Sprint-plan approval workflow:
- Request explicit approval after creating issues
- Present scope summary (branches, files, dependencies)
- User must type "approve sprint N" to authorize
- Record approval in milestone description with timestamp
Sprint-start verification:
- Check milestone for "## Sprint Approval" section
- If missing, STOP and direct to /sprint-plan
- Extract approved scope (branches, files)
- Enforce scope during execution
Orchestrator scope enforcement:
- Verify approval before any execution
- Check each operation against approved scope
- Operations outside scope require re-approval
This separates planning (review) from execution (action),
preventing agents from executing without explicit user consent.
Closes#233
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Executor checkpointing:
- Standard checkpoint comment format with branch, commit, phase
- Files modified with status (created, modified)
- Completed and pending steps tracking
- State notes for resumption context
- Save checkpoint after major steps, before stopping
Orchestrator resume detection:
- Scan issue comments for "## Checkpoint" markers
- Offer resume options: resume, start fresh, review details
- Verify branch exists and files match before resuming
- Dispatch executor with checkpoint context
Sprint-start integration:
- Checkpoint detection as first workflow step
- Resume flow documentation with example
- Checkpoint format specification
This enables resuming work after:
- Budget exhaustion (100 tool call limit)
- Agent failure/circuit breaker
- Manual interruption
- Session timeout
Closes#237
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add pre-dispatch conflict detection:
- Analyze target files for each task before parallel dispatch
- Check for file overlap between tasks in same batch
- If overlap detected, sequentialize those specific tasks
- Example analysis showing conflict detection workflow
Branch isolation protocol:
- Each task MUST have its own branch
- Never have two agents work on the same branch
- Sequential merge after completion (not simultaneous)
- Handle merge conflicts by stopping second task
Conflict resolution rules:
- Same file → MUST sequentialize
- Same directory → Usually safe, review
- Shared config → Sequentialize
- Shared test fixture → Sequentialize or assign different files
This prevents parallel agents from modifying the same files
and causing git merge conflicts.
Closes#234
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add structured progress comment format for agents:
- Standard format with Status, Phase, Tool Calls budget
- Completed/In-Progress/Blockers/Next sections
- Clear examples for starting, blocked, and failed states
- Guidance on when to post (every 20-30 tool calls)
Update sprint-status.md:
- Document how to parse progress comments
- Show enhanced in-progress display with tool call tracking
- Add progress comment detection to blocker analysis
This enables users to see:
- Real-time agent progress
- Tool call budget consumption
- Current phase and step
- Blockers as they occur
Closes#232
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add Status/In-Progress, Status/Blocked, Status/Failed, Status/Deferred labels
- Update orchestrator.md with Status Label Management section
- Update executor.md with honest Status Reporting requirements
- Update labels-reference.md with Status detection guidelines
Status labels enable accurate tracking:
- In-Progress: Work actively being done
- Blocked: Waiting for dependency/external factor
- Failed: Attempted but couldn't complete
- Deferred: Moved to future sprint
Agents must report honestly - never say "completed" when blocked/failed.
Closes#235
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The _get_current_branch() method was running git commands from the
installed plugin directory instead of the user's project directory.
This caused incorrect branch detection (always seeing 'main' from
the marketplace repo instead of the user's actual branch).
Fix: Use CLAUDE_PROJECT_DIR environment variable to get the correct
project directory and pass it as cwd to subprocess.run().
Fixes#231
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Implements PostToolUse hook to warn about breaking interface changes:
- Detects changes to plugin.json, hooks.json, .mcp.json, agents/*.md
- Compares with git HEAD to find removed/changed elements
- Warns on: removed hooks, changed matchers, removed MCP servers
- Warns on: plugin name changes, major version bumps
- Non-blocking, configurable via CONTRACT_VALIDATOR_BREAKING_WARN
Depends on #229 (SessionStart auto-validate infrastructure).
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Implements smart SessionStart hook for plugin validation:
- Validates plugin contracts only when files change (smart mode)
- Caches file hashes to avoid redundant checks
- Non-blocking warnings for compatibility issues
- Configurable via CONTRACT_VALIDATOR_AUTO_CHECK
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Implements UserPromptSubmit hook to detect vague prompts:
- Checks for short prompts without context
- Detects ambiguous phrases ("fix it", "help me with")
- Suggests /clarity-assist when beneficial
- Non-blocking, configurable via CLARITY_ASSIST_AUTO_SUGGEST
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Implements PostToolUse hook to warn about potentially breaking schema changes:
- DROP COLUMN/TABLE/INDEX detection
- Column type changes (ALTER TYPE, MODIFY COLUMN)
- NOT NULL constraint additions
- RENAME operations
- ORM model field removals
Non-blocking - outputs warnings only.
Configurable via DATA_PLATFORM_SCHEMA_WARN env var.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The Claude plugin cache at ~/.claude/plugins/cache/leo-claude-mktplace/
holds versioned copies of .mcp.json files. When we update .mcp.json
to use run.sh instead of direct python paths, the cache retains old
versions, causing MCP servers to fail.
Now post-update.sh clears this cache, forcing Claude to read fresh
configs from the installed marketplace on next session start.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The branch permission check was only allowing feat/, feature/, and dev/
prefixes for write operations. This blocked PR creation from fix/*
branches, forcing fallback to direct API calls.
Added patterns:
- fix/, bugfix/, hotfix/ - for bug fixes
- chore/, refactor/ - for maintenance
- docs/, test/ - for documentation and tests
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Replace direct python path in .mcp.json with run.sh wrapper scripts
that automatically locate the venv in cache or local directory.
Problem: .venv symlinks are gitignored, causing them to be wiped on
every git operation. The SessionStart hook should recreate them but
this was unreliable, leading to repeated MCP server failures.
Solution: run.sh scripts that:
- First check ~/.cache/claude-mcp-venvs/leo-claude-mktplace/{server}/.venv
- Fallback to local .venv if exists
- Exit with helpful error if neither found
This eliminates dependency on symlinks entirely - the scripts are
tracked in git and will always be present after clone/pull/update.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add missing create_pull_request tool to Gitea MCP server. This completes
the PR lifecycle - previously only had list/get/review/comment tools.
- Add create_pull_request to GiteaClient
- Add async wrapper to PullRequestTools with branch permissions
- Register tool in server.py with proper schema
- Parameters: title, body, head, base, labels (optional)
- Branch-aware security: only allowed on development/feature branches
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Replace the old approach (create venvs in marketplace directory) with
the new venv-repair.sh approach (symlinks to external cache).
This ensures post-update.sh:
- Instantly restores symlinks if cache exists
- Only does full pip install on first run
- Works correctly after marketplace updates
Flow after this fix:
Update marketplace → post-update.sh → venv-repair → Session works
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Problem:
- Venvs in marketplace directory got deleted on every update
- Users had to manually run setup.sh and wait for full pip install
- This caused MCP servers to fail until manually fixed
Solution:
- Store venvs in external cache (~/.cache/claude-mcp-venvs/)
- Auto-repair symlinks via SessionStart hook (instant operation)
- Only run pip install on first use or when requirements change
Architecture:
Cache (runtime) → Marketplaces → External venv cache
The chain of symlinks ensures all three locations work:
1. ~/.claude/plugins/cache/.../mcp-servers/* (runtime)
2. ~/.claude/plugins/marketplaces/.../mcp-servers/* (install)
3. ~/.cache/claude-mcp-venvs/* (persistent venvs)
Performance:
- First install: ~2-3 min (unchanged)
- After marketplace update: 0.03 sec (was 2-3 min)
Files:
- scripts/venv-repair.sh: Fast symlink restoration for hooks
- scripts/setup-venvs.sh: Full setup with external cache
- plugins/projman/hooks/startup-check.sh: Auto-repair on session start
- .gitignore: Ignore .venv symlinks
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Previously post-update.sh would only warn when venvs were missing,
requiring a separate setup.sh run. Now it automatically creates
missing venvs and installs dependencies including editable packages.
Also added viz-platform and contract-validator to the update list.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The filter tool was adding an __index_level_0__ column to results
because pandas query() preserves the original index, which gets
converted to a column when storing the DataFrame.
Added .reset_index(drop=True) after query() to drop the preserved
index and create a clean sequential index.
Fixes#203
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
