[jonahgabriel]

OMN-20606

What

Operator, 2026-10-05: show whether a delegation is an escalation or a fallback, and why its parent failed. Measured on the h201 dev lane (ledger TERMINAL 2026-10-05T16:10:11Z lane=byok-routing-fix-9143): 93 of 94 failed delegations that carried a session were answered by a later in-process delegation within 120 s, and the answering row had a new correlation id, an empty caller_lane and nothing naming the failed parent.

Two defects, both fixed here:

1. Nothing carried lineage. delegation_events gains parent_correlation_id, lineage_kind (fallback or escalation) and parent_failure_cause (migration 0055, nullable TEXT, partial index on the parent). One spelling in models/delegation/delegation_lineage.py; a pure fold (HandlerDelegationLineageFold) returns all three columns or a named refusal, and both writers (sync HandlerProjectionDelegation and the deployed DelegationProjectionRunner) persist what it returns. A partial or malformed lineage stores none of it and never dead-letters the row; a lineage-less re-emit keeps stored lineage.
2. The in-process evidence terminal named no caller. On the in-process path port_local_delegation_dispatch._project_evidence builds the only terminal that reaches the bus (OMN-20154), and it never copied the request's caller_lane or ticket_id, although the handler's own terminal carries both. That is the cause of the empty caller_lane on every in-process fallback row. The handler now hands the port the validated caller lane, ticket and lineage (attribution, passed only when a request names any), and the port stamps them onto the evidence terminal.

The delegate-skill response tolerates and drops the three lineage keys before declaring them (OMN-18868 consumer-first order); the projection model declares and keeps them. Declaring them on the response and stamping them from the deployed handler is the second half, after a release carrying this one.

Dependency

Needs the omnibase_infra release carrying omnibase_infra#4592 (vendored migration 0055, and the dispatch port's attribution keyword this handler passes). Pin bump to that release goes on this branch once it is out.

Proof

• RED then GREEN: the two unit files fail with the changed sources checked out from origin/dev (10 failed) and pass on the branch.
  • uv run pytest tests/unit/projection/test_delegation_lineage_omn20606.py tests/unit/nodes/node_delegate_skill_orchestrator/test_delegate_skill_attribution_omn20606.py -q (32 passed)
  • Real Postgres twin, PG16: tests/test_omn20606_lineage_projection_real_postgres.py (3 passed: a fallback row joins to its failed parent, re-emit keeps lineage, malformed writes none, migration shape and index); positive control tests/test_omn20162_backend_id_and_host_real_postgres.py passes on the same database.
  • Neighbours: caller lane, backend_id/host, attempt ladder seam, runtime dispatch port, local evidence publish, relation inventory (136 passed).
• Lab readback on dev-202 (h202), 2026-10-05 ~17:31Z. The dev-202 delegation writer was hotpatched with this branch (originals verified identical to origin/dev, backed up, restored after), migration 0055 applied to dev-202 omnidash_analytics, and two in-process onex delegate runs made on h202 publishing to the dev-202 broker:
  • parent e850644a-5411-4e29-9e66-65f9831b4e9c forced to fail (one-token budget): row terminal_ok=f, caller_lane=deleg-fallback-mark-9143, ticket_id=OMN-20606 (both were empty on every in-process row before this change);
  • fallback 1333fb52-eb7f-4bf8-944d-1fe81623cc2e: row terminal_ok=t, parent_correlation_id=e850644a-..., lineage_kind=fallback, parent_failure_cause=quality_gate_refused, same caller lane and ticket.
  • The dev-202 tenant mirror had no row for the h202 install tenant, so its terminals dead-lettered (OMN-16831); the lane seeded that one mirror row from the h201 copy.
• RLS/reader paths: the writer change adds columns to an existing tenant-scoped upsert; the lab pass did not go through onex-api.

Not in this PR

• The deployed handler's own terminal carrying lineage (needs the response to declare the keys, after this release).
• A separate defect seen while proving: on the h201 dev lane the acceptance-judge request handler dead-letters every in-process delegate-skill-completed terminal (ModelDelegateSkillCompleted refuses actual_score, causation_id, emitted_at). Reported in the lane TERMINAL, not changed here.

Evidence-Ticket: OMN-20606
Evidence-Source: OCC#12962

[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 — DEGRADED

Blocking findings (critical/error): 0
Total findings: 0
Models succeeded: none

Action required: fewer than 2 reviewer models succeeded (infra unavailable or timeout). This blocks merge (OMN-15110) -- re-run once the reviewer endpoint is reachable/uncontended, or investigate.

---

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)

[onexbot-occ-writer]

OCC autobind rebound this PR's evidence-source stamp line to OCC#12962.

The receipt gate's own eligibility validator, run against the change-control tree the gate pins for OCC#12962, returned eligible for head 5ecc6467875b2bd36cd04d85de4c5c099de5693b. The body now carries exactly one stamp line. Displaced: OCC#12945.

Reported by occ_companion_emitter (OMN-18853).

[jonahgabriel]

Superseded: this branch has no changed files against dev; the OMN-20606 work already landed through #3472. Closing (found by lane cause-omnimarket-ci-summary).

---

Generated by Claude Code