agent-loop-auto
Autonomous variant of /agent-loop for ONE repo: thin coordinator + parallel implementer subagents + a coordinator-dispatched INDEPENDENT reviewer, plus an…
Skill by PlatformRelay
Chalk Labs
Need help deploying agent-loop-auto?
Chalk Labs sets up agents like this inside real teams: connected to your tools, tested on your data, with a human in the loop where it matters.
Install
Works with Claude Code, Claude apps, Codex, Cursor, Gemini CLI, GitHub Copilot, VS Code, OpenCode, Goose, Windsurf. Read the repository before you give any tool access to your data.
npx agentmarket install skill/platformrelay-agent-loop-auto
Asks which of your tools to add it to (Claude Code, Codex, Cursor, VS Code, Windsurf, Gemini CLI, Claude Desktop) and shows the change first. See https://agentmart.dev/cli.
Get the weekly diff
The weekly diff: one email a week with new agent tools, notable updates and grade changes, plus security alerts for tools you follow. Double opt-in, one-click unsubscribe.
About
Use cases: devops, agent-building
Reviews
Write a reviewNo reviews yet. Installed it? Reviews come from verified installers only. How reviews work
Permissions
What this skill can do and what it asks for, from its SKILL.md. Checked Oct 25, 2026.
- Tools
- No tools found
- Secrets
- None asked for
- Install scripts
- None found
Network hosts it talks to 1
codeload.github.com· found in install command
Files it touches
~/.claude/skills/agent-loop-auto
What's inside
- Skill name
agent-loop-auto- Description
- Autonomous variant of /agent-loop for ONE repo: thin coordinator + parallel implementer subagents + a coordinator-dispatched INDEPENDENT reviewer, plus an Integrator subagent that auto-merges each PR (rebase) without operator input once gates + independent APPROVE + CI are green — then loops to the next lanes until the backlog is exhausted or a stop condition hits. Invoking this skill is the operator's explicit authorization to merge autonomously. Use for hands-off progress on your OWN backlog; never for external-contributor PRs.
- License
- MIT
- Instructions
- 80 lines in SKILL.md · view on GitHub
Files (1)
- SKILL.md
SKILL.md (first lines)
# agent-loop-auto — autonomous parallel loop (auto-merges)
Everything `/agent-loop` does (thin coordinator + **parallel** implementers + **coordinator-dispatched
independent review**, per-repo, TDD → gates → review), **plus**: it **auto-merges** each lane's PR when
it passes, then keeps picking new lanes until done. Read `/agent-loop` for the implementer procedure,
the independent-review step (3b), and the role schemas — this file only states what's different: the
auto-merge gate and a dedicated **Integrator subagent** that keeps merge work off the coordinator.
> **Autonomy is operator-authorized by invoking this skill.** It overrides the base loop's
> "operator merges" step — but only within the guardrails below. This is *your own* backlog being
> advanced hands-off; it is **not** a path to merge other people's code.
## When to use
- You want the harness to work through a repo's backlog and land the results without babysitting.
- **Not** for external-contributor PRs — those always need operator-visible review (`/merge-open-prs`).
## Auto-merge gate (ALL required, per lane)
A lane's PR auto-merges only when **all** hold:
1. Full gate matrix green (repo config) — `task verify/lint/test/coverage/scrub` or `pnpm build/export`.
2. The **coordinator-dispatched independent reviewer** (base Step 3b — a fresh subagent that saw only
the diff, never the implementer's context) returned **APPROVE** with **no P0/P1** findings.
3. CI is **green on the PR's head commit** (`gh pr checks`), and the PR is **MERGEABLE** (no conflicts).
4. The lane's author is the harness/operator (never an external contributor's PR).
Then the coordinator hands the lane to an **Integrator subagent** (below) — it does not merge inline.
> **Permission-layer veto (learned 2026-07-16, mkurator):** in auto-permission/background environments
> the classifier may categorically deny `gh pr merge` to the default branch (public-deploy-triggering
> merges especially) — even when this gate is fully satisfied. Do not route around a denial. Fallback
> per gate-clean lane: record the PR in the repo INBOX 🟢 with the exact merge command, and ask the
> operator directly if they are live. True auto-merge requires the operator to add a `gh pr merge`
> allow rule to the repo's `.claude/settings.local.json` (mkurator has one since 2026-07-16).
## Integrator subagent (keeps merge work off the coordinator)
For each gate-clean lane the coordinator dispatches **one Integrator subagent** — the coordinator does
not rebase/merge inline. The integrator:
1. **Rebase onto current `main` first — isolation worktrees can come up on a STALE base.** An
`isolation:"worktree"` worktree may be created at an earlier session HEAD, not the live `main` (seen
this run: a worktree 7 commits behind). `git rebase main` the lane head; **never ff-merge an
unrebased head**. (Tell implementers to `git merge --ff-only main` at lane start too.)
2. **Re-run the full gate matrix on the rebased head** (file-disjoint lanes rebase clean — verify anyway).
3. `gh pr merge <n> --rebase --delete-branch`. **If the classifier denies the merge (veto note above),
do not route around it** — return `{merged:false, error:"classifier-veto"}` so the coordinator runs
the INBOX-record fallback.
4. **Return** `{ lane, merged, mainSha, error? }`. On success the coordinator moves the board row to
**Integrated**, clears any matching INBOX task, and updates `SESSION-HANDOFF`; on `error` it records
and skips. The coordinator routes on this schema — it does not read the merge output itself.
Integrators run **serialized** (one merge to `main` at a time); implementers stay parallel.
## Stop conditions (halt the loop and hand back)
Stop and surface via the repo's INBOX (a **DECISION** with options) when any occur:
- Backlog empty, or no remaining **non-overlapping** lanes to parallelize.
- A lane hits **BLOCK / REQUEST CHANGES**, red or pending CI, or a merge conflict — leave that PR open,
don't merge, record why.
- A change touches **security, public API/CRD, or release** — surface for a human look before merging
even if gates pass (don't silently auto-merge high-blast-radius changes).
- `main` CI goes red after a merge — stop starting lanes, diagnose.
- A configured ceiling is reached: max lanes, max iterations, or the turn's token budget.
## Flow
1. **Steps 1–3b of `/agent-loop`** — orient, pick N non-overlapping lanes, dispatch parallel
implementers, then dispatch the **independent reviewer** per lane; collect the structured returns.
2. For every lane meeting the auto-merge gate, **dispatch an Integrator subagent** (above); skip +
record the rest. The coordinator routes on each integrator's `{merged, error?}` return — it does not
rebase or merge inline.
3. **Loop:** pick the next batch of independent lanes and repeat, until a stop condition.
4. **Report + `/retrospective`:** what merged, what was skipped and why; log notable decisions to
`decisions.md` (including any dissent). Leave the repo green.
## Do not
- Auto-merge on anything less than **gates green + independent APPROVE + green CI + mergeable**.
- Merge inline in the coordinator, or run integrators in parallel — merges to `main` are serialized.
- Auto-merge an **external contributor's** PR, or a security/API/release change without surfacing it.
- Run past the stop conditions / ceilings, or force-merge / use `--admin` / bypass branch protection.
- Parallelize machine-locked suites (e2e/integration), or coordinate across repos.
- Merge partial or self-unreviewed work — every merge still passes the same independent review gate as
`/agent-loop` (a fresh reviewer that never saw the author's context).From PlatformRelay/Kaddy, MIT.
More from PlatformRelay/Kaddy
- 0.497score
write-story
PlatformRelay
Turn an epic, feature idea, or vague ask into INVEST-compliant vertical-slice stories with Given-When-Then acceptance criteria, mapped onto the target repo's…
SkillCFreshness: C★ 1Claude Code +9 - 0.497score
agent-loop-local
PlatformRelay
Maximum-autonomy variant of /agent-loop-auto for ONE repo: no PRs — each lane is gated by a coordinator-dispatched INDEPENDENT reviewer, then a serialized…
SkillCFreshness: C★ 1Claude Code +9 - 0.497score
brainstorm
PlatformRelay
Structured diverge → cluster → converge ideation that prevents premature fixation on the first idea.
SkillCFreshness: C★ 1Claude Code +9 - 0.497score
changelog
PlatformRelay
Turn gitmoji-conventional commit history into categorized, user-facing release notes for a PlatformRelay repo.
SkillCFreshness: C★ 1Claude Code +9 - 0.497score
design-architecture
PlatformRelay
Design a change and record the decision: ADR/AgDR, C4 sketch, weighted trade-off matrix, or a spec-first contract.
SkillCFreshness: C★ 1Claude Code +9 - 0.497score
roast
PlatformRelay
Roast me — adversarial critique of a plan, design, or diff to find where it breaks, via pre-mortem, red-team, and steelman-then-destroy.
SkillCFreshness: C★ 1Claude Code +9 - 0.497score
security-review
PlatformRelay
Dedicated data-flow security pass over a diff or a repo — deeper than the security dimension of /tech-review.
SkillCFreshness: C★ 1Claude Code +9 - 0.497score
tech-review
PlatformRelay
Review a lane/branch before opening (or while awaiting) a PR in any PlatformRelay repo.
SkillCFreshness: C★ 1Claude Code +9 - 0.497score
agent-loop
PlatformRelay
Start or resume implementation in one PlatformRelay repo.
SkillCFreshness: C★ 1Claude Code +9 - 0.497score
handover
PlatformRelay
Wrap up a session cleanly and produce a copy-pasteable prompt to start the NEXT one.
SkillCFreshness: C★ 1Claude Code +9 - 0.497score
merge-open-prs
PlatformRelay
Review and merge all open PRs/MRs in ONE repo after checking CI.
SkillCFreshness: C★ 1Claude Code +9 - 0.497score
operator-inbox
PlatformRelay
Scan each PlatformRelay repo's INBOX independently and give the operator a per-repo digest of what needs them — decisions (with options), operator tasks…
SkillCFreshness: C★ 1Claude Code +9
Signals
- Stars
- 1
- Forks
- 0
- Open issues
- 0
- Last commit (this folder)
- Jul 20, 2026 (3 months ago)
- Last release
- Jul 21, 2026 (3 months ago)
From GitHub (fetched Oct 24, 2026).
Details
- License
- MIT
- Repository
- github.com/PlatformRelay/Kaddy
- First seen
- Oct 24, 2026
- Last checked
- Oct 24, 2026
- Source
- GitHubgithub-graphql
More from PlatformRelay
- 0.497score
agent-loop-local
PlatformRelay
Maximum-autonomy variant of /agent-loop-auto for ONE repo: no PRs — each lane is gated by a coordinator-dispatched INDEPENDENT reviewer, then a serialized…
SkillCFreshness: C★ 1Claude Code +9 - 0.497score
agent-loop
PlatformRelay
Start or resume implementation in one PlatformRelay repo.
SkillCFreshness: C★ 1Claude Code +9 - 0.497score
changelog
PlatformRelay
Turn gitmoji-conventional commit history into categorized, user-facing release notes for a PlatformRelay repo.
SkillCFreshness: C★ 1Claude Code +9 - 0.497score
design-architecture
PlatformRelay
Design a change and record the decision: ADR/AgDR, C4 sketch, weighted trade-off matrix, or a spec-first contract.
SkillCFreshness: C★ 1Claude Code +9 - 0.497score
brainstorm
PlatformRelay
Structured diverge → cluster → converge ideation that prevents premature fixation on the first idea.
SkillCFreshness: C★ 1Claude Code +9 - 0.497score
handover
PlatformRelay
Wrap up a session cleanly and produce a copy-pasteable prompt to start the NEXT one.
SkillCFreshness: C★ 1Claude Code +9
Done for you
Rather not set it up yourself?
Chalk Labs scopes, builds and runs agents for your team: the task audit, the eval harness, the integrations and the weekly ops report. Fixed scope, priced before we start.
See how we build AI agents →Read the case studies
Chalk Labs is a two-founder AI and growth studio. We build AI agents, CRMs and MVPs for Web3 and AI companies, and have shipped 50+ product and token launches.
- launches shipped
- 50+
- launches shipped
- media placements
- 500+
- media placements
- founders on every account
- 2
- founders on every account
Press we've placed: Forbes · CoinDesk · Decrypt · Cointelegraph · The Block
Rahil Jain, Co-founder, Chalk Labs. Ex-Goldman Sachs engineer. You talk to the person who builds it.
Get agent-loop-auto running in your business
Tell us what you want to automate. A founder replies within one working day with a plan, not a pitch deck.
Cite this page
agent-loop-auto skill. Agent Mart. https://agentmart.dev/skills/platformrelay-agent-loop-auto (last updated 2026-10-24)