A collection of agent configurations across AI services
This repository also configures its own agents. See what agents-skills tells them →
Agent Claude Code
Fast-forward lane. Loops on an interval and keeps the PRIMARY checkout's integration branch at the remote tip with a safe git pull --ff-only — so every worktree an agent cuts starts from a current base (hooks, scripts, skills). Purely mechanical — run the tick, match the decision table, report one line. When the…
Agent Claude Code
Surgical 1–2 file edit — typo fixes, single-function changes, mechanical renames, format-preserving tweaks. Hard-refuses 3+ file scope, new files (unless asked), or new features. Returns a tight diff receipt. Use for bounded, obvious edits; not for cross-file refactors or anything needing design.
Agent Claude Code
CI-stability medic and the SOLE flake hunter for a repo. Continuously mines CI history for real flakes (shards that intermittently red the per-PR gate or the mergegroup speculative build and eject correct PRs), reproduces them harsher-than-CI, root-causes the real race, and fixes them FORWARD — never masks (no skip…
Agent Claude Code
CI-latency medic and the SOLE speed hunter for a repo. Continuously mines recent CI history (per-PR gate + mergegroup) to find where CI is SLOW — long-pole jobs, the wall-clock critical path, under-sharded matrices, cold caches, and tier gates that fail to skip work — then makes CI faster WITHOUT losing coverage.…
Agent Claude Code
Per-domain Figma design mapper for /figma-init-spec and /figma-refresh-spec — pulls a component set or screen flow via the Figma MCP (getvariabledefs, getcode, getimage) and writes/updates its design specs anchored to stable node IDs and component keys, with a repo-owned design→code map found by name-matching the…
Agent Claude Code
Read-only repro and root-cause worker for the /…-use-aws and /…-use-browser commands and recon — drives the browser via whichever browser MCP is connected (Chrome, Playwright, Cypress, or Chrome DevTools) in its own context and/or digs AWS logs, config, and CloudTrail with strictly read-only calls, then returns…
Agent Claude Code
Implements one groomed issue end-to-end — works in its own worktree/branch, writes tests first, keeps scope to the issue, runs the pre-push gate, opens a PR with auto-merge armed, and returns the completion contract. The dev agent /dispatch spawns into a dev slot.
Agent Claude Code
CI + conflict + merge-queue medic watcher lane. Loops on an interval over open agent PRs, fixes failing required checks and DIRTY/conflicted branches, and runs full merge-queue health every tick — head-green is not queue-green. Dequeues + drafts poison-pill PRs that drop an ALLGREEN group, and re-arms innocent PRs…
Agent Claude Code
Post-merge janitor watcher lane. Loops on an interval over CLOSED agent PRs only — deletes merged local branches, removes spent worktrees, closes the linked GitHub/Jira issue on merge, and releases the claim lock. A PR closed without merging frees the lock but leaves the issue open and re-dispatchable. Needed because…
Agent Claude Code
Review-thread-closer watcher lane. Loops on an interval over open agent PRs, drives bot + human review threads to resolved (fix or reasoned counter, resolve only after the fix is pushed), and owns the initial arm-after-bot-review of auto-merge. Re-arming after a merge-queue ejection is the pr-checks lane's, not this…
Agent Claude Code
Unsticks a stuck PR — reads mergeStateStatus, triages CI from per-job logs (transient vs real), resolves bot and human review threads, recovers DIRTY/conflicted branches, and re-arms auto-merge. The worker /rescue and dispatch rescue slots use. Diagnoses before it mutates.
Agent Claude Code
Adversarial diff/branch/PR reviewer. One line per finding, severity-tagged, no praise, no scope creep, format path:line: : . .. Use for 'review this PR/diff/file'. Skips style nits unless they change meaning.
Agent Claude Code
Read-only code locator. Given "where is X", "what calls Y", "list uses of Z", or "map this area", returns a tight file:line map in caveman shorthand and nothing else — no fixes, no opinions. Use to offload mechanical lookups off the main session onto the cheap tier (see MODEL-DEFAULTS.md).
Agent Claude Code
Per-domain codebase mapper for /spelunking-init-spec and /spelunking-refresh-spec — runs the spelunk loop (framework introspection + machine-name following + runtime observation) over one assigned domain and writes its numbered spec files with verified file:line / command evidence. Writes only under the specs dir…
Agent Claude Code
Verify-then-trust gate. Confirms a claimed result actually holds before it's trusted — the PR exists and is the linked one, required CI is green, acceptance criteria are met, and any cited file:line / recon / spec claim is real. Returns a pass/fail verdict with evidence. Read-only; an unverifiable claim is a FAIL.
Agent Claude Code
Slot-filler watcher lane — the continuous form of /dispatch's dev slots. Each tick it claims groomed, unblocked, unclaimed issues, spawns an implementer per issue, tracks each to an open PR, hands that PR to the watcher lanes, and refills the freed slot. Honors the label control plane (do-not-dispatch holds…
Agent Claude Code
Deploy-health medic watcher lane. Loops on an interval over the integration environment's deploy workflow (deploy step, migrations/bootstrap, post-deploy smoke) and keeps it green by fixing FORWARD — never reverts, never skips tests, never masks. Triages transient (concurrent-deploy state-lock race, CI-infra blip…
At most 3 mods per repository are shown here, and a mod shipped inside a plugin is left to that plugin's page — the rest are on their repository pages: