents
- agent_name added to AgentEvent variants, AgentInfo, start/stop/subscribe APIs
- GET /agents/config returns roster, POST /agents/config/reload hot-reloads
- POST /agents/start accepts optional agent_name, /agents/stop requires it
- SSE route updated to /agents/:story_id/:agent_name/stream
- Frontend: roster badges, agent selector dropdown, composite-key state
- Project root initialized to cwd at startup so config endpoints work immediately
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add git worktree isolation for concurrent story agents. Each agent now
runs in its own worktree with setup/teardown commands driven by
.story_kit/project.toml config. Agents stream output via SSE and support
start/stop lifecycle with Pending/Running/Completed/Failed statuses.
Backend: config.rs (TOML parsing), worktree.rs (git worktree lifecycle),
refactored agents.rs (broadcast streaming), agents_sse.rs (SSE endpoint).
Frontend: AgentPanel.tsx with Run/Stop buttons and streaming output log.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
