name: ralph-loop
description: Autonomously fix open GitHub issues one at a time using fresh-context subagents. Loops until no eligible issues remain, a fix-issue invocation is halted by the complexity gate or an error, or a HITL checkpoint is encountered. AFK-safe: no human input required during the loop. Use when asked to "run the ralph loop", work through all open issues, or fix issues autonomously.
user-invocable: false
Ralph Loop
Configuration
Model Override: Set to true to use default models (conserves token limits). Set to false to use specified models for optimal performance.
USE_DEFAULT_MODELS = false
Orchestrate repeated fix-issue runs — one issue per fresh-context subagent — until a stopping condition is met.
AFK contract: this skill requires no human input during execution. The fix-issue Phase A step 4 HITL confirmation is pre-approved by the user invoking this skill. If any other HITL checkpoint is reached inside fix-issue, the loop halts and reports to the user.
Fresh context contract: each issue is handled by an isolated subagent. The orchestrating agent (you, running this skill) never reads source files, writes code, or runs tests. It only manages the queue and delegates. This prevents context pollution across a long session.
Phase 0 — Build the candidate pool
Fetch all open AFK candidates once, at the start of the session:
gh issue list --label ready-for-agent --state open \
--json number,title,body,labels \
--jq '[.[] | {number, title, body, labels: [.labels[].name]}]'
If gh is not available: use the github-pull-request_doSearch tool with query is:issue is:open label:ready-for-agent repo:owner/name. Always include the repo: qualifier — the tool searches all repos by default. See docs/agents/issue-tracker.md for the full fallback table.
All issues returned by this query are AFK by definition — ready-for-agent means AFK (see docs/agents/triage-labels.md). No further type filtering is needed.
Sort candidates by issue number, ascending. This is the candidate pool.
Do not check blockers yet — that happens dynamically inside the loop, so newly unblocked issues are discovered after each fix.
Initialise two session-level sets:
resolved_this_session: issue numbers for which a PR was created in this loop run (initially empty)attempted: issue numbers already processed in this run (initially empty)
If the candidate pool is empty, output:
Ralph Loop: no eligible AFK issues found. Nothing to do.
and stop.
Phase 1 — Loop
Repeat until a stopping condition is met:
Step 0 — Select next eligible issue
From the candidate pool, exclude all issues in attempted. For each remaining candidate (in ascending number order), check whether it is unblocked:
Parse the ## Blocked by section. For every blocker issue number listed (e.g. #12):
- First check: is
12 in resolved_this_session? If yes, treat it as resolved without a network call.
- Otherwise run:
gh issue view 12 --json state --jq '.state'
If gh is not available: use github-pull-request_issue_fetch with issueNumber: 12; check the .state field of the returned object.
The candidate is eligible if every blocker is in resolved_this_session OR returns CLOSED (or the section reads "None — can start immediately").
Pick the lowest-numbered eligible candidate.
- If no candidates remain in the pool (excluding
attempted): proceed to Phase 2 with halt reason QUEUE_EXHAUSTED.
- If candidates remain but none are eligible (all are still blocked): proceed to Phase 2 with halt reason
ALL_BLOCKED.
Step 1 — Spawn a fresh-context subagent
Invoke a subagent using the runSubagent tool. If USE_DEFAULT_MODELS is false, specify model: "Claude Sonnet 4.5 (copilot)"; if true, omit the model parameter to use the default. Pass it the following prompt, substituting {number} and {title}:
Read the fix-issue skill at .github/skills/fix-issue/SKILL.md in full, then execute it for issue #{number} ("{title}").
Skip Phase A entirely. The Ralph loop orchestrator has already selected this issue, confirmed it is AFK type, and verified it has no open blockers. Begin at Phase B (branch creation).
Complete all remaining phases (B through E) and return a structured result with these fields:
- outcome: one of
PR_CREATED, COMPLEXITY_GATE, ERROR, HITL_REQUIRED
- pr_url: the pull request URL (if outcome is PR_CREATED), otherwise null
- detail: a one-paragraph plain-language summary of what happened
Step 2 — Evaluate the result
Add {number} to attempted regardless of outcome.
| Subagent outcome | Loop action |
|---|
PR_CREATED | Add {number} to resolved_this_session; log the PR URL; return to Step 0 |
COMPLEXITY_GATE | Proceed to Phase 2 with halt reason COMPLEXITY_GATE |
ERROR | Proceed to Phase 2 with halt reason ERROR |
HITL_REQUIRED | Proceed to Phase 2 with halt reason HITL_REQUIRED |
Phase 2 — End-of-loop report
After the loop ends for any reason, output this summary to the user:
Ralph Loop Summary
==================
Issues attempted : {N}
PRs created : {list of PR URLs, one per line, or "none"}
Stopped because : {reason — see table below}
| Halt reason | Message |
|---|
QUEUE_EXHAUSTED | "All eligible issues processed — nothing more to do." |
ALL_BLOCKED | "Remaining candidates are still blocked. PRs may need to be merged before they can proceed." |
COMPLEXITY_GATE | "Issue #{number} exceeded the complexity gate. Run the to-issues skill to break it into smaller child issues." |
ERROR | "An unrecoverable error occurred on issue #{number}. Detail: {subagent detail field}" |
HITL_REQUIRED | "Issue #{number} requires human input. Detail: {subagent detail field}" |
Do not attempt to retry a failed iteration. Proceed to Phase 2 and stop.
