- New skill: token-budget-report.md with MCP overhead and skill loading
estimation model based on known file sizes and static overhead tables
- Sprint-plan displays token report after planning complete
- Sprint-close displays token report after lessons captured
- Visual-output.md updated with token report format
- .gitignore exception for token-budget-report.md (not API tokens)
Helps identify which phases and components consume the most context budget.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
BREAKING CHANGE: All generic command names now prefixed with plugin namespace.
See CHANGELOG.md v7.0.0 for complete rename map.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Separate cognitive work from mechanical API execution to reduce
skill-related token consumption by ~76-83% during sprint workflows.
Changes:
- Add batch-execution.md skill with 4-phase protocol
- Promote mcp-tools-reference and batch-execution to frontmatter
for planner and orchestrator agents (auto-injected, zero re-read)
- Replace "Skills to Load" with phase-based "Skill Loading Protocol"
- Restructure planning-workflow.md Steps 8-10 for batch execution
- Update agent matrix in CLAUDE.md and docs/CONFIGURATION.md
- Add Phase-Based Skill Loading documentation section
- Clean up .gitignore (transient files, dev symlinks)
Token impact:
- 6-issue sprint planning: ~76% reduction
- 10-issue sprint planning: ~80% reduction
- 8-issue status updates: ~83% reduction
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Test Fixes:
- Fix mock_config fixture to use 'owner/repo' format (was separate fields)
- Update test_client_initialization to match current client API
- Add required 'org' argument to get_org_labels, list_repos, aggregate_issues tests
- Update error message assertion in test_no_repo_specified_error
- Fix test_create_issue to mock is_org_repo and label resolution
- Update aggregate_issues tests in test_issues.py with org argument
Documentation Updates:
- Expand tools table from 8 to 36 tools (organized by category)
- Update directory structure to show all 6 tool files
- Remove unused GITEA_OWNER from configuration docs
- Add automatic repository detection documentation
- Add project directory detection strategies
- Update test count from 42 to 64
- Create CHANGELOG.md with full version history
All 64 tests now pass. No production code changes.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Extract tool definitions and dispatcher from server.py into tool_registry.py
to enable transport-agnostic reuse. External consumers (e.g., HTTP transport
in gitea-mcp-remote) can now import and use the Gitea MCP tools without
duplicating code.
Changes:
- Create pyproject.toml with PEP 621 compliant package manifest (hatchling)
- Create tool_registry.py with get_tool_definitions() and create_tool_dispatcher()
- Refactor server.py to use registry (1100 -> 93 lines)
- Update __init__.py with package exports and __version__
The tool_filter parameter enables selective tool exposure for remote servers.
Stdio transport behavior is unchanged - all 36 tools still work identically.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Plugin installation scripts for consumer projects
- MCP server mapping via mcp_servers field in plugin.json
- CLAUDE.md section markers for install/uninstall
- Agent model selection (25 agents with model frontmatter)
- Agent frontmatter standardization
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The mcp_servers key is not a valid key in the Claude plugin manifest
schema. MCP servers are configured in the root .mcp.json file instead.
Affected plugins:
- cmdb-assistant
- contract-validator
- data-platform
- pr-review
- projman
- viz-platform
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Note: This fix may not help because MCP servers fail BEFORE hooks run.
See lessons learned wiki page for full debug trace.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Plugin load failures were caused by version mismatch between
marketplace.json and individual plugin.json files:
- contract-validator: 1.2.0 vs 1.1.0
- git-flow: 1.2.0 vs 1.0.0
- projman: 3.4.0 vs 3.3.0
- clarity-assist: 1.2.0 vs 1.0.0
- doc-guardian: 1.1.0 vs 1.0.0
Claude Code silently fails to load plugins when versions don't match.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Startup hooks in data-platform and pr-review were checking for venvs
at the marketplace path (~/.claude/plugins/marketplaces/.../mcp-servers/)
which gets wiped on updates. The actual venvs live in the cache directory
(~/.cache/claude-mcp-venvs/) which survives updates.
This caused false "MCP venv missing" errors even when venvs existed,
wasting hours of debugging time.
Fixed hooks now check cache path first, matching the pattern used
by run.sh scripts.
Also updated docs/CANONICAL-PATHS.md with the correct venv path pattern
to prevent future occurrences.
Lesson learned: lessons/patterns/startup-hooks-must-check-venv-cache-path-first
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The startup hooks were looking for MCP venvs relative to the plugin
directory instead of the marketplace root, causing false "venv missing"
errors on every session start.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add per-agent model selection using Claude Code's now-supported `model`
frontmatter field, and standardize all agent frontmatter across the
marketplace.
Changes:
- Add `model` field to all 25 agents (18 sonnet, 7 haiku)
- Fix viz-platform/data-platform agents using `agent:` instead of `name:`
- Remove non-standard `triggers:` field from domain agents
- Add missing frontmatter to 13 agents
- Document model selection in CLAUDE.md and CONFIGURATION.md
- Fix undocumented commands in README.md
Model assignments based on reasoning depth, tool complexity, and latency:
- sonnet: Planner, Orchestrator, Executor, Coordinator, Security Reviewers
- haiku: Maintainability Auditor, Test Validator, Git Assistant, etc.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
