Issue Handler

Issue Handler — End-to-End Orchestrator

This is the high-level scenario skill for handling a single GitHub issue. It does not do the detailed work itself; it sequences the leaf skills into one iterative pipeline and reports the result. Each stage's mechanics live in its own skill — read and follow that skill when you reach its stage.

Every agent-produced diff is a proposal. This skill commits the staged fix onto a dedicated local branch agent/fix-issue-<N> (single bug) or agent/fix-issue-<N>-<seq>-<slug> (per batch sub-item) and records it in a fix_result*.json. It never pushes, tags, or opens a PR — the invoking workflow reads each fix_result*.json, exports base_sha..branch as a patch, and a human applies it. The leaves themselves only stage (they never commit).

Contents

• Pipeline overview
• Inputs
• Execution modes
• Stage 0: Re-run gate
• Stage 1: Triage
• Stage 2: Reproduce
• Stage 1u: Batch-issue fan-out (skip-list + heterogeneous)
• Stage 3: Root cause
• Stage 4: Implement
• Stage 5: Verify
• Stage 6: Report
• Iterative loop bounds
• Issue-body status contract

Pipeline overview

Single-bug path (default for issue_type=single-bug):

triage → reproduce → root-cause → implement → verify → report
                       ↑                          |
                       └─── loop up to 3 times ───┘

Batch path (issue_type=batch-bug, either batch_kind) — fan out the single-bug pipeline per sub-item, each on its own fix branch:

triage → [preflight: install nightly wheel once]
       → for each sub-item:
             reset + reproduce
               ├─ NOT_REPRODUCED → heterogeneous: ALREADY_FIXED · skip-list: STALE_SKIP (+follow-up)
               └─ REPRODUCED → branch agent/fix-issue-<N>-<seq>-<slug>
                               → root-cause → implement → verify
             (any failure marks the sub-item and continues the batch)
       → fan-out report

skip-list vs heterogeneous differ only in how a NOT_REPRODUCED sub-item is labeled; the loop is identical.

Inputs

• A GitHub issue on intel/torch-xpu-ops or pytorch/pytorch (URL, number, or raw body).
• pytorch_dir — path to a local pytorch checkout, resolved as described in fix-reproduce Prepare. If absent, this skill lets fix-reproduce / fix-root-cause clone it into $XPU_OPS_ROOT/agent_space_xpu/pytorch/.
• Mode (see below).

Execution modes

The pipeline runs in one of two modes — interactive (default) or pipeline — which changes how every stage reports results and whether it writes to the GitHub issue. Decide the mode at the start and pass it to every leaf. See references/execution-modes.md for the full contract.

• Interactive (default): ask the user when blocked; report conversationally; do not touch the issue body / labels / comments unless the user asks.
• Pipeline (explicit): no human to interrupt — advance the issue's agent:status marker, update stage labels, let leaf skills leave their <!-- agent:<name> --> comments, and stop when the pipeline settles on a terminal verdict.

Stage 0 — Re-run gate (first-run vs re-run)

Pipeline mode only; skip in interactive mode (a human is already driving, so just run the full pipeline). An agent:active issue is frequently re-triggered — a maintainer leaves feedback, or the bot is re-invoked with no new information. This gate decides, before spending a build, whether this is a fresh run, a human-feedback re-run (highest priority), or a bare re-run (skip work that already ran and is still valid).

Step 0.1: Detect prior agent activity

Find the most recent agent comment and its timestamp:

last_agent_ts=$(gh issue view "$N" --repo "$OWNER/$REPO" --json comments \
  --jq '[.comments[] | select(.body | test("<!-- agent:"))] | last | .createdAt // ""')

Empty last_agent_ts → first run. Skip the rest of Stage 0 and go to Stage 1 normally.

Step 0.2: Detect new human feedback

A comment is human feedback when it is authored by a non-bot account and created after last_agent_ts. (The bot account is the one that authored the <!-- agent:* --> comments; exclude it by login.)

new_human=$(gh issue view "$N" --repo "$OWNER/$REPO" --json comments \
  --jq --arg ts "$last_agent_ts" --arg bot "$BOT_LOGIN" \
  '[.comments[] | select(.author.login != $bot and .createdAt > $ts)] | length')

• new_human > 0 → human-feedback re-run. Human feedback is the highest priority signal. Read every such comment verbatim and prepend it to the failure description you hand to Stage 3 (fix-root-cause takes a free-form failure description; a leading "Maintainer feedback since last run: ..." block steers the re-analysis without any new leaf input). Run the full pipeline from Stage 1; do not take any of the skip fast-paths below. A human saying "still wrong" or "change X" overrides any cached verdict.
• new_human == 0 → bare re-run. Continue to Step 0.3.

Step 0.3: Bare re-run — skip what is still valid

No human pointed anything out since the last agent comment, so re-doing the whole pipeline would just repeat identical work. Re-run only the cheap front of the pipeline and compare against last time:

1. Run Stage 1 (triage) and Stage 2 (reproduce) as normal.
2. Compare the reproduce result to the previous run. The previous refined_command + verdict are recoverable from the last <!-- agent:root-cause --> comment's analyzed_sha context, or re-derived by reading the prior <!-- agent:reproduce --> / sweep comment. "Identical" means same verdict and same refined_command.
   • Reproduce differs (now passes, or a different command reproduces) → the situation changed on its own; resume the full pipeline from Stage 3 (root-cause) with the new reproduce result. Do not reuse the cached root-cause.
   • Reproduce identical → nothing observable changed. Hand off to Stage 3, which runs fix-root-cause's own <!-- agent:root-cause --> analyzed_sha fast-path: if target_repo HEAD sha equals the recorded analyzed_sha, that leaf re-emits the prior verdict verbatim and this orchestrator stops (the earlier outcome — fix already staged, or NEEDS_HUMAN — still stands; there is nothing new to do). If the sha moved, fix-root-cause re-analyzes and the pipeline continues from Stage 3 as usual.

This never skips Stage 1/Stage 2 — they are cheap (text + nightly wheel) and are the only way to notice the failure went away. It only avoids the expensive Stage 3-5 rebuild+fix when both the observable failure and the analyzed code are unchanged.

Stage 1 — Triage (issue-triage)

Call issue-triage on the issue body + comments. It emits issue_type (single-bug / batch-bug / nonbug), batch_kind (skip-list / heterogeneous / null), reproduction_missing (yes / no), scope, runtime_dependencies, and a preliminary handling (agent-fixable / needs-human).

Branch on issue_type (triage already made the batch-vs-nonbug call; no re-detection here):

• nonbug → stop the fix pipeline; skip to Stage 6 Report with SKIPPED(reason=nonbug).
• single-bug with reproduction_missing=yes → stop; Stage 6 Report with NEEDS_HUMAN(reason=reproduction_missing). issue-triage's own comment already asks the reporter for a reproducer.
• single-bug with reproduction_missing=no → continue to Stage 2.
• batch-bug → Stage 1u, passing through batch_kind (the loop uses it only to label a NOT_REPRODUCED sub-item).

Stage 2 — Reproduce (fix-reproduce)

Only for the single-bug path (issue_type=single-bug). Call fix-reproduce with:

• reproducer_command — extracted by issue-triage from the issue body.
• stage=nightly — reproduce against the nightly wheel.
• ci_repo — inferred from repo (torch-xpu-ops for issues on intel/torch-xpu-ops, pytorch for pytorch/pytorch), or the value the bot passes explicitly.

Branch on its verdict:

• REPRODUCED → continue to Stage 3. Record the refined_command and base for downstream stages.
• NOT_REPRODUCED → the issue is stale; Stage 6 Report with SKIPPED(reason=no_longer_reproduces).
• NO_REPRODUCER → Stage 6 Report with NEEDS_HUMAN(reason=no_reproducer).
• CANNOT_VERIFY → Stage 6 Report with NEEDS_HUMAN(reason=cannot_verify) and the blocker field.

Stage 1u — Batch-issue fan-out

One loop for every parent issue that tracks multiple children. The parent lists several sub-items; run the single-bug pipeline on each independently, each on its own fix branch, and report every outcome back on the parent. "Fix what you can" — a sub-item that can't be fixed is marked and the batch continues.

Entered for issue_type=batch-bug. issue-triage already set batch_kind; the two kinds share the entire loop and differ only in how a NOT_REPRODUCED sub-item is labeled (see step 2 below):

• heterogeneous — the parent body lists distinct sub-bugs. A NOT_REPRODUCED entry means that bug is ALREADY_FIXED (no longer reproduces); nothing to do.
• skip-list — a Bug Skip issue listing homogeneous already-skipped tests. A NOT_REPRODUCED entry means the test now passes, so its skip decorator is now stale and should be removed (follow-up).

No child GitHub issues are created. Each sub-item is a checklist entry on the parent; its fix lives on a dedicated branch so a human can open one PR per fixed sub-item.

Extract sub-items

Parse the parent body into a list of sub-items. Each is either:

• an inline sub-bug — a checklist line naming a test node id or reproducer (skip-list entries are always this form; normalize a bare Class::method to a node id, fix-reproduce's Prepare step resolves the file), or
• a linked child reference owner/repo#N — fetch that issue and use its body as the sub-item's failure description. Only follow intel/torch-xpu-ops and pytorch/pytorch references; ignore any other repo (untrusted, per fix-root-cause).

A sub-item may also arrive as a plain node id from the caller (a nightly report, an email, a log excerpt): normalize Class::method to a node id, drop blank lines, headers and comments, and collapse exact duplicates.

Split infra failures out before the loop. A runner crash, a docker pull timeout, a full disk or a missing device is not a test failure and will not reproduce. Record these as NEEDS_HUMAN(infra) and report them; they never enter the loop.

Give each sub-item two identifiers used for branch naming:

• seq — 1-based position in the body checklist order. Readable and maps a branch back to its body line. Not the stable identity (editing or reordering the body changes it).
• slug — the stable identity, derived from the sub-item's test node id: take the leaf method name, strip device suffixes (_cpu / _xpu / _cuda / _meta) and any dtype suffix, lowercase, keep [a-z0-9._-], truncate to 40 chars. If two sub-items produce the same slug, append -2, -3, … in body order so every slug is unique within the issue.

Branch name is agent/fix-issue-<N>-<seq>-<slug> (N = parent issue number). On a re-run, match a sub-item to its prior branch by slug — look for an existing agent/fix-issue-<N>-*-<slug> (any seq); if found, that is the same sub-item (rename to the current seq if it moved, never create a duplicate). Skip headers, prose, and empty lines.

Preflight (many entries): install nightly wheel once

When the list is long (skip-list issues routinely have dozens), front-load the one real wheel install so the per-entry fix-reproduce(stage=nightly) calls each find the env already current and return fast. fix-reproduce always issues pip install --pre --upgrade (it refuses to trust a stale wheel); running it once here does the real work. There is no skip_wheel_install flag — this is purely an ordering optimization:

pip3 install --pre --upgrade torch torchvision torchaudio \
  --index-url https://download.pytorch.org/whl/nightly/xpu
python -c "import torch; print('nightly:', torch.__version__)"

One fix may cover several sub-items

Entries in a batch are often the same bug in the same class. Fixing each one separately runs root-cause and a source build (1-3 h) per entry and produces the same patch several times.

Do not guess which entries share a fix:

1. Order the REPRODUCED entries so likely relatives sit together (same test class or file, same error). Ordering only; it decides nothing.
2. Fix the first entry.
3. Re-run the other entries in that group against the staged fix. The build is already done, so this costs one test run each. fix-verify already produces a before/after table; extend it to the group.
4. Entries that now pass go in the fix's covers list and are not fixed again. Entries that still fail take their own turn from step 2.

One fix gets one branch and one fix_result.json, naming in covers every entry it also fixed.

Per-sub-item loop

Capture the two base SHAs once, then for each sub-item reset both checkouts per the shared reset-between-entries recipe — a prior sub-item's staged diff must not bleed into the next.

For each sub-item:

1. Reset both checkouts to the base SHAs (shared recipe).

2. Reproduce (fix-reproduce). It detaches the pytorch tree to its base (origin/main or the ci_commit fallback) and returns that base plus a refined_command.

   stage follows the issue, like allow_skip:

   • CI break (pytorch-ci-failure, or a mirrored DISABLED test) → stage=auto. Many CI failures only reproduce on a source build. stage=nightly returns NOT_REPRODUCED as soon as the nightly wheel passes and does not fall through, which records unfixed work as already fixed.
   • Everything else → stage=nightly.

   Branch on the verdict:

   • REPRODUCED → continue to step 3; keep its base and refined_command.
   • NOT_REPRODUCED → nothing to fix; record it and go to the next sub-item. The label depends on batch_kind:
     • heterogeneous → ALREADY_FIXED: the reported bug no longer reproduces on latest nightly (upstream fixed it, or it was flaky). No action needed.
     • skip-list → STALE_SKIP: the test now passes, so its skip decorator is obsolete. Record a follow-up to remove the decorator — the orchestrator does not delete it here (see "Stale skips" below).
   • NO_REPRODUCER → INVALID_ENTRY (renamed/removed test, or malformed). Record, continue.
   • CANNOT_VERIFY → UNVERIFIED (environmental). Record, continue.

3. Root-cause, then branch. Run Stage 3 (fix-root-cause) first — it returns target_repo, which decides which checkout the fix (and thus the branch) lives in:

   • target_repo=pytorch → target_repo_dir = pytorch_dir, base is the reproduce base on the pytorch tree.
   • target_repo=torch-xpu-ops → target_repo_dir = pytorch_dir/third_party/torch-xpu-ops, base is xpu_ops_base (the submodule's pinned commit from the shared recipe). fix-implement / fix-verify handle the xpu.txt pin rewrite internally; the orchestrator only creates the branch.

   Create the isolated fix branch on target_repo_dir off its base so the diff is pushable on its own:

   git -C "$target_repo_dir" checkout -B "agent/fix-issue-${N}-${seq}-${slug}" "$base"
   

   Then run Stage 4 → 5 (same contract and 3-attempt bound as the single-bug path). fix-implement expects exactly this: a fresh branch the orchestrator just created, clean worktree.

4. On any leaf NEEDS_HUMAN / CANNOT_VERIFY / FAILED (or attempts exhausted): mark this sub-item blocked with the reason, record it, and continue to the next sub-item — never abort the whole batch on one hard sub-item.

5. On fix-verify PASSED: do NOT push or open a PR. Update the sub-item's fix_result-${slug}.json in place (written at verdict=PENDING_VERIFY when the branch was committed) to verdict=PASSED with fix_repo_dir, branch, base_sha (the sub-item's base from step 3), changed_files, so the workflow can export base_sha..branch per sub-item. As in the single-bug path, commit the branch and write the PENDING_VERIFY record as soon as fix-implement returns READY, not only on PASSED, so a crash mid-sub-item leaves a recoverable record instead of an orphan branch.

Stale skips (batch_kind=skip-list only)

A STALE_SKIP sub-item's skip decorator can be removed, but this orchestrator does not delete it. Deleting a skip is itself a code change that needs its own verify + PR; folding it into this sweep would mix "the skip is obsolete" with "and here's the removal diff" and obscure the batch outcome. Instead, surface every STALE_SKIP in the report as an explicit follow-up (candidate for a human, or a separately invoked fix-implement run to remove the decorator). The report is the hand-off; nothing is auto-removed.

Fan-out report

Post one summary comment on the parent (or surface to the user in interactive mode), and mirror it into the parent's checklist:

<!-- agent:batch-fanout -->

## Batch fan-out results

Base: <torch nightly version or base sha>

| Sub-item | Outcome | Branch / Reason |
|---|---|---|
| test_bar_xpu_float32 | FIXED | agent/fix-issue-4321-1-test_bar |
| test_baz | NEEDS_HUMAN | cross_repo_coordinated |
| test_qux | NEEDS_HUMAN | attempts_exhausted |
| test_new | ALREADY_FIXED | no longer reproduces on latest nightly |
| test_old | STALE_SKIP | follow-up: remove skip decorator |
| test_dup | COVERED | by test_bar's fix, agent/fix-issue-4321-1-test_bar |
| test_hard | SKIPPED | skip added, tracking issue intel/torch-xpu-ops#1234 |
| test_gone | INVALID_ENTRY | does not collect |

- **FIXED:** N sub-items — one branch each, ready for a human to open
  a PR.
- **NEEDS_HUMAN:** M sub-items — see per-sub-item reason.
- **ALREADY_FIXED:** J sub-items (heterogeneous only) — no longer
  reproduce; the reported bug is resolved, no action needed.
- **STALE_SKIP:** K sub-items (skip-list only) — the test now passes;
  follow up to remove the obsolete skip decorator.
- **INVALID_ENTRY / UNVERIFIED:** P sub-items — malformed/renamed, or
  environmental during reproduce.

*Automated by issue-handler.*

Omit category rows that have no members (a heterogeneous batch has ALREADY_FIXED but never STALE_SKIP, and vice versa). The <!-- agent:batch-fanout --> marker lets a re-run locate and update this same comment in place. On a re-run (Stage 0), only re-process sub-items that are not already FIXED on a live branch, unless human feedback (Stage 0.2) reopens a specific one.

After the loop, go to Stage 6 Report with the aggregate outcome (IMPLEMENTING(fix_verified) if any sub-item was fixed; NEEDS_HUMAN only if every actionable sub-item needed a human — STALE_SKIP follow-ups do not by themselves force NEEDS_HUMAN).

Machine-readable outputs

Written on both paths: the single-bug Stage 4/5 writes fix_result.json, and a batch fan-out writes batch_summary.json plus one fix_result-<slug>.json per fixed sub-item. The <!-- agent:batch-fanout --> comment is for humans; these files let the invoking bot workflow export patches and drive re-verification without re-parsing a comment.

They live under $AGENT_SPACE (the gitignored scratch dir). In pipeline mode the bot workflow always sets it. In interactive mode it is usually unset, which would resolve to /fix_result.json — so fall back to <repo>/agent_space_xpu/ (the same gitignored dir AGENTS.md defines) and report the path you used:

agent_space="${AGENT_SPACE:-$(git -C "$target_repo_dir" rev-parse --show-toplevel)/agent_space_xpu}"
mkdir -p "$agent_space"

• batch_summary.json — one file listing every sub-item:

  {
    "issue": 4321,
    "kind": "batch-bug",
    "batch_kind": "heterogeneous",
    "sub_items": [
      { "seq": 1, "slug": "test_bar", "outcome": "FIXED",
        "branch": "agent/fix-issue-4321-1-test_bar",
        "target_repo": "torch-xpu-ops",
        "fix_result": "fix_result-test_bar.json",
        "summary": "one-line what/why" },
      { "seq": 2, "slug": "test_baz", "outcome": "NEEDS_HUMAN",
        "branch": null, "reason": "cross_repo_coordinated" },
      { "seq": 3, "slug": "test_new", "outcome": "ALREADY_FIXED",
        "branch": null, "reason": "no longer reproduces on latest nightly" }
    ]
  }
  

• fix_result-<slug>.json — for each FIXED sub-item, the same schema the single-bug Stage 5 writes as fix_result.json. It MUST include the keys the workflow's schema gate and patch-export read:

  • verdict — PASSED / FAILED / CANNOT_VERIFY / PENDING_VERIFY. Written incrementally: Stage 4 writes PENDING_VERIFY when it commits the branch, Stage 5 updates it to the terminal verdict. Only PASSED is exported as a normal patch; a non-PASSED verdict that still names a reachable branch + base_sha is salvaged as an unverified/ patch so the committed work is not lost with the runner. Keep branch and base_sha accurate even on a non-PASSED record.
  • target_repo — torch-xpu-ops | pytorch,
  • fix_repo_dir — absolute path of the git repo holding the fix commit,
  • branch — agent/fix-issue-<N>-<seq>-<slug> (single bug: agent/fix-issue-<N>),
  • base_sha — the commit the fix branch was started from,
  • changed_files — list of the files the fix touched, plus needs_build, refined_command, notes. Suffixed by slug so each sub-item can be exported / re-verified independently.

The patch-export step iterates every fix_result*.json and emits one patch series per unit from base_sha..branch (the branches are not pushed); batch_summary.json enriches it (target_repo, summary line). A unit that reports a verified fix but yields no patch fails the step.

Stage 3 — Root cause (fix-root-cause)

Called on the single-bug path, and once per sub-item from Stage 1u. Call fix-root-cause with the failure description and the refined_command from Stage 2.

Branch on its verdict:

• IMPLEMENTING(reason=ok) → record target_repo, domain, analyzed_sha, root_cause, fix_strategy, then create the fix branch before Stage 4 — target_repo is what decides which checkout it lives in, and fix-implement expects a fresh branch with a clean worktree ($base is Stage 2's reproduce base; the batch path already did this in its own step 3, with its own branch name):

  git -C "$target_repo_dir" checkout -B "agent/fix-issue-${N}" "$base"
  

• NEEDS_HUMAN → Stage 6 Report with the specific reason (task_or_feature / feature_gap / hardware_specific / cross_repo_coordinated / no_registered_domain / etc.). Each reason maps to a different final agent:status value; see execution-modes.md.

Stage 4 — Implement (fix-implement)

Call fix-implement with triage_result, pytorch_dir, target_repo_dir (derived from target_repo), and allow_skip:

allow_skip follows the issue:

• CI break — the issue carries the pytorch-ci-failure label, or mirrors a DISABLED test from pytorch/pytorch. allow_skip=true. Fix it in place where you can, in pytorch or here. Skip only when the fix needs a dependency, information you do not have, or feature-sized work; fix-implement then files a tracking issue for the real fix.
• Everything else — allow_skip=false. Never add a skip decorator.

If the caller states the flag, the caller wins.

Branch on the verdict:

• READY(reason=ok) → commit the staged fix onto the branch created in Stage 3, and write the incremental fix_result.json, then continue to Stage 5. Do NOT wait for Stage 5 to persist — if verify then crashes or the context runs out, the committed branch would otherwise be an orphan the Export step flags as a lost fix. Committing here keeps the branch and the hand-off record in lock-step:

  git -C "$target_repo_dir" commit -m "fix: <one-line summary> (#${N})"
  

  Then write fix_result.json (see "Machine-readable outputs" for the path and its interactive-mode fallback) with the fields known so far and verdict=PENDING_VERIFY: target_repo, fix_repo_dir, branch, base_sha, changed_files, analyzed_sha, root_cause, notes. The branch is local and dies with the runner, so this record is what lets the workflow salvage the commit as an unverified/ patch if verification never finishes — without it, a crash after this point loses the work outright.

• NEEDS_HUMAN → Stage 6 Report. The specific reason (skip_outside_target_repo / skip_guard_rejected / no_fix_possible / etc.) drives the final label.

Stage 5 — Verify (fix-verify)

Call fix-verify with refined_command (from Stage 2), target_repo_dir, and changed_files (from Stage 4). fix-verify unconditionally produces the FAIL->PASS before/after table and runs spin fixlint on a passing result — no flags to pass.

Branch on the verdict:

• PASSED(reason=ok) → the fix is verified. The branch and fix_result.json already exist from Stage 4; update the record in place — set verdict=PASSED and add needs_build, the before/after result, and any final notes — then go to Stage 6 with IMPLEMENTING(fix_verified). fix-verify leaves its spin fixlint fixes staged: amend them into the Stage 4 commit (git commit --amend --no-edit) so they reach the exported patch — never a second commit. The workflow exports base_sha..branch as the patch. Never push or open a PR.
• FAILED → loop back to Stage 4 with the failure output as additional context (see "Iterative loop bounds"). On a retry the branch already carries the previous attempt's commit, so amend it (git commit --amend --no-edit after staging the new edits) rather than creating a second commit or a new branch — base_sha..branch must stay a single reviewable fix, and Stage 5's clean-worktree assertion requires nothing left uncommitted. If attempts are exhausted, update fix_result.json to verdict=FAILED with the last failure in notes so the record explains the branch, then Stage 6 Report NEEDS_HUMAN(reason=attempts_exhausted).
• CANNOT_VERIFY → update fix_result.json to verdict=CANNOT_VERIFY with the blocker in notes, then Stage 6 Report NEEDS_HUMAN(reason=<verify's reason>). Do not loop on CANNOT_VERIFY — the environment problem will not fix itself.

Stage 6 — Report

At the end, summarize the outcome. In interactive mode present this to the user in plain language. In pipeline mode advance the issue's agent:status to the terminal stage (DONE / NEEDS_HUMAN / SKIPPED) and update the checklist per execution-modes.md; a batch fan-out run also posts the <!-- agent:batch-fanout --> summary.

Always include in the summary:

• Issue: link/number and one-line title.
• Path: single-bug / batch (with batch_kind).
• Outcome: IMPLEMENTING(fix_verified) / NEEDS_HUMAN(<reason>) / SKIPPED(<reason>).
• Root cause (from Stage 3, if reached).
• Files changed (from Stage 4, if reached).
• Verification (from Stage 5, if reached).
• For a batch: per-sub-item outcome + branch name (FIXED) or reason (NEEDS_HUMAN / ALREADY_FIXED / INVALID_ENTRY / UNVERIFIED), plus any STALE_SKIP follow-ups when batch_kind=skip-list.

If the outcome is IMPLEMENTING(fix_verified), the fix is committed on its agent/fix-issue-<N> branch and recorded in fix_result.json. The invoking workflow reads that, exports base_sha..branch as a patch artifact, and a human applies it. Do not push or open the PR from this skill.

Review request block

When the job ends, the fix workflow appends a template to the session comment for a human to review the fix. Do not write it yourself:

**strong-accepted** <!-- review: strong-accepted -->
**weak-accepted** <!-- review: weak-accepted -->
**weak-rejected** <!-- review: weak-rejected -->
**strong-rejected** <!-- review: strong-rejected -->

notes:

A reviewer copies it into a new comment, keeps one line, and says why.

Iterative loop bounds

The pipeline is not strictly linear. Loop when a later stage invalidates an earlier assumption:

• Stage 5 FAILED → return to Stage 4 (refine the fix).
• Stage 4's Step 3.5 skip-guard rejects → the leaf itself re-runs once; a second rejection returns NEEDS_HUMAN and the orchestrator does not retry.

Bound: maximum 3 fix attempts (Stage 4 → Stage 5 → Stage 4 …). This matches the legacy pipeline's max_agent_attempts. When you stop without success, report NEEDS_HUMAN(reason=attempts_exhausted) with the last fix-verify failure output in reason_detail.

Do not loop on:

• CANNOT_VERIFY at any stage (environment problem, not fix problem).
• NEEDS_HUMAN from any leaf (contract: the leaf already decided it needs a human).
• Stage 3 no_registered_domain (domain registry is a fixed set, looping won't unstick it).

Issue-body status contract

Pipeline mode only. In interactive mode, do not touch the issue body/markers/labels unless the user asks — report to the user instead.

This orchestrator owns advancing the overall <!-- agent:status:X --> marker through:

DISCOVERED → TRIAGING → REPRODUCING → TRIAGED → IMPLEMENTING →
VERIFYING → DONE

with terminal alternates NEEDS_HUMAN and SKIPPED. Stage-by-stage mapping to labels is in references/execution-modes.md; each leaf skill owns its own <!-- agent:<name> --> comment/log slot, this orchestrator owns the overall agent:status marker + the Action Items checklist.