Implement — Apply the Fix

Takes triage output and makes the code change. Does not run tests (that is fix-verify's job); staging-only, see HARD RULES.

Contents

• Inputs
• Step 0: Verify environment
• Step 1: Read the triage output
• Step 2: Implement the fix
• Step 3: Stage changes
• Step 3.5: Skip-guard review (allow_skip=false)
• Output
• HARD RULES

Inputs

• triage_result — JSON output from fix-root-cause (root_cause, fix_strategy, target_repo, domain, analyzed_sha).
• PYTORCH_DIR — path to local PyTorch checkout.
• target_repo_dir — path to the checkout that will be edited. Derived from PYTORCH_DIR and triage_result.target_repo:
  • target_repo == "pytorch" → target_repo_dir = PYTORCH_DIR.
  • target_repo == "torch-xpu-ops" → target_repo_dir = <PYTORCH_DIR>/third_party/torch-xpu-ops (per AGENTS.md "Commit Pin & Development Override"; the caller is expected to have cloned the working branch there ahead of time). All git and edit operations in this skill run against target_repo_dir, never against PYTORCH_DIR when the two differ.
• allow_skip — controls skip decorator strategy:
  • false (UT and everything else): never add skip decorators; must unskip and really fix.
  • true (CI break: pytorch-ci-failure, or a mirrored DISABLED test): may add a skip with a tracking issue when the fix needs dependencies, information you do not have, or feature-sized work. Repair in place first; the skip is the fallback. Stale skips must still be removed.

Step 0: Verify environment

basename $(git -C $target_repo_dir rev-parse --show-toplevel)  # confirm which repo
git -C $target_repo_dir status                                 # inspect current state

On the first attempt (fresh branch just created by the orchestrator), the worktree is clean.

On a loop-back attempt (orchestrator returned here after fix-verify FAILED or the reviewer requested changes), staged changes from the previous attempt are still present by design. Do NOT abort. Refine those changes on top; do not git reset or git clean — the orchestrator would have done it if a fresh start were intended.

Step 1: Read the triage output

Read triage_result carefully before touching any file. Understand:

• Which files/functions need to change
• Why the failure occurs (if applicable, why CUDA works but XPU fails)
• Whether the fix is in pytorch or the backend repo
• Which analyzed_sha the triage was against — your edits go on top of that base

If the issue is not yet triaged, run fix-root-cause first.

Step 2: Implement the fix

Key rules

• Minimal changes — fix only what's broken; every changed line must trace to the triage output.
• Stay in your repo — see the domain reference file loaded during triage (../domain-knowledge/domain-<name>.md, per the domain registry) for path conventions.
• Never modify unrelated files.

Fix strategies by category

See fix-root-cause Step 1 for domain routing. Common strategies:

• Tolerance: match upstream atol/rtol values exactly.
• Regression: find the guilty commit (git log --oneline -20 -- <file>), apply a fix aligned with upstream intent; document any divergence in comments.
• Newly added test: enable backend support. If allow_skip=false and support is genuinely missing, report NEEDS_HUMAN — do not add a skip. If allow_skip=true, follow the "Add a new skip" recipe in "Skip operations" below.
• Unknown root cause: compare with upstream backend behavior.

Skip operations

Skip decorators for a failing test live wherever that test lives: the pytorch test tree for upstream tests, test/xpu/ inside torch-xpu-ops for its own tests.

The skip must live inside target_repo_dir. One invocation produces one repo's diff (a two-repo fix is two invocations), and the orchestrator commits (or reads) only target_repo_dir. If the skip would have to be added to a file outside target_repo_dir — e.g. target_repo == "torch-xpu-ops" but the failing test is in pytorch's test/ tree — do NOT edit it: that change would be left uncommitted and then wiped by the orchestrator's reset between failures. Return NEEDS_HUMAN(reason=skip_outside_target_repo) naming the file that would need the skip.

Removing a stale skip (a @skipIfXpu / xfail decorator whose underlying failure this run is fixing): read the test file, delete the decorator lines, save. Nothing else to do — the change goes through the normal git add in Step 3.

Adding a new skip (allow_skip=true only). File a tracking issue first so the skip has a follow-up owner, then edit:

# 1. Create the tracking issue and capture the URL.
issue_url=$(gh issue create \
  --repo intel/torch-xpu-ops \
  --title "[upstream_ut] <test name>: <one-line symptom>" \
  --body "Auto-added by fix-implement (allow_skip=true) while handling
  intel/torch-xpu-ops#<the issue number this run was triggered on>.

  Test: <test_id>
  Original failure: <one-line failure summary from triage_result>
  Root cause: <root_cause from triage_result>
  Reason for skip: <why the actual fix requires human follow-up>
  Base analyzed: <target_repo>@<short_sha from analyzed_sha>
  " \
  --label "test: ut" \
  | tail -1)

# 2. Add the decorator with a comment citing $issue_url so a human
#    can find the tracking issue by grepping the skip in place.
#    Example additions (choose the one matching the test's shape):
#
#      @skipIfXpu(f"see {issue_url}")
#      def test_foo(self): ...
#
#      # or in an OpInfo skips tuple:
#      DecorateInfo(unittest.skip(f"see {issue_url}"), 'TestFoo', 'test_bar', ...)

Emit the resulting issue_url as tracking_issue in the JSON output (see Output section). skip_added becomes true.

Keep the title prefix, the label, and the bare intel/torch-xpu-ops#<N> exactly as above: a label this repo does not have fails the call, and backticking the issue number kills the cross-reference back to it.

Step 3: Stage changes

git -C $target_repo_dir add <your_files>
git -C $target_repo_dir diff --cached --stat   # verify only intended files are staged

Never stage unrelated files. In particular, never stage third_party/xpu.txt — it is a submodule pin managed by build tooling, not part of any bug fix (see HARD RULES).

Step 3.5: Skip-guard review (only when allow_skip=false)

Skip this step entirely when allow_skip=true (the CI-break flow).

When allow_skip=false, spawn a fresh-context subagent via the Task tool (subagent_type=general-purpose) to inspect the staged diff for skip-shaped workarounds before returning to the orchestrator. The implementer must not review its own diff for this specific bias — the gatekeeper here is a separate agent with no memory of the reasoning that produced the diff.

Pass the reviewer:

• The full staged diff: git -C $target_repo_dir diff --cached.
• The triage_result (so the reviewer knows what the root cause is supposed to be).
• The value of allow_skip (always false when this step runs).

Instruct the reviewer to reject the diff (return REQUEST_CHANGES) if any of the following appears anywhere in added lines:

• A new @skipIfXpu / @skipXPU / @skipCUDAIf / @skipMPS / @unittest.skip / @unittest.skipIf / @pytest.mark.skip / @pytest.mark.skipif / @pytest.mark.xfail / @expectedFailureXPU / @expectedFailureCUDA / @expectedFailureMPS / @expectedFailure decorator on a test.
• A new DecorateInfo(unittest.skip, ...) / DecorateInfo(skipIfXpu, ...) / DecorateInfo(unittest.expectedFailure, ...) entry in an OpInfo / ModuleInfo skips/decorators list.
• A new xfail(...) / skip(...) entry in an instantiate_device_type_tests skip dict for XPU.
• Bare raise unittest.SkipTest(...) / self.skipTest(...) inserted into a previously-running test to short-circuit it on XPU.
• Loosening atol / rtol on assertEqual (or any tolerance-carrying assertion) by more than an order of magnitude, when the diff has no quantitative justification for the new value in a comment.
• Hardcoded set_rng_seed(...) / torch.manual_seed(...) / random.seed(...) inserted into a previously-random test purely to dodge a failure region.
• Deleting or commenting out the failing assertion / the failing test function.
• A broad try / except Exception: pass (or equivalent) wrapping the call that used to fail.

Existing skips being removed by the diff are fine — that is a legitimate root-cause fix in the "stale test expectation" category. The rule only fires on added skip-shaped constructs.

The reviewer returns one of:

• APPROVE — no skip-shaped workaround found. Continue to Step 4 (output).
• REQUEST_CHANGES — cite each offending hunk (file + line + which rule it matched). The implementer MUST address every citation (either replace the workaround with a real root-cause fix or, if no root-cause fix is possible within this run's scope, unstage the offending change and return NEEDS_HUMAN(reason=skip_guard_rejected) to the orchestrator with the reviewer's citations attached).

Do not loop this step more than once. If a second run of Step 3.5 still returns REQUEST_CHANGES, unstage the offending change and return NEEDS_HUMAN(reason=skip_guard_rejected) — that is a signal the fix cannot be produced without a workaround and a human should take it.

This step is intentionally narrower than the orchestrator's Stage 5.5 review. Stage 5.5 checks the entire diff for correctness, minimalism, and root-cause alignment; Step 3.5 checks only for the specific class of "hide the failure instead of fixing it" workarounds that allow_skip=false is meant to forbid. Both run; they do not replace each other.

Output

Return to the orchestrator a report (markdown block plus JSON block). The skill does not commit, push, or open PRs — the caller consumes stdout and decides what to do, per the pattern established by issue-triage, fix-reproduce, and fix-root-cause.

Include the <!-- agent:implement --> marker on the first line of the markdown block so a downstream caller can locate its own previous implement comment (if any) and update it in place. Comment location and update is the caller's responsibility.

<!-- agent:implement -->

## Implement Result

- **Target repo:** <pytorch | torch-xpu-ops>
- **Analyzed at:** <target_repo>@<short_sha>
- **What I changed:** <per file: what changed in it, with the line count>
- **Why:** <one sentence connecting each change to the triage root cause; must cite file:line for every upstream/CUDA comparison — see the hard rule below>
- **Skip added:** <yes (tracking: intel/torch-xpu-ops#N, url: <url>) | no>
- **Ready for verify:** <yes | no>

```diff
<git diff --cached, or the key hunks when the full diff would blow the
65000-char comment limit>
```

*Automated by fix-implement.*

{
  "target_repo": "pytorch or torch-xpu-ops",
  "analyzed_sha": "<full 40-char sha inherited from triage_result>",
  "changed_files": ["path/to/file1.py", "src/ATen/native/xpu/Foo.cpp"],
  "skip_added": false,
  "tracking_issue": null,
  "allow_skip": false,
  "covers": [],
  "ready_for_verify": true,
  "verdict": "READY or NEEDS_HUMAN",
  "reason": "<enumerated reason code, see below>",
  "reason_detail": "one-line human-readable detail"
}

Field contract

• target_repo / analyzed_sha — echo from triage_result so downstream stages have a self-contained record without re-reading triage output.
• changed_files — list of paths (relative to target_repo_dir) that are staged. fix-verify reads this to decide whether a C++/SYCL rebuild is required. Must equal git -C $target_repo_dir diff --cached --name-only at output time. Re-run that command immediately before emitting the JSON block; do not cache a pre-Step-3.5 file list, because Step 3.5's reviewer may have unstaged offending files.
• covers — other batch entries this patch also fixes, confirmed by re-running them against the staged fix (see issue-handler, "One fix may cover several sub-items"). Node ids or issue numbers; empty on the single-bug path. They are not fixed again.
• skip_added — true only when this run added a new skip decorator under allow_skip=true. Removing a stale skip is NOT skip_added.
• tracking_issue — issue URL from the "Add a new skip" recipe in Step 2 when skip_added=true; null otherwise. The orchestrator reads this for the "Skipped (with tracking issue)" rows of its fan-out report.
• allow_skip — echo the input flag verbatim, so a reviewer can tell by looking at the output alone whether Step 3.5 ran.
• ready_for_verify — true when Step 3.5 (if it ran) returned APPROVE and staged changes exist; false if the implementer decided to bail out with NEEDS_HUMAN (in which case the orchestrator should not call fix-verify).
• verdict — READY (staged diff exists, fix-verify may run) or NEEDS_HUMAN (see reason).

reason values

On verdict=READY: ok.

On verdict=NEEDS_HUMAN:

• skip_outside_target_repo — the skip decorator that would need to be added lives outside target_repo_dir; see Step 2's "Skip operations".
• skip_guard_rejected — Step 3.5's subagent reviewer returned REQUEST_CHANGES twice, or the implementer chose to bail out rather than fight the reviewer.
• no_fix_possible — the fix strategy in triage_result is not implementable without either widening the diff outside target_repo or adding a skip when allow_skip=false.
• other — fallback; put full explanation in reason_detail.

Contract: this leaf leaves the changes staged (git add) and does not commit — leaves never commit, branch, or push. What the caller does with the staged diff (commit it onto a branch, or hand the staged diff straight to its workflow) is the caller's business; this skill only stages.

HARD RULE: every upstream/CUDA claim in "Why" must cite a source

The "Why" line written here will be read by reviewers as factual statements. Before writing any of the following phrases, you MUST look up the specific file and line number that backs it up:

• "consistent with upstream"
• "upstream does X" / "upstream already handles this"
• "same as CUDA/MPS/ROCm"
• "mirrors upstream behavior"
• "aligned with upstream"

If you cannot find a specific file:line to cite, do not write the phrase. Replace it with a direct statement of the observable fact, e.g.:

A "Why" that contains an unsubstantiated upstream comparison is worse than one that only cites what you directly observed — it inflates reviewer confidence in a claim that was never verified.

HARD RULES

• NEVER add skip decorators when allow_skip=false.
• NEVER edit a file outside target_repo_dir — including when the skip or fix "belongs" in the other repo. That diff cannot be committed or read back by the orchestrator; return NEEDS_HUMAN instead.
• When allow_skip=false, Step 3.5 (skip-guard reviewer subagent) is MANDATORY before returning to the orchestrator. Do not skip it, do not run it inline in your own context.
• NEVER stage third_party/xpu.txt. It is a submodule pin managed by build tooling; staging it is never part of a bug fix, regardless of domain (xpu-kernel / inductor / upstream-pytorch).
• NEVER cherry-pick upstream commits. If a fix already landed on trunk, rebase (git rebase origin/main) instead.
• NEVER commit, push, tag, or open a PR. This skill only stages the diff (git add); the workflow that invoked it reads the staged diff after fix-verify passes and exports it as a patch (no branch, no push) for a human to apply.