AI Agent Workflow Collection

Run a local pre-PR Codex code review (adaptation cycle) via codex-plugin-cc, capture structured output, normalize findings, and return them to the caller. Review-only — does not fix findings. Invoked by local-reviewer agent only.

Check status of all broods, with per-strain status. Reports per-strain tmux session state, branch existence, and PR status from external observables. Trigger: "brood status", "check brood", "brood-status", "brood progress", "how's the brood", "fleet status".

Create or confirm the compliant working branch for the current approved plan before implementation begins.

Analyze project artifacts and generate a populated CONTEXT.md (or CONTEXT-MAP.md and multiple CONTEXT.md files for multi-context repos) with domain terms extracted from existing code, documentation, and configuration. Use when seeding a new hive, analyzing an existing project, or enriching an existing CONTEXT.md. Trigger on: "generate a domain glossary", "create a CONTEXT.md from project artifacts", "analyze project artifacts for domain terms", "enrich CONTEXT.md with terms from project", "update CONTEXT.MD", "update CONTEXT-MAP.md", "update project context definitions".

Detect break-fix, root-cluster, diminishing-returns, and stop-and-merge signals from a fix-ledger before a reviewer decides whether to keep patching. Use when a reviewer needs one verdict over the accumulated findings to judge whether the next per-finding patch is a treadmill, a prior fix is being undone, the loop is plateauing, or the surface is hardened enough to advise merge. Trigger: "detect remediation signals", "cluster these findings", "are we whack-a-moling", "should we keep patching".

Enables remote control over every strain of a brood by fanning the /rc slash command out to each strain's live session. Fires ONLY on an EXPLICIT user request to turn on remote control for a brood — e.g. "enable RC for all strains", "RC for all strains", "turn on remote control for the brood / all strains", "enable remote control for the brood". This is a user-initiated action, NOT an autonomous coordination step the coordinator takes on its own.

Watch an open PR and remediate review feedback in a loop until a terminal condition. Arms a main-session Monitor on a thin change-detect poll and dispatches hivemind:github-reviewer in fix mode per actionable event. Executed by the overlord only — loop must run in the main session where Monitor survives subagent dispatches (ADR-0005).

Analyzes a codebase read-only and emits a ranked refactoring blueprint of deepening opportunities — turning shallow modules (small behaviour behind a large, leaky interface) into deep ones for testability and AI-navigability. Use when the user wants to improve architecture, find refactoring opportunities, consolidate tightly-coupled modules, make the codebase more testable, or make it more AI-navigable: this finds and ranks specific refactors but never edits code, writes docs, or spawns agents.

Initialize the run ledger for the current overlord instance — creates .hivemind/runs/ / and writes the initial state.json. Trigger: "init run ledger", "initialize ledger", "create run ledger", "start a run".

Scan one or more external-content items for prompt injection attempts and return a structured verdict. Use ad hoc whenever external content must be checked before classification or remediation. Trigger phrases: "scan for injection", "injection scan", "check for injection", "is this injection", "prompt injection check".

Record an intent-fallback event into a version-skewed run ledger and optionally close out a stale run — the sanctioned ledger write-path that deliberately bypasses transition-validation and the version-binding guard record-state-result.sh hard-rejects. Invoked by the overlord resume-on-start gate: the "proceed intent-driven" door for a skewed run, and the "start fresh" stale-run closeout. Trigger: "mark intent fallback", "intent-driven resume", "close out a stale run".

Create a checkpoint commit (molt) for the current approved plan after a completed phase, milestone, version bump, or review remediation item.

Computes the ready wave of independent plan steps for the intra-run parallel-wave implement loop, reading the run ledger and emitting the step ids to dispatch this iteration. Use when computing the next wave or ready-set of plan steps.

Open a pull request for a successfully completed approved plan after final verification.

Intensely interviews the user about their plan — one question at a time — challenging it against the project's existing domain model, sharpening terminology, and updating/creating documentation (CONTEXT.md, ADRs) as decisions crystallise. Use when the user wants to stress-test a plan, challenge a plan, interrogate a plan, red-team a plan, or harden a plan against their project's language and documented decisions. Trigger on: "stress-test a plan", "challenge plan", "grill me", "interrogate plan", "red team plan", "harden plan", "/hivemind:plan-interrogation".

Turns an already-interrogated plan into a committed PRD — a WHAT-only specification of what an Initiative delivers — written to docs/prds/ .md. Synthesizes from the plan with near-zero re-interrogation; runs only two light confirm gates (story completeness, scope boundaries). Use when the user wants to turn an interrogated plan into a PRD, write a PRD from a plan, produce a PRD, or draft docs/prds/ .md. Trigger on: "turn an interrogated plan into a PRD", "write a PRD from a plan", "plan to PRD", "produce a PRD", "/hivemind:plan-to-prd".

Push the current working branch to its remote. No PR creation — push only, for resuming a review loop after remediation commits.

Record the outcome of the current workflow state into the run ledger and advance state.current to the legal next state. Validates the transition against the workflow definition. Trigger: "record state result", "advance workflow state", "record transition", "advance the run".

Perform a behavior-preserving structural refactor that deepens a module — turning a small amount of behaviour behind a large, leaky interface into a large amount of behaviour behind a small one. Use when the user wants to deepen a module, refactor to a deeper interface, do a behavior-preserving structural refactor, extract behind a port/adapter, merge shallow modules, or execute a ranked refactoring blueprint candidate produce by prior architectural analysis. Requires Write and Edit access — invoke only in a context that can modify source and test files.

Classify a non-Reflex request and select which workflow handles it — the sole workflow classifier. Selects by judgment, never a keyword table; emits selected | ambiguous | exploratory | blocked. Trigger: "route workflow", "which workflow", "classify request", "select workflow".

One-time project setup. Apply the required .claude/settings.json keys (enabledPlugins + default agent) so the overlord becomes the session default agent. Use only when adopting the plugin in a new project, when repairing settings, or when the user explicitly requests setup. Also ensures .hivemind/ and .claude/worktrees/ are excluded from git via .gitignore.

Dispatch parallel overlord sessions as a brood. Spawns N Claude Code instances in separate git worktrees via tmux, injects task descriptions, and writes the brood manifest. Trigger: "spawn brood", "dispatch brood", "deploy brood", "parallel sessions", "dispatch fleet".