Resolves issue #86 by having the spawn-issues orchestrator create worktrees
upfront and pass the worktree paths to agents, instead of having agents
create their own worktrees in sibling directories outside the sandbox.
Changes:
- spawn-issues orchestrator creates all worktrees before spawning agents
- issue-worker, pr-fixer, code-reviewer accept optional WORKTREE_PATH
- When WORKTREE_PATH is provided, agents work directly in that directory
- Backward compatible: agents still support creating their own worktrees
if WORKTREE_PATH is not provided
- Orchestrator handles all worktree cleanup after agents complete
- Eliminates permission denied errors from agents trying to access
sibling worktree directories
This ensures agents operate within their sandbox while still being able to
work with isolated git worktrees for parallel implementation.
Closes#86
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Set explicit model preferences to optimize for speed vs capability:
- haiku: 11 commands, 2 agents (issue-worker, pr-fixer), 10 skills
Fast execution for straightforward tasks
- sonnet: 4 commands (groom, improve, plan-issues, review-pr),
1 agent (code-reviewer)
Better judgment for analysis and review tasks
- opus: 2 commands (arch-refine-issue, arch-review-repo),
1 agent (software-architect)
Deep reasoning for architectural analysis
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add linter detection logic that checks for common linter config files
(ESLint, Ruff, Flake8, Pylint, golangci-lint, Clippy, RuboCop)
- Add instructions to run linter on changed files only
- Add "Lint Issues" section to review output format
- Clearly distinguish lint issues from logic/security issues
- Document that lint issues alone should not block PRs
Closes#25
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add model: sonnet to issue-worker agent (balanced for implementation)
- Add model: sonnet to pr-fixer agent (balanced for feedback iteration)
- Add model: haiku to /dashboard command (read-only display)
- Add model: haiku to /roadmap command (read-only categorization)
- Document rationale for each model selection in frontmatter comments
Closes#72
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add Architecture Beliefs section to manifesto with outcome-focused beliefs:
auditability, business language in code, independent evolution, explicit over implicit
- Create software-architecture.md as human-readable documentation
- Enhance software-architecture skill with beliefs→patterns mapping (DDD, Event
Sourcing, event-driven communication) and auto-trigger description
- Update work-issue command to reference skill and check project architecture
- Update issue-worker agent with software-architecture skill
- Add Architecture section template to vision-management skill
The skill is now auto-triggered when implementing, reviewing, or planning
architectural work. Project-level architecture choices go in vision.md.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Create the software-architect agent that performs deep architectural
analysis on codebases. The agent:
- References software-architecture skill for patterns and checklists
- Supports three analysis types: repo-audit, issue-refine, pr-review
- Analyzes codebase structure and patterns
- Applies architectural review checklists from the skill
- Identifies anti-patterns (god packages, circular deps, etc.)
- Generates prioritized recommendations (P0-P3)
- Returns structured ARCHITECT_ANALYSIS_RESULT for calling commands
Closes#57
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
spawn-issues now orchestrates the full workflow:
- Phase 1: Spawn issue-workers in parallel, wait for completion
- Phase 2: Review loop - spawn code-reviewer, if needs work spawn pr-fixer
- Phase 3: Report final status
issue-worker simplified:
- Removed Task tool and review loop
- Just implements, creates PR, cleans up
- Returns structured result for orchestrator to parse
Benefits:
- Better visibility into progress
- Reuses pr-fixer agent
- Clean separation of concerns
- Orchestrator controls review cycle
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add Task tool to spawn code-reviewer synchronously
- Add review loop: fix issues and re-review until approved (max 3 iterations)
- Add final summary format for cleaner output to spawning process
- Reviewer works in same worktree, cleanup only after review completes
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Background agents need skills specified in frontmatter rather than
using @ syntax which may not expand for Task-spawned agents.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
New command that spawns background agents to work on multiple
issues simultaneously, each in an isolated git worktree.
- commands/spawn-issues.md: Entry point, parses args, spawns agents
- agents/issue-worker/agent.md: Autonomous agent that implements
a single issue (worktree setup, implement, PR, cleanup)
Worktrees are automatically cleaned up after PR creation.
Branch remains on remote for follow-up work if needed.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Transforms the product-manager from a reactive backlog manager into a
vision-driven system with continuous improvement capabilities.
New components:
- vision-management skill: How to create, maintain, and evolve product vision
- /vision command: View, create, or update product vision (syncs to Gitea)
- /improve command: Identify gaps between vision goals and backlog
Enhanced existing components:
- product-manager agent: Now vision-aware with strategic prioritization
- /retro command: Connects learnings back to vision updates
- /plan-issues command: Shows vision alignment for planned work
The vision lives in two places: vision.md (source of truth) and a Gitea
issue labeled "vision" for integration with the issue workflow.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The approval step was failing on self-authored PRs and stopping the
merge flow. Since LGTM verdict already indicates approval, just merge
directly without the separate tea pulls approve command.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The tea CLI doesn't have a command to output PR diff content directly.
The -f diff flag only returns a URL. Updated docs to use tea pulls
checkout followed by git diff main...HEAD.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The skills are already listed in frontmatter (skills: gitea, issue-writing,
backlog-grooming, roadmap-planning), so the "Use the gitea skill"
instruction was unnecessary.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add Comments section to gitea skill with examples
- Update code-reviewer agent with explicit tea CLI commands
for posting comments, approving, and merging PRs
Closes#6🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Replace fj CLI with tea CLI across all commands
- Create new gitea skill, remove forgejo skill
- Update all agents to use gitea skill
- Update commands to use skill-based approach (reference skills instead of embedding CLI commands)
- Update all documentation (README, ARCHITECTURE, VISION, writing guides)
- Swap git remotes: origin now points to git.flowmade.one (Gitea)
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add explicit example and warning to prevent agents from misinterpreting
the fj pr view diff command. The correct format is:
fj pr view <number> diff
NOT:
fj pr diff <number>
Closes#25🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
When the code-reviewer agent determines a PR is LGTM with no blocking
issues, it now automatically merges using rebase strategy and deletes
the branch. PRs with "Needs Changes" or "Blocking Issues" verdicts
are left for the user to address.
Closes#23🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Both agents now use `sonnet` instead of `inherit`:
- code-reviewer: Sonnet provides good code understanding for
structured review tasks without needing opus-level reasoning
- product-manager: Planning and issue-writing follow patterns
from skills; sonnet handles these well
Each agent includes a comment explaining the model choice.
Closes#18🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Update product-manager agent to use official Claude Code format
- Update ARCHITECTURE.md to document frontmatter fields
- Add code-reviewer agent and code-review skill to docs
- Update README.md project structure
This aligns our agent documentation with the official Claude Code
agent format which uses YAML frontmatter for name, description,
model, skills, and tools.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Follow official Claude Code agent format with:
- name, description, model, skills in frontmatter
- System prompt as markdown content
- Streamlined documentation focused on agent behavior
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Create specialized agent for automated PR code review that:
- Fetches PR diffs via fj CLI
- Analyzes code for quality, bugs, security, style, and test coverage
- Posts structured review comments to PRs
Includes code-review skill with guidelines for reviewing code changes.
Closes#9🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
