adept

Discover and record a repository's instructions, default branch, working- tree state, guardrail commands, GitHub authentication, decision-record coupling, and parallel-work context before changing code. Use when starting implementation, issue, pull-request, shipping, or campaign work.

Mine GitHub workflow telemetry for cycle time, review iterations, scope accuracy, and status-label history, then write a grounded process- retrospective report with proposed tuning. Use for retrospective analysis of status labels and WORK annotations over a label set or date range.

Verify, deduplicate, draft, and file a structured GitHub issue grounded in actual code, with triage labels and native sub-issue support. Use when asked to create or file an issue, turn a reported problem into an issue, or decompose an existing issue or epic into child issues.

Orchestrate a set of GitHub issues until each is closed as already fixed or resolved by a merged pull request, including triage, dependency planning, per-issue execution, serial merging, and newly discovered work. Use when asked to run a campaign, clear an issue set, or drive a batch of issues to done.

Sweep local branches whose work has landed: classify every local branch, delete the merged and squash-merged ones after one confirmation, remove their worktrees, and report every deletion and skip. Use when asked to clean merged or gone branches or prune stale local branch state.

Reverse a bad merged PR: judge the live damage, choose git revert vs fix-forward using the risk taxonomy's reversal-cost axis, write truthful tracking state, and drive the urgent fix through a reduced pipeline whose review is compressed but never skipped. Use when a merged PR turned out bad.

Push a feature branch, create or update a pull request, and drive it to green CI and a mergeable GitHub state. Use when asked to ship, publish, or prepare completed work for merge, including as the shipping phase of an issue flow.

Find the cause of a failure before changing anything. Use on a failing test, a reported bug, a broken build, or behaviour you cannot account for — and before proposing a fix for any of them.

Run a diff-scoped security pass by inventorying touched trust boundaries and checking validation, authorization, bounding, destination encoding, leaks, secrets, cryptography, deserialization, supply chain, permissions, and security defaults. Use when a branch is security-relevant or the user asks for a threat scan.

Apply a behavior-preserving simplification pass over recent changes: remove speculative abstraction, dead code, cleverness, redundancy, and noise, then verify with repository guardrails. Use when asked to simplify a diff, branch, PR, or recent work.

Estimate a GitHub issue's blast radius, change hazards, complexity, and decomposition before implementation. Use when asked to scope an issue, assess whether it fits one pull request, or provide grounded sizing to issue and work workflows.

Implement an approved plan with test-driven development, direct or worker-driven execution, focused verification, and the repository guardrail suite. Use when asked to build with TDD, execute an implementation plan, or continue the build phase of an issue workflow.

Use when asked to challenge, pressure-test, hostile-review, or adversarially review code changes, git diffs, specs, RFCs, plans, design docs, or files before shipping; also use for hidden assumptions, failure modes, edge cases, rollout risk, pre-ship skepticism, or structured JSON findings written to an output path.

Capture a non-obvious, verified problem solved in the current session as a durable solution document and, when available, a short memory-system recall pointer. Use when asked to compound learning, record a solution, or preserve a recurring debugging discovery.

Audit a complete reviewed design against its frozen scope before implementation, producing a human-readable verdict and approved surface without authorizing scope.

Implement a GitHub issue end to end through scoping, feature-branch setup, design, TDD, adversarial review, threat scan, simplification, pull-request creation, CI, and merge handoff. Use when asked to work, implement, or resolve a specific GitHub issue with the repository's full workflow.

Use when tracking GitHub issue/PR work across sessions — defines the status: label state machine, the WORK:* annotation-comment convention, and cross-issue linking so pipeline state lives on GitHub and any session can resume it. Referenced by $quest, $campaign, $deliver, $return-to-town, $sort-board, $bards-tale, $bounty, $saga, $resurrection, $warding.

Audit Dependabot configuration, evaluate open dependency-update pull requests against a clean baseline, group overlapping updates, test each unit, and merge only verified safe updates serially. Use when asked to review, batch, or merge Dependabot PRs for a GitHub repository.

Reconcile GitHub status labels with actual pull-request and branch state, reset stale orphaned work, and remove residual labels from closed issues. Use between work or campaign runs when issue workflow state may be stale.

Hand off a green, mergeable pull request or merge it when explicitly authorized, then clean merged branches and worktrees. Use after shipping a PR or when asked to merge and clean up completed work.

Interview a rough feature idea into a GitHub epic containing a PRD and dependency-ordered, born-triaged native sub-issues. Use when a feature spans multiple pull requests and the user wants an epic, PRD, decomposition, or campaign-ready issue set.

Rank the status:ready GitHub issue queue and recommend the next issue to work, excluding untriaged, occupied, or blocked candidates. Use when asked to pick the next issue, choose what to work on next, or select from the ready backlog before running $quest.

Analyze GitHub issues and apply a consistent type, priority, status, risk, and effort taxonomy, including bootstrapping missing labels. Use when asked to triage issues, label a backlog, prepare issues for work selection, or create the label set required by issue workflows.

Design a non-trivial code change before implementation by writing a spec or ADR, adversarially reviewing the ADR and spec, producing an implementation plan, and reviewing the plan. Use for issues or changes involving public contracts, schemas, auth, concurrency, migrations, persistence, dependencies, AI surfaces, security boundaries, or external services.

Offload high-volume, well-specified generation to parallel OpenAI Codex CLI workers. Use when a job splits into independent worker-sized tasks (content batches, test generation, bulk refactors) and Codex CLI is installed and logged in — The current Codex session decomposes, briefs, spawns worker `codex exec` runs in parallel, and review-gates every output. Not for judgment-dense work this session should do itself, or for coordinating workers on one complex change.

Use when a repo needs immutable numbered records — Architecture Decision Records or deferred-work records — that stay auditable and hard to erase. Defines the docs/adr/ and docs/debt/ record formats and installs the CI gate that fails a PR carrying a malformed record or one that stopped being a record. Referenced by $trial-loop, which writes deferral records for its deferred-tracked disposition.

Iteratively run an adversarial challenge review, fix or disposition defensible findings, write deferral records when needed, and re-review until approved or bounded stop conditions fire. Use for code, diffs, specs, ADRs, plans, and pre-ship review loops.

Run a proactive repository maintenance sweep for stale issues, dependency freshness outside Dependabot, supply-chain advisories, and deferral records past review-by. Use when asked to groom, sweep, or surface maintenance work without doing the work.