[jonahgabriel]

Summary

Fix a crash in the delegated code edit loop when a delegate fails (OMN-20390).

DelegatedCodeEditPorts.delegate (src/omnimarket/delegated_code_edit/loop_ports.py) read the
delegate receipt as receipt["receipt"]["result"]["terminal_payload"]["payload"]. A failed
delegate's receipt carries "terminal_payload": null (key present, value None); because
dict.get(key, {}) returns None for a present null key, the next .get raised
AttributeError at loop_ports.py:435. The exception escaped
HandlerDelegatedCodeEditOrchestrator.run, so no loop receipt was written and
<state root>/runs/<correlation id>/loop_claim stayed on disk. First seen on a failed run on
.202 (run d642a86f-8780-45cd-8c4f-412fa938dc68).

• Base: dev · Head: codeedit-loop-fix-2a21/OMN-20390-null-terminal

Changes

1. Null-safe receipt reader (loop_ports.py, _read_delegate_receipt). Any level
   (receipt, result, terminal_payload, payload, metrics) may be null or not a dict;
   receipt.json may be unreadable or not JSON; non-numeric token counts become 0. It also
   builds a one-line failure description from the receipt (status, terminal_failure_cause,
   terminal_failure_reason, failure_reason, result.error, result.runtime_error_type,
   and terminal_payload is null).
2. Typed failed reply from delegate. DelegatedCodeEditPorts.delegate returns a typed
   failed ModelTurnReply (ok=False, run id, model and tokens from the receipt) whenever
   there is no result text, with the receipt failure description appended to invalid_reason.
3. Claim/receipt hygiene. claim_loop_receipt refuses when a loop receipt already exists
   (checked before and after the exclusive create). write_loop_receipt writes atomically
   (temp file then os.replace) and always releases loop_claim in a finally. A receipt's
   existence is what refuses a rerun.
4. Receipt on every raising exit path. HandlerDelegatedCodeEditOrchestrator.run catches
   any unexpected exception from the turn phase as status infra_error with detail
   unexpected <Type>: <message>, and guards the diff and score ports, so a loop receipt is
   written on every exit path that raises an exception.

Proof (RED/GREEN)

Test paths:

• tests/unit/delegated_code_edit/test_loop_ports.py

• tests/nodes/node_delegated_code_edit_orchestrator/test_handler_delegated_code_edit_orchestrator.py

• RED — the same tests against the dev source (origin/dev at dadcc91b7):
  42 failed, 57 passed. The headline test fails there with
  AttributeError: 'NoneType' object has no attribute 'get' at loop_ports.py:435.

• GREEN — on this branch, a focused run of those two paths: 99 tests pass.

Lab runs

Real onex code-edit run, in-process delegate, on host .201, task: make add return the
sum, max 6 turns.

• dev source: loop run ada0a179-449d-4e0e-9d20-996f4833f540 raised AttributeError at
  loop_ports.py:435; loop_claim left on disk; no loop_receipt.json.
• this branch: loop run d584b9f2-8865-4f4d-8f10-89e9ac54b81c ended status
  delegate_failed after 2 turns, loop_receipt.json written, no loop_claim; delegate runs
  44c4aeb3-e3ee-4544-b269-14dc76087db9 and bfbf2ec9-0bd7-4e3e-8c82-9212bb2f9dbe. The real
  receipt had terminal_payload null and failure_reason
  "DelegatePhaseTimedBus.subscribe() takes 2 positional arguments but 4 positional arguments (and 1 keyword-only argument) were given";
  the loop receipt now names it.
• dev lane (deployed locus): branch run 0040ad07-4dfc-423f-954e-a992bee7801a ended
  delegate_failed with a receipt and no claim.

Notes

Resuming a delegate_failed loop and keeping the full error in the receipt are tracked
separately under OMN-20391 and are not in this change.

Evidence-Ticket: OMN-20390
Evidence-Source: OCC#12445

[github-actions]

✅ Architectural Review — PASSED

Errors: 0
Warnings: 0

What this checks

Rule	Description
ARCH-TOPIC-001	No hardcoded Kafka topic strings in handler code
ARCH-DI-001	No event_bus=None bypass
ARCH-DI-002	No direct Handler instantiation outside workflow_runner/adapters
ARCH-DI-003	No reinvented DI containers
ARCH-TOPIC-002	contract.yaml topics follow `onex.{cmd

---

No architectural violations found.

Static architectural lint — no model inference (OMN-14176).

[github-actions]

✅ Hostile Reviewer — PASSED

Blocking findings (critical/error): 0
Total findings: 13
Models succeeded: qwen3-review,gpt-oss-review

---

Gate semantics

Verdict	Meaning	Blocks merge?
passed	>=2 models succeeded, no critical/error findings	No
blocked	CRITICAL or ERROR findings found	Yes
degraded	Fewer than 2 models succeeded (infra unavailable/timeout)	Yes (OMN-15110)

Powered by omniintelligence.review_pairing.cli_review — node-based adversarial review via HandlerLlmCliSubprocess (OMN-8524)

[jonahgabriel]

Lab: host=h201 (hostname omninode-pc); lane=codeedit-loop-fix-2a21; head=9dff5631a5ea38ac772a55f360a34cb274811fcb; command: uv run onex code-edit run --worktree --task-file --writable 'src/*.py' --check 'tests= -m pytest -q tests/test_m.py' --omnibase-path $OMNI_HOME --delegate-in-process --max-turns 6 (real onex delegate, in-process); observed on dev source (dadcc91): loop ada0a179-449d-4e0e-9d20-996f4833f540 raised AttributeError at loop_ports.py:435, loop_claim left on disk, no loop_receipt.json; observed on this head: loop d584b9f2-8865-4f4d-8f10-89e9ac54b81c ended delegate_failed after 2 turns with loop_receipt.json written and no loop_claim; the receipt named the real cause (terminal_payload is null; DelegatePhaseTimedBus.subscribe() takes 2 positional arguments but 4 positional arguments (and 1 keyword-only argument) were given).