Session isn't finished until delivered work is written back to the project's
canonical planning store (cp-cli / local checkout / offline staging), and
roll-up docs are checked for contradictions. Motivated by the agent-runtimes
work-items repo drifting 6 weeks behind delivered reality.
Wraps scripts/scan-ralph-candidates to inject scan output at load time.
Shows unmerged agent-fork branches with changed files and inferred spec IDs.
Designed for /loop 5m /scan-ralph continuous polling.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Walks CONTEXT.md and CLAUDE.md files from a given folder up to the git
root, assembles them into a context bundle via shell redirects (file
contents never enter the calling session), then delegates a one-line
question to MiniMax M2.7. Token cost to the main session: paths only.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Implements DS-1..DS-13 from spec/dispatch-skill.md. Reads
.agent-tasks.json, validates structure, submits tasks to the control
plane in dependency-ordered waves, and records cp_task_ids atomically
after each submission. Supports --wait-wave, --continue-on-failure, and
--as-manifest flags.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The skill now loads mechanical-test-generation.md as a 6th core file alongside
spec-driven-development, test-driven-development, api-design, llm-code-security,
and security-architecture. The previous "load related files based on the stack"
hint is replaced with an explicit conditional triage table covering agent-repos,
ai-parallel-agents, api-integration, secrets-management, database-selection,
docker, kubernetes, and validation, plus a skip list of implementation-time
topics that don't add value at spec-review time.
Adds a new "Review: mechanical test derivation" section with seven checks
(module layout, integration boundaries, library semantics, concrete interfaces,
error messages as test data, parametric pattern tables, pre-dispatch testability)
and a corresponding scorecard row.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- skills/log/SKILL.md: pre-gathers unprocessed transcript metadata (compact
JSON via list-transcripts-here.sh) rather than full content; spawns a
Sonnet subagent (not Haiku — gotcha detection requires judgment) to read
JSONL backups, write a companion -transcripts.md log, and mark backups
processed in tracking.json; adds Agent and Bash(bash|pwd|python3) to
allowed-tools
- skills/switch-mode/SKILL.md: add previously untracked skill to repo
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Run a workflow template from a Claude Code session via the local control
plane. Handles topological wave execution, artifact passing, sentinel
resolution, and human review gates locally — no CP-side workflow
expansion required.
The pre-gathered "Available workflows" context uses python3 (not a
ls|sed pipe) so the Bash permission checker accepts it without needing
sed in allowed-tools — piped commands fail the check even when both
sides match an allowed pattern individually.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Two new read-only skills that review plans and specs against best practices:
- /review-plan: Checks plans against 6 areas (security, API design, LLM code
security, spec-driven dev, TDD, operational readiness). Outputs scorecard
with critical gaps and recommendations.
- /review-spec: Checks specs for structure quality, requirement testability,
security coverage, and API design patterns. Scores 5 dimensions and
identifies missing requirements and scenarios.
Both load api-design.md, llm-code-security.md, spec-driven-development.md,
test-driven-development.md, and security-architecture.md from best-practices.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Adds decisions and process-lessons from recent reflections.
Updates decompose and orchestrate SKILL.md with operational improvements.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
All paths now point to ~/dev/claude/projects/best-practices/ instead of
claude-foundations/best-practices/. The settings.yaml remains in
claude-foundations (where the pipeline config lives).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
New skills for container agent task orchestration:
- /decompose: break tasks into subtasks with dependency graph, write .agent-tasks.json
- /orchestrate: check task state, launch container agents in git worktrees, poll for completion
Updated with learnings from first real run (agent-runtimes M3, 9 tasks):
- Prompts must end with "run pytest and fix failures"
- State import conventions explicitly in prompts
- Warn about uncommitted files before decomposing (worktrees need committed content)
- Copy dependency outputs into downstream worktrees before launching
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replace non-ASCII characters in log and distill-best-practices
frontmatter that silently prevented skill loading. Simplify distill
skill instructions. Update install.sh for new skills.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Scan mode checks formatter setup; run mode formats all project files.
Works with the claude-foundations formatter infrastructure.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Gathers CLAUDE.md, CONTEXT.md, MEMORY.md, BESTPRACTICES.md, directory
trees, and git status from the directory hierarchy after /clear.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
unreflected-logs, git-status-report, check-skills, and validate-skill
all exit 1 as a status signal when issues are found. The skill's bang-
command syntax treated this as a failure, aborting mid-run. Append
`|| true` so informational output is always captured.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Claude Code's Bash permission checker rejects \${VAR} parameter
substitution in skill ! commands. Replace all occurrences with \$VAR
(identical shell expansion) across log, reflect-logs, and
distill-best-practices skills. Rewrite \${VAR:-default} to use
test -n / || pattern instead.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Skills used absolute paths (~/dev/claude/projects/claude-foundations/...)
which get blocked by Claude Code's sandbox when running from other projects.
Changed to relative paths (../claude-foundations/) with local-first fallback
and inline defaults so skills work regardless of sandbox restrictions.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Matches the move of best-practices/INDEX.md to BESTPRACTICES.md in
claude-foundations root for context-load auto-discovery.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- /log: end-of-session capture of decisions, gotchas, open questions to memory/log/
- /reflect-logs: processes session logs into topic memory files with staleness detection
- /distill-best-practices: cross-project best practices extraction with interactive proposals
- README.md: updated skill table with all three new skills
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The !`command` blocks in SKILL.md go through Claude Code's Bash permission
checker, which rejects $() command substitution. Simplified the git log
command, replaced dynamic verify script path with a Read tool instruction,
and narrowed allowed-tools to specific command patterns. Also documented
the constraints in CLAUDE.md for future skill development.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Reusable Claude Code skills shared across projects via symlinks
into ~/.claude/skills/. Includes the /reflect skill for structured
milestone reflections.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
