forge-loop

name	forge-loop
description	Run the full forge pipeline end-to-end within this Claude Code session, processing multiple issues autonomously until a stop condition is met
disable-model-invocation	true

Loop — Long-Horizon Autonomous Development

Run the complete forge pipeline entirely within this Claude Code session. Pick issues, implement them, process review feedback, wait for CI, and merge — then repeat. Continue until a stop condition is reached or you are interrupted.

This is the agent-native alternative to the GitHub Actions pipeline. Each major step runs as a sub-agent with focused context. The loop itself is a thin orchestrator that manages state and sequences the steps.

Input

$ARGUMENTS — flags controlling the run.

/forge-loop                               # Run until no more backlog/ready issues
/forge-loop --max-issues 5                # Stop after 5 issues
/forge-loop --max-hours 8                 # Stop after 8 hours
/forge-loop --issues "#42,#43,#44"        # Process specific issues in this order
/forge-loop --sweep                       # Run /sweep-issues before starting
/forge-loop --no-merge                    # Stop before merge (human merges)
/forge-loop --dry-run                     # Plan only: show what would run, no changes

Flags can be combined: /forge-loop --max-hours 4 --sweep --no-merge

---

Step 0 — Read Configuration

Read forge.toml:

[project]
repo = "org/project"
base_branch = "main"

[stack]
build_command = "..."
test_command = "..."

[gating]
human_merge_required = true
max_copilot_iterations = 3

[loop]
max_failures_before_stop = 3
ci_poll_interval_minutes = 5
ci_timeout_minutes = 60

Read CLAUDE.md for project conventions. This applies to all implementation work across all issues in the run.

---

Step 1 — Initialize Loop State

Create or resume .forge/loop-state.json at the repo root:

{
  "session_start": "<ISO timestamp>",
  "flags": {
    "max_issues": null,
    "max_hours": null,
    "issue_list": [],
    "sweep": false,
    "no_merge": false,
    "dry_run": false
  },
  "issues_completed": 0,
  "consecutive_failures": 0,
  "current": null,
  "completed": [],
  "failed": []
}

If .forge/loop-state.json already exists with a non-null current entry, a previous run was interrupted mid-issue. Ask the user:

Found in-progress issue from previous run:
  Issue #$N: $TITLE (phase: $PHASE, PR: #$PR)

Resume from where it left off? [Y/n]

If yes: skip to the appropriate step for that phase (Step 3, 4, 5, or 6). If no: clean up the stale state (remove agent/implementing label, delete worktree) and start fresh.

---

Step 2 — Optional Backlog Sweep

If --sweep flag is set, run /sweep-issues before the main loop. This ensures the backlog has quality-checked, ready issues before implementation begins.

In dry-run mode, report what /sweep-issues would do but don't execute it.

---

Step 3 — Pick Next Issue

Determine the next issue to work on:

If --issues list was provided: take the next uncompleted issue from the list.

Otherwise: use /dispatch to pick the highest-priority backlog/ready issue.

If no issues are available (list exhausted or no backlog/ready issues), the loop ends cleanly:

Loop complete — no more issues to process.
  Issues completed this session: $N
  Total time: $ELAPSED

Save final state and exit.

In dry-run mode, show which issue would be picked and stop.

---

Step 4 — Implement

Update loop state:

"current": { "issue": $N, "title": "...", "phase": "implement", "started_at": "..." }

Run /autodev $ISSUE_NUMBER as a sub-agent to implement the issue, create the worktree, run tests, and open the PR. Pass the branch name from /dispatch if already created.

On success: update state with the PR number:

"current": { "issue": $N, "pr": $PR_NUMBER, "phase": "review" }

On failure (blocked, tests failing after retries):

"failed": [..., { "issue": $N, "phase": "implement", "reason": "...", "timestamp": "..." }]
"consecutive_failures": $N + 1

Check circuit breaker (Step 8). Then proceed to next issue.

---

Step 5 — Review Loop

Two-phase review: a sub-agent posts review comments on the PR, then the main thread addresses them. This replaces waiting for Copilot (which takes 40–60 minutes and consistently arrives after merge in local runs).

5a — Sub-agent review

Spawn a fresh-context sub-agent (general-purpose) to perform and post a code review directly on the PR. The sub-agent should:

1. Read the PR diff: gh pr diff $PR_NUMBER --repo $REPO
2. Read the full content of every changed file (not just the diff hunks)
3. Review for:
   • Correctness and logic errors
   • Security (XSS, injection, path traversal, DoS vectors)
   • Error handling gaps
   • Test quality and coverage
   • Adherence to project conventions from CLAUDE.md
4. Post findings as inline PR review comments using the GitHub API:

   gh api "repos/$REPO/pulls/$PR_NUMBER/reviews" \
     -X POST \
     -f event="COMMENT" \
     -f body="Review by forge-loop sub-agent" \
     --jq .id \
     -f 'comments[][path]=<file>' \
     -f 'comments[][position]=<diff position>' \
     -f 'comments[][body]=<finding>'
   

   If there are no findings, post a single approving review comment noting the PR looks clean.
5. Return a summary: count of findings by severity, and whether any are blocking.

The sub-agent must NOT modify any files — it only reads and posts comments.

5b — Address review findings

The main thread (orchestrator) processes the sub-agent's findings:

1. Run /review-pr $PR_NUMBER to read the posted comments, fix addressable issues, reply in-thread, and create follow-up issues for deferred items.
2. Push fixes if any were made.

5c — Update state

Update state:

"current": { ..., "phase": "await-ci" }

On failure: if the review pass fails after 2 attempts, add human/blocked, log failure, check circuit breaker, continue to next issue.

---

Step 6 — Await CI

Run /await-ci $PR_NUMBER to wait for all CI checks to pass.

While waiting, report progress:

Waiting for CI on PR #$PR_NUMBER... ($ELAPSED_MINUTES min elapsed)
  passing: $N  pending: $N  failing: $N

On CI pass: proceed to Step 7.

On CI failure: inspect the failure output. If the failure is clearly fixable (test assertion mismatch, lint error, type error): attempt a fix and push. Re-run /await-ci. Maximum 2 CI fix attempts.

If fix attempts are exhausted or the failure requires deeper investigation:

gh pr edit $PR_NUMBER --repo $REPO --add-label "human/blocked"

Log the failure, check circuit breaker, continue to next issue.

Update state:

"current": { ..., "phase": "merge" }

---

Step 7 — Merge

Run /merge-pr $PR_NUMBER.

If human_merge_required = true or --no-merge flag:

PR #$PR_NUMBER is ready to merge.
  Labeled: human/review-merge
  URL: $PR_URL

Record as completed (merge is human's responsibility):

"completed": [..., { "issue": $N, "pr": $PR_NUMBER, "merged": false, "ready_at": "..." }]

If auto-merge is enabled:

After successful merge, record:

"completed": [..., { "issue": $N, "pr": $PR_NUMBER, "merged": true, "merged_at": "..." }]

Clean up the worktree:

git worktree remove ".worktrees/issue-$N-$SLUG" 2>/dev/null || true

Print a progress line:

✓ Issue #$N merged (PR #$PR_NUMBER) — $ELAPSED_MINUTES min total
  Session progress: $COMPLETED / $TOTAL issues

Update state:

"current": null,
"issues_completed": $N + 1,
"consecutive_failures": 0

---

Step 8 — Circuit Breaker

After every failure, check stop conditions:

consecutive_failures >= max_failures_before_stop

If triggered:

Circuit breaker: $N consecutive failures. Stopping the loop.

Failed issues:
  - #$N1: $REASON
  - #$N2: $REASON
  - #$N3: $REASON

These have been labeled human/blocked. The loop state is saved at .forge/loop-state.json.
Resume with /forge-loop when the underlying issues are resolved.

Save state and exit.

---

Step 9 — Check Global Stop Conditions

After each completed (or failed) issue, check:

Condition	Check
--max-issues reached	issues_completed >= max_issues
--max-hours reached	now - session_start >= max_hours * 3600
--issues list exhausted	All listed issues processed
No more ready issues	/dispatch returns nothing

On any stop condition met:

Loop stopping: $REASON.

Session summary:
  Duration:   $ELAPSED
  Completed:  $COMPLETED issues
  Merged:     $MERGED PRs
  Blocked:    $BLOCKED issues (labeled human/blocked)
  Next ready: $NEXT_ISSUE (if any)

---

Step 10 — Loop

If no stop condition is met, return to Step 3.

---

Execution model — sub-agents

Each major step (implement, review, await-ci, merge) runs as a sub-agent to keep the loop orchestrator's context small across long runs. The loop orchestrator:

• Reads state from .forge/loop-state.json
• Spawns the appropriate sub-agent
• Reads the result
• Updates state
• Decides what to do next

This means the loop can run many issues without the orchestrator's context window growing unboundedly.

---

State file location

.forge/loop-state.json at repo root. This file is gitignored (add .forge/ to .gitignore if not already present). It persists across interrupted runs and is the checkpoint for resume.

---

Dry-run mode

With --dry-run:

• Read all state (backlog, open PRs, CI status)
• Report what the loop would do: which issues, in which order, estimated duration
• Do not create branches, open PRs, push code, or modify any labels
• Output a plan table:

Dry-run plan — forge loop

  Would process:
    1. #42: $TITLE (backlog/ready, priority/high)
    2. #43: $TITLE (backlog/ready)
    3. #38: $TITLE (via/autodev PR #95 — in review-loop, 1/3 iterations)

  Stop condition: --max-issues 3
  Estimated time: 2–4 hours (at current pipeline velocity)

  Run /forge-loop --max-issues 3 to execute.

---

Guardrails

• One loop at a time: Check for a running loop (stale .forge/loop-state.json with a recent timestamp) before starting. If one appears active, warn and ask to confirm.
• Never modify protected files: Check forge.toml [protected_files] patterns before every commit.
• Never push to base branch: All work on autodev/issue-* branches only.
• Always reply to review comments: Every review pass must reply in-thread.
• Respect CI: Never merge with failing checks, even if human_merge_required = false.
• Save state on interrupt: If the user stops the loop (Ctrl+C), attempt to write the current state to .forge/loop-state.json before exiting.