[jonahgabriel]

Ticket: OMN-20291

Lab: hosts h201 and h202, model turns on the .201 dev lane (Qwen3.8-27B). The OMN-20291 comparison's own run2_<id>.sh replay scripts were run with this branch first on PYTHONPATH, twice per task. Results are in the table below. The comparison's original outputs were set aside and restored afterwards.

Summary

The OMN-20291 comparison replayed 26 crush tasks through the delegated code-edit node. The tool_use rubric passed 18 node runs and 17 crush runs, and the node failed 8 tasks. Reading each failure's loop receipt and turn replies turned up four causes in the node itself. This PR fixes all four:

cause	replay evidence	change
Read loop	a16a3138: 40 turns, 132 views, 0 edits. e3a2c923: 40 turns, 215 views, 1 edit. From turn 3 on, every prompt sat at the 90,068-character cap and its history held only the last 1 to 3 turns. One turn's reads (8 to 12 views of up to 16 KB) were larger than the whole history, so even the newest turn was cut.	Every earlier turn keeps its action lines. The newest turns keep their output, a re-viewed window points at the later turn, and a view of a file changed later is marked stale. One turn's reads total at most 30,000 characters. After 3 turns that read and change no file, the next turn's reads are refused, and its prompt says so.
Absolute path inside the worktree refused	58744096 t6a1 grep path=/data/.../wt/58744096/omnimarket/src was refused with leaves the worktree. The model copied the path from the loop's own echo of the check command ($ env PYTHONPATH=/data/.../src ...).	Such a path is used and recorded as its worktree-relative form. Any other absolute path is still refused.
Edit anchor indentation slip	ea3ee264 t3a7 and t5a1: the view prints 96| text, and the model copied the separator's blank into every line. In re-run 2 it did so on the first line only (' def umbrella_tickets('). t7a2 re-sent an edit it had applied in turn 3.	An old_string that matches exactly one place up to that shift is applied, and new_string moves back the same way. An edit whose result is already in the file is reported applied and writes nothing, so a re-sent insertion is not inserted twice.
Bulk-edit cap counted the glob (#3270)	In the #3270 AC2 replay (loop 7ea45ffe), replace_in_files glob=src/omnimarket/nodes/*/contract.yaml was refused with targets over 200 files. The model fell back to write /tmp/... and 117 single edits, which is the 8e1b5f72 failure.	The cap counts the files the glob would edit, meaning those that hold old_string.

Two parts of the history change depend on the code that #3276 and #3270 brought to dev. History turns are now structured, so the resume block from #3276 stores them as JSON. replace_in_files counts as a file change for the read-pause streak.

How the read-loop fix got here

Each step was driven by a lab re-run:

1. History rendering alone. e3a2c923 was still only viewing at turn 30, because the newest turn's own reads overflowed the history.
2. History plus a per-turn read bound. The model kept choosing to read until the turn cap.
3. History, read bound and read pause. a16a3138 completed: accepted in 29 turns, every declared check passed, against 40 turns and no edits before. e3a2c923 completed: accepted in 38 turns with both checks passing, against budget_exhausted before.

Both tasks still exceed the rubric's 80-tool-call budget. Crush finishes them in 8 and 5 turns. The remaining gap is prompt capacity: onex delegate takes the prompt as one argv word, which caps it at 120 KB. That is a separate omnibase_infra change, and the report describes it.

Tests: RED, then GREEN

The focused tests are all in tests/nodes/node_delegated_code_edit_orchestrator/test_handler_delegated_code_edit_orchestrator.py. Each group failed before its change:

• History, absolute path and anchors: 8 failed. The uniqueness guard already passed.
• Read bound: 2 failed.
• Read pause: 2 failed.
• First-line anchor slip: 1 failed.
• Glob cap: 1 failed.

The resume tests from #3276 that asserted string history blocks now assert structured entries. The cut-not-dropped history test now builds HistoryTurns.

GREEN, focused only: uv run pytest tests/nodes/node_delegated_code_edit_orchestrator/ tests/unit/delegated_code_edit tests/unit/cli/test_cli_code_edit.py tests/nodes/node_delegation_rubric_check_compute -q gave 504 passed. ruff format, ruff check and mypy --strict on the node and delegated_code_edit are clean. pre-commit passed on the changed files.

ea3ee264 re-run 2's turn-3 replies, replayed offline against the base file, apply all nine edits with the file's indentation.

Lab re-runs

The code under test is e793efaee. The PR head 82534fca4 is that code plus dev, merged in by the OCC bot, and the ticket contract. Each task ran twice through its own run2_<id>.sh. Loop ids are in brackets.

task	class	node before	fixed, run 1	fixed, run 2
ea3ee264	edit anchors	FAIL edits_apply	PASS, 7 turns / 404 s (00979abd)	PASS, 10 / 390 s (cd48d7c5)
58744096	absolute path	FAIL no_phantom_paths	PASS, 6 / 162 s (068a6dff)	PASS, 7 / 194 s (0fbff7d4)
a16a3138	read loop	budget_exhausted at 40 turns, 0 edits	accepted, all checks pass; FAIL within_budget (90 calls > 80), 22 / 864 s (a88d8355)	PASS, 18 / 801 s (aa8c5957)
e3a2c923	read loop	budget_exhausted at 40 turns, 1 edit	budget_exhausted at 40 turns, 33 edits (6df81ab0)	void: the dev lane failed the delegate twice in a row at turns 23, 37 and 38 (dc0909e6, d36c82e6, 0315be42)
8e1b5f72	bulk-edit cap	FAIL, 3 criteria, 23 turns	accepted, both checks pass; FAIL edits_apply (a named-list replace_in_files with files lacking old_string), 7 / 394 s (a37e15a5)	no_progress; a glob reached files outside the task's list (69210626)
f76301f0	bulk-edit cap	FAIL, 3 criteria, 15 turns	accepted, both checks pass; FAIL on one refused helper-script write, 5 / 454 s (c7f97755)	no_progress; a glob changed 5 contracts outside the list (dbd4ab25)

• Fixed: edit anchors and absolute paths, with 4 of 4 runs passing.
• Read loop: a16a3138 now completes, and passed the rubric once. e3a2c923 now edits but still runs out of turns.
• Bulk class: replace_in_files runs now. Its remaining failures come from how the model uses it, and the report covers them.
• Earlier commits on this branch: their runs (58744096 6/6 PASS, ea3ee264 3/4 PASS) are in the report.

Report with the full 26-task table and every failure classified: knowledge-base-internal reports/2026-10-03-code-edit-node-vs-crush-replay.md (knowledge-base-internal#986).

Evidence-Ticket: OMN-20291
Evidence-Source: OCC#12497

[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: 10
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)