crew-research

Dispatch Codex to review all unreviewed repository work, then verify its correlated findings ticket or clean result. Use when requesting an independent Codex review, auditing all changes, checking work since the last review, or waiting for Codex findings. Trigger: dispatch codex review, independent review, review all changes, check codex findings, run review agent.

Always-on: when a project has tickets, work the frontier (first unblocked open ticket), mark done, update plan status.

Incremental and full-history repository review with resumable adoption batches. Use when reviewing new work, initializing review tracking, reviewing all work to date, resuming historical review, inspecting unreviewed commits, or reviewing changed tickets. Trigger: review new work, adopt repository, review all history, resume review, review marker, unreviewed commits.

End-of-session handoff that captures current state for the next session. Use when ending a work session, switching context, or before a long break.

Consolidate project artifacts — promote scratch to memory, deduplicate memory, process decisions, organize scripts, update steering/skills accuracy. Use periodically or when the project feels cluttered.

Run and interpret crew-research evals — run.sh flags, scores.jsonl fields, activation TPR/FPR verdicts, resuming interrupted runs. Use when running evals, reading eval results, diagnosing a failing definition, or resuming a dead run. Trigger: run the evals, eval results, scores.jsonl, activation test, TPR, delta threshold, resume the run, skip-completed, known gap.

Best practices for writing tool and helper scripts. Use when creating bash scripts, automation tools, CLI utilities, or any executable that will be run repeatedly by humans or agents.

How to process images in kiro-cli sessions — kiro-cli vs Claude multimodal limits, sizing, and the fresh-session workaround for context compaction losing inline images.

Run and interpret crew-research deploy tooling — init.sh, doctor.sh, catalog.sh, prune semantics, manifests, ADR 0009 reference placement. Use when deploying tiers, diagnosing skill/steering deployment, reading doctor warnings, or verifying deploy idempotency. Trigger: deploy the tier, init.sh, doctor output, pruned, unmanaged skill, crew-skills manifest, steering references, redeploy.

Run and interpret session transcript analysis — skill activation rates, steering compliance, tool usage across kiro sessions. Use when measuring skill field usage, checking recall-check compliance, reviewing what skills actually activate, or judging skill retirement. Trigger: session review, skill usage, field compliance, steering compliance, never activated, session:skills, session:parse.

Style guide for writing behavioral eval criteria that produce consistent LLM-judged scores. Use when creating, reviewing, or modifying eval definitions.

Write concise, effective AGENTS.md files for projects. Use when creating, rewriting, or trimming an AGENTS.md. Trigger: AGENTS.md, agent guidance, project instructions, agent-facing docs, too long, bloated.

Partition MCP servers across a lean default agent and specialist agents to avoid tool bloat. Use when sessions feel slow or tool selection misfires, when adding an MCP server, when designing agent configs, or when validating that agents actually load their tools. Trigger: tool bloat, too many tools, mcp server placement, agent config, specialist agent, slim default session, tools whitelist.

Subagent dispatch reliability patterns. Expect failures, design around prompt size limits, use write-then-read for synthesis tasks.

Probe the current session for self-improvement opportunities — corrections, friction, gotchas, and repetition that should become updates to project-local skills, AGENTS.md, steering, or tool-script guides. Invoke manually (/guidance-sync) periodically during a session or before wrapping up. Trigger: guidance sync, self improvement, what should we update, capture learnings, sync guidance, improvement opportunities, session retro.

Drift analysis — check if deployed skills, steering, and AGENTS.md still match project reality. Use periodically or after major refactors.

Write Architecture Decision Records when making significant technical choices. Use when selecting tools, patterns, frameworks, or making structural changes that future developers need to understand.

Find deepening opportunities and run architecture review sessions — refactors that turn shallow modules into deep ones, plus the acceptance loop where the user approves or rejects refactoring candidates. Use when improving architecture, consolidating tightly-coupled modules, reducing complexity, or mid-review when the user accepts/approves candidates ("accept candidate 3", "looks good, keep it", "sounds right"). Trigger: "shallow modules", "deepen", "architectural friction", "consolidate", "simplify interfaces", "refactoring review", "accept candidate", "approve this refactor", "architecture review".

Design interrogation with evidence-backed recommendations. Researches each question via web search before presenting options. Updates CONTEXT.md and offers ADRs inline. Dispatches spikes for empirical validation. Trigger: grill, grill me, stress-test this plan, interrogate this design, poke holes, challenge my assumptions.

Structure work through phased planning — brainstorm, validate, scope, finalize. Use when starting new features, breaking down epics, or when a task feels too vague to implement directly.

Plan projects using PLAN.md and per-feature spec files. Use when starting a new project, breaking a project into phases, writing specs for features, or when someone says 'let's plan this out' or 'write a spec'. Trigger: PLAN.md, spec, feature spec, project plan, phases, task graph, validation criteria.

Workspace conventions enforced every turn: glossary upkeep, document placement (.scratch vs .memory vs docs), validation contract for scripts, and git discipline for convention projects.

Cross-session memory recall. Use when asked about past decisions, prior work, what was discussed previously, or to continue from where a session left off. Use to persist decisions and lessons learned. Trigger: "what did we decide", "last session", "previously", "recall", "remind me", "continue from where we left off", "what was the decision".

Write and improve agent-loadable skills. Use when creating a new skill, improving an existing skill's activation, restructuring a skill that's too broad, writing skill frontmatter, or diagnosing why a skill doesn't trigger. Trigger: new skill, write a skill, skill format, skill template, activation trigger, skill description, SKILL.md.

Cut tagged releases safely with SemVer, changelog roll, and validation gates. Use when releasing, bumping version, tagging, shipping, or deciding what version to use. Trigger: release, tag, version bump, ship, publish, cut a release, what version.

Start-of-session orientation — read the handoff and continue work. Use when starting a new session, resuming after a break, or picking up someone else's work.

Run and interpret crew-research platform proofs — run.sh adapter proofs (A/C/G series), run-proof.sh subagent reliability proofs (S series), result reading. Use when validating tool adapter behavior, testing skill discovery/isolation, or checking platform assumptions. Trigger: run proofs, proof harness, adapter proof, A-series, S-series, platform assumption, skill discovery test, subagent isolation.

When to search cross-session memory before answering.

Install missing development tools (kiro-cli, codex, agy, crush, tk, yq, mise, jq, bc) with OS detection, per-OS commands, corporate-proxy recipes, and verification. Use when a command fails with 'not found' or a prerequisite tool is missing. Trigger: not found, install tool, missing tool, winget, brew, go install proxy error.

Build a pass/fail signal BEFORE attempting fixes. Use when debugging, diagnosing failures, tests failing, errors recurring, or when a fix didn't work. Trigger: debug, failing test, broken, not working, TypeError, error output, diagnose.

Quick reference for all available skills and workflows. Use when you need a reminder of what's available, or when you don't know where to start.

Code quality rules that catch common AI-generation artifacts. Apply when writing code, generating functions, implementing features, creating scripts, or producing any code output. Catches: redundant defensive checks, gratuitous logging, restating comments, unnecessary casts, over-abstraction, verbose implementations. Always relevant during code generation tasks.

Manage context as a finite resource. Compress between phases, reinforce objectives in long sessions, decay old detail. Use when sessions are long, multi-phase, or approaching limits.

Validate artifacts with multiple independent AI tools — images, code, documents, designs. Use when a single perspective has blind spots, for visual output validation, or when validators disagree. Trigger: multi-agent validation, independent validators, cross-check with another tool, second opinion, validation crew.

Plan project milestones using the right methodology for your uncertainty level. Routes between spikes (learn), tracer bullets (prove architecture), and vertical slices (deliver features). Use when stuck, planning next milestone, or project has broad progress but no end-to-end path working. Trigger: plan a milestone, what should I work on next, project feels stuck, too many things to do, replan, tracer bullet.

Select the right enforcement mechanism for agent behavior rules. Use when configuring agent constraints, writing AGENTS.md, designing skills, or deciding how to enforce a new rule.

Structured research process for investigating topics and producing findings. Use when exploring unfamiliar domains, evaluating options, or gathering evidence to inform decisions.

Batch study all repos/tools in .references/. Dispatches parallel subagents to document each, then distills learnings into steering/skills. Use when onboarding to a project with multiple reference repos.

Deep-dive a tool or repo in .references/. Documents purpose, usage, novel patterns, and distills learnings into steering/skills. Use when onboarding a new reference repo or tool.

Two-axis code review: Standards (code quality, smells, security) and Spec (does it match requirements?). Use when reviewing a diff, PR, branch, or implementation. Trigger: review this code, review the diff, PR review, code review, check this implementation, review since main.