[jonahgabriel]

Ticket: OMN-20392

Lab: host=h201 lane=codeedit-bulk-2a21 head=7dc39e3cd7d983693836b027742a0ba46c3d630a command=in-process code-edit loop (real handler, worktree, checks; Qwen3.8-27B on the h201 model server) observed=loop e58a5fbb-afa0-4c8c-aafe-44615e78b96a accepted, 110 files changed in 4 turns, rubric PASS

Summary

The delegated code edit node allowed at most 12 actions per turn and had no multi-file action, so the crush replay's 105-file and 114-file tasks (f76301f0, 8e1b5f72) needed at least 9 and 10 turns of single edits. This blocks cutover parity (OMN-20291, OMN-20292).

This change adds one typed action, replace_in_files, which replaces every occurrence of old_string with new_string across many files in a single action.

Design

One typed action replace_in_files, declared in the node contract description, TOOL_SCHEMAS and RESPONSE_CONTRACT.

• Arguments: exactly one of file_paths (array, at most MAX_BULK_FILES=200) or glob (same syntax as writable globs, expanded over the worktree manifest); plus old_string (required) and new_string (may be empty).
• Replaces every occurrence of old_string in each file.

Per-file all-or-nothing — content is computed in memory and written once. A named file that is outside the worktree, outside the writable globs, lacks old_string, would exceed MAX_WRITE_BYTES, or whose write is refused is reported FAILED <path>: <reason> and left untouched. Refusals count into the result's refusals.

Glob is a scope — only writable manifest files are targeted; files without old_string are skipped and counted; files matching the glob but outside the writable globs are counted; a glob that edits nothing is an error.

Feedback to the model lists failures first, then each edited file with its occurrence count, capped at MAX_OBSERVATION_BYTES — so a cap cuts successes only.

No shell, no generated script — only this typed action, same confinement and writable globs. The 12-actions-per-turn cap is unchanged.

Rubric:

• replace_in_files added to edits_apply edit_tools.
• file_paths added to no_phantom_paths path_argument_names; no_phantom_paths now reads list-valued path arguments.
• One call is recorded per action (status error when any file failed), because within_budget caps tool calls at 80 — per-file calls would break it for a 105-file edit.

Also fixed: DelegatedCodeEditPorts.delegate raised AttributeError on a failed delegate receipt whose terminal_payload is null; it is now a failed turn reply (test added).

Tests

Focused run of tests/nodes/node_delegated_code_edit_orchestrator, tests/unit/delegated_code_edit and tests/nodes/node_delegation_rubric_check_compute: 382 passed. ruff check and format clean, mypy --strict on src/omnimarket clean, runtime sweep import check and metadata dependency check pass.

New tests cover:

• 30 files in one action and one turn
• glob scope
• glob that edits nothing
• partial failure (one error call, rubric edits_apply fails)
• per-file write error
• over-cap and zero-target refusals
• every-occurrence count
• failures-before-successes ordering
• parser rejections
• schema array types
• rubric edits_apply and no_phantom_paths on replace_in_files

Lab proof

Real Qwen3.8-27B on the h201 lab model server, the real loop handler, worktree and checks, in-process:

• Loop run f477e734-4b9a-49f1-8c73-a1d6562d1304 — 41-file fixture (40 files to migrate): status accepted in 3 turns, one replace_in_files action by glob, 40 files changed, check passed, rubric outcome PASS.
• Loop run e58a5fbb-afa0-4c8c-aafe-44615e78b96a — 110-file fixture across two directories: status accepted in 4 turns, 110 files changed, check passed, rubric outcome PASS.
• Turn delegate run ids of the first: 14ed37af-fe89-4149-bc35-a7789e3bf5f6, baaeb8b9-9d51-46b9-8397-388a6e525a01, 4f5ddbcd-b9c7-4dd5-8726-56463f9e5131 (in the lane scratch state root).

Notes

Limit of the proof: onex code-edit run itself could not reach a model from h201 today. The dev lane failed with "delegate receipt carries no resolvable delegation terminal" (loop f86b9d8f-33a2-4a20-b14e-f1ab7562a4ab), and the in-process route failed with the DelegatePhaseTimedBus.subscribe TypeError (delegate run 141fdc8d-13eb-4c0c-a6e2-dba8ef6a7d8e, loop 1974e6bc-4140-4dc0-a7a1-a1506347d1bd). The lab runs therefore replaced only the onex delegate hop with a direct call to the model server; the prompt, response parsing, apply, checks, receipt and rubric score are the shipped code.

Follow-up: re-running through onex code-edit run on the dev lane once the delegate route is back.

Not covered: the crush replay's actual 105-file and 114-file tasks were not replayed.

Evidence-Ticket: OMN-20392
Evidence-Source: OCC#12449

[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)