[jonahgabriel]

Summary

The worst answer-quality class in the last 24 h of lab delegation is code_generation turns from the delegated code-edit loop refused no_schema_conforming_json: the model answers a prose line instead of the declared JSON object. This declares supports_response_format_json_schema: true on local-coder, so the code-edit response contract goes out as a provider-native json_schema constraint instead of prose only, and closes each action in that contract over its own tool's arguments so the gate and the loop's parser accept the same replies. On the same 20 failing turn prompts, first-attempt acceptance on local-coder went from 12/20 to 20/20.

Ranking (lab dev lane, last 24 h)

task_type	events	gate failed	failure rate	local-rung climbs
code_generation	938	43	4.6%	468
document	611	34	5.6%	16
review	117	29	24.8%	0
summarization	235	14	6.0%	31
test	22	3	13.6%	10

Most terminal failures in document and review are timeouts, provider errors and runtime shutdowns, not answer quality. In code_generation the largest answer-quality class is DELIVERABLE_EXTRACTION: no_schema_conforming_json: 304 attempts across 170 events, every one on a code-edit loop turn; local-coder 234 of its 747 code-edit attempts, local-omnipc2-chat 48 of 485.

Cause

The response contract reaches local-coder only as prose inside a 20-30k token prompt, and the model drifts into narration instead of emitting schema-conforming JSON.

• Read 20 failing code-edit turn prompts from the projection and replayed them on the .201 endpoint the way local-coder sends them (same system prompt composition, enable_thinking false, no response_format): 18 of 23 replies refused no_schema_conforming_json. The replies were one prose line such as "I'll write a Python script to do both steps, then run it." or "I need to see the remaining sections of the files I'll edit before making changes.", or empty content; one repeated "</no_think>" lines.
• The prompt states the JSON reply format once at the top of a 20-30k token prompt whose HISTORY renders actions as "> view(path=...) -> ok"; the model drifts into narration.
• Category: the response contract. It reaches local-coder only as prose. local-coder does not declare supports_response_format_json_schema, while local-heavy-reasoning, on the same endpoint (BIFROST_LOCAL_CODER_ENDPOINT_URL) and served model Qwen3.8-27B, does.
• Same replay with response_format json_schema: 0 of 20 refused; 2 of 20 still rejected by the loop's parse_turn_reply ("action 6 (grep) takes no name, summary") because the flat item schema allowed every argument on every tool; with a per-tool contract both parsed.
• The earlier lane's code_review prompt-template change was not repeated; this is a capability declaration and a response-contract change, not a template change.

Change

• src/omnimarket/configs/bifrost_delegation.yaml: local-coder declares supports_response_format_json_schema: true, with the probe evidence in a comment.
• node_delegated_code_edit_orchestrator/handlers/turn_protocol.py: RESPONSE_CONTRACT items are an anyOf of one closed object per tool (two for replace_in_files: file_paths or glob), derived from ALLOWED_ARGUMENTS and REQUIRED_ARGUMENTS; parse_turn_reply, build_turn_prompt and TOOL_SCHEMAS unchanged.
• Tests updated for the new contract shape; new tests for contract/parser agreement and for local-coder declaring the flag.

A/B: same 20 prompts through onex delegate

All arms: onex delegate --task-type code_generation --backend-id local-coder --response-contract <code-edit contract> over the same 20 failing code-edit turn prompts; pass = the first attempt accepted by the quality gate.

arm	route	config	first attempt accepted	other runs	runs ending accepted (on local-coder)
before	dev lane (--bus kafka --lane dev --locus deployed-lane)	deployed omnimarket 0.4.283	12/20	8 first attempts refused no_schema_conforming_json; 29 attempts in all	19/20 (15/20)
before	in-process (--bus inmemory --locus in-process)	packaged contract	12/20	2 accepted on a retry, 5 failed after 3 local-coder attempts, 1 cancelled at the 240 s budget	14/20 (14/20)
after	in-process	this branch's bifrost_delegation.yaml and RESPONSE_CONTRACT	20/20	none; 20 attempts in all	20/20 (20/20)
control	in-process	this branch's RESPONSE_CONTRACT, flag off	9/20	7 accepted on a retry, 4 failed after 3 local-coder attempts	16/20 (16/20)

The in-process arms raise local-coder's max_grounded_input_tokens from 8000 to 60000 in both before and after, because the in-process path refuses these 20k-token prompts at the 8000 grounding budget and the dev lane does not; nothing else differs between before and after. The after arm cannot run on the dev lane until this deploys, because the deployed lane runs the released contract and a pin to local-heavy-reasoning for code_generation is refused by routing; the control arm shows the json_schema directive carries the gain.

Lab readback

Runtime-affecting (routing contract). Lab evidence above is from h201 against the .201 endpoint, after-arm run ids d7086335-43b2-4125-a15d-fa2ec38966a6, 0b47cc17-7670-43a6-8fcf-a6268354532f, ce33b7d1-7b19-46e9-9c75-c92296e6bbf0 (first three of twenty). After deploy, the readback is the same 20 prompts on the dev lane pinned to local-coder, expecting first-attempt acceptance near 20/20, and the no_schema_conforming_json count on local-coder in delegation_events falling.

Tests

Focused: test_handler_delegated_code_edit_orchestrator.py, test_declared_label_set_membership_omn18989.py, test_loop_ports.py, test_tool_use_rubric.py: 336 passed. Routing/config files (same-model spread, in-process spread, weighted spread, runtime dispatch port, wire completion fidelity, backend id model family, routing authority gate, bifrost renderer reducer seam): 75 passed. ruff check, ruff format --check and mypy --strict on turn_protocol.py clean; pre-commit passed.

Ticket: OMN-17427

Evidence-Source: OCC#12518
Evidence-Ticket: OMN-17427

[github-actions]

⚠️ Architectural Review — WARNINGS

Errors: 0
Warnings: 1

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

---

ARCH-DI-002: Direct Handler instantiation (warning)\n\n```\nsrc/omnimarket/cli/cli_local.py:236: HandlerLocalDashboardServe().handle(

src/omnimarket/nodes/node_delegate_skill_orchestrator/ports/port_local_delegation_dispatch.py:965: self._effect_handler = effect_handler or HandlerLlmDelegationCall()
src/omnimarket/nodes/node_delegate_skill_orchestrator/ports/port_local_delegation_dispatch.py:966: self._projection_handler = projection_handler or HandlerProjectionDelegation(
src/omnimarket/nodes/node_delegate_skill_orchestrator/ports/port_local_delegation_dispatch.py:3289: HandlerProjectionLlmCost().project(call_event, self._evidence_db)
src/omnimarket/nodes/node_delegation_orchestrator/handlers/handler_delegation_workflow.py:1370: loop = HandlerComplianceLoop()
src/omnimarket/nodes/node_delegation_orchestrator/handlers/handler_delegation_workflow.py:2265: HandlerEscalationDecision()
src/omnimarket/nodes/node_local_dashboard_serve_effect/handlers/handler_local_dashboard_serve.py:226: handler=HandlerProjectionRead(topic_map=topics, row_source=source),\n```\n\nHandler classes should be wired only in workflow_runner.py or adapter modules. Direct instantiation elsewhere bypasses the contract system.\n\n

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

[github-actions]

✅ Hostile Reviewer — PASSED

Blocking findings (critical/error): 0
Total findings: 12
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 lane=deleg-quality-2-2a21 head=397b42253 rebase and retarget complete
Rebased onto dev (0c3748b) and retargeted to dev, new head 397b422
RESPONSE_CONTRACT at head is byte-identical to the one measured in the body
Measured against omnimarket 0.4.283@117b144 (dev at #3285), both on dev lane and in-process
Command: onex delegate --task-type code_generation --backend-id local-coder --response-contract
Observed: 12/20 before (dev lane), 12/20 before (in-process), 20/20 after (in-process), 9/20 control
Focused tests after rebase: 351 passed; ruff, ruff format and mypy --strict clean