GPT-6 gateway lane on the all-engines OmniRoute instance: static headers and a one-slash model

[seathatflowsinourveins]

Scope

• What this PR changes: the landscape-sweep GPT-6 lane can target the framework OmniRoute instance on 127.0.0.1:20129.
  • That instance has all engines and output styles on and is chained through node sharedgw to the shared gateway on 20128.
  • The shared 20128 stays unchanged for the memory services and the other sessions.
• Base commit: 7fa3849c (Pre-push registry gate: run the three registry tests on every pushed commit #438). Reviewed as 55fc8d17..854803b0; the rebase changed no reviewed line, and docs/harness-defaults.md keeps Pre-push registry gate: run the three registry tests on every pushed commit #438's row beside this PR's two rows.
• Head: 74c233d7
• Lane: lane:foundation
• Paths (6 files):
  • tools/sota-convergence/landscape-sweep/{build_args.py,codex_job.py,README.md}
  • tests/test_landscape_sweep_harness.py
  • docs/harness-defaults.md (two anti-pattern rows)
  • manifests/evidence.json (registration)
  • No generated report changes.

Changes

• Static provider headers. build_args.py --omniroute-header NAME=VALUE stages Codex's model_providers.<id>.http_headers, for example x-omniroute-compression=allow-lossy.
  • Only OmniRoute 3.8.51's four per-request switches are accepted: x-omniroute-compression, -no-cache, -no-memory and -strip-reasoning.
  • Other x-omniroute-* headers can carry a secret under a name with no credential word, so a word filter cannot keep them out: x-omniroute-self-hop, -video-bridge-broker and -lease-owner, for example. Staged values are recorded in staged.json and in each job's inputs.json.
  • A refused name or value is never echoed.
  • codex_job.py applies the same allowlist to staged.json. It checks the lane home's table against it and records the headers per job.
  • The runner refuses staged headers when tomllib is missing (Python 3.9 and 3.10). A header-less lane still runs on 3.9. The builder's read-back is unconditional, because staging already needs 3.11.
• The model has at most one provider segment, of letters, digits, _ and -. On 20129 it is sharedgw/gpt-6-astra-max.
  • Codex rust-v0.157.1 strips one namespace, and only one of those characters, when it looks up model metadata (codex-rs/models-manager/src/manager.rs L763-780).
  • Any other slug gets generic fallback metadata (model_info.rs L99-150): a different prompt template, no Responses Lite, no multi-agent, a 272k context window and a 10,000-byte tool-output truncation. That covers sharedgw/cx/gpt-6-astra and my.gw/gpt-6-astra-max.
  • main already refused a second slash. This PR adds the fallback-metadata diagnostic and the namespace character rule to both entry points.
• README:
  • Manual commands must pass -m sharedgw/gpt-6-astra-max, because under -p stack-worker the profile's model otherwise wins.
  • The parity check compares the rendered items' content, without id and create_time and with the model masked. Its commands were run as written (evidence below).
  • Slashless aliases: the 401 observation is kept, without a mechanism, since none was established from source.
  • Effort evidence:
    • A gateway's outbound request is pipelinePayloads.providerRequest in its detailed call log, when detailed pipeline logging is on.
    • requestBody is the body that gateway received.
    • A non-null reasoning_effort_upstream is positive evidence; only a null effort column proves nothing.
• Anti-pattern rows:
  • The two-slash slug on a chained gateway, checked through both real entry points.
  • A null call_logs effort column read as "effort not sent", and the first correction's inbound field. No repository check enforces this; the row says so.

SOTA sources

• openai/codex rust-v0.157.1:
  • codex-rs/models-manager/src/manager.rs L763-780: the namespace rule. model_info.rs L99-150: the fallback metadata.
  • codex-rs/core/src/prompt_debug.rs: prompt-input.
  • codex-rs/model-provider-info/src/lib.rs:
    • L166-168: http_headers, whose values are a RedactedString.
    • L385-412: build_header_map skips an invalid name or value.
    • L135: deny_unknown_fields is a JSON-schema attribute only.
• OmniRoute 3.8.51 (build dd6e9607e), installed source:
  • The switches:
    • open-sse/handlers/chatCore/headers.ts: L18-33 no-memory, L35-46 compression, L48-65 strip-reasoning.
    • src/lib/semanticCache.ts L483-486 and L501-504: no-cache.
    • open-sse/services/compression/lossyRequestPolicy.ts L29-40: allow-lossy.
  • The secret-carrying headers:
    • open-sse/utils/selfHop.ts L1-12
    • src/lib/guardrails/videoBridgeBrokerAuth.ts L7-19
    • open-sse/utils/requestLogger.ts L100 and L109-111: lease-owner is redacted.
  • Effort evidence:
    • open-sse/handlers/chatCore/attemptLogging.ts L513-516: providerRequest into pipelinePayloads. L575-582: requestBody from the client body.
    • src/lib/usage/callLogs.ts: L646-653 for the effort columns; L1047-1053 for the detail route.
    • src/lib/db/detailedLogs.ts L56-64: detailed logging.
    • open-sse/executors/codex.ts L1451-1465: the effort rewrite.

Evidence-class table

Claim	Evidence class	Command / result
The allowlist, the namespace rule, real-entry-point refusals and the tomllib refusal	local integration test, fail-first	The final tests on the pre-repair code (d19e3a69, the rebased 854803b0): FAILED (failures=6, errors=1). After the repair: tests.test_landscape_sweep_harness Ran 122 tests ... OK (skipped=3), again on the rebased head
Each new guard is needed	local mutation check	10 named mutations, each targeted test run, file restored and sha256-checked: 10 caught. They cover each allowlist, each namespace check, MODEL_NAME, the value and type checks, the tomllib refusal, a job directory created before settings, and an echoed refused name
The one-slash model renders the same prompt as the control	native, no model call (this round)	The README's parity block as written, on a lane home staged by this branch for 20129. Both prompt-input commands exit 0 with 4 items each (3 developer, 1 user). cmp of the normalized files exits 0; neither output contains the model slug. Earlier coordinator run (installed profile, prompt argument): 5 against 5 items, identical after the same normalization, and the negative control sharedgw/cx/gpt-6-astra differed (3 items)
20129 serves the one-slash model at max	historical live host observation (2026-09-27, this host)	A live call returned 200 with reasoning. A probe without -m sent the profile's gpt-6-astra and got 401 six times, which is why -m is required
Registry tests, validation and the pre-push gate	local and native git push	57 registry tests OK; scripts/validate.py passed after re-registration per docs/lanes.md. This push went through #438's gate: pre-push: running the registry tests on 74c233d7…, then OK

Review round (review-439; one review, one repair)

#	Finding	Disposition
1	Credential-carrying OmniRoute headers passed the word filter	Fixed: allowlist of four switches in both entry points. Negative tests for self-hop, video-bridge-broker, lease-owner and session-id, through the builder, settings and codex_call.sh start
2	requestBody is inbound, not outbound	Fixed after checking the cited lines: README and row 88 name pipelinePayloads.providerRequest and the detailed-logging condition. The column rule is narrowed
3	A . passed in the provider segment	Fixed: the namespace rule and diagnostic in both entry points; my.gw/gpt-6-astra-max is tested
4	The no-partial-state test called the pure settings()	Fixed: it runs codex_call.sh start (mutation M8 is caught)
5	"Failed first" was not reproducible from git	Fixed: row 87 says main already refused a second slash, and the test failed first on the diagnostic and the dotted namespace
6	The README stated tomllib checks as unconditional	Fixed: the runner fails closed with staged headers, and the README states the 3.9 behaviour
7	The parity check compared counts; its commands had no recorded run	Fixed: content comparison, run as written (above)
8	The alias mechanism was unsourced	Fixed: the mechanism is removed and the observation kept
9	Body errors (NAME:VALUE, base, correlation-id citation, paths)	Fixed in this body

Residual risks

• Without tomllib, a header-less lane still skips the comparison, so a hand-edited lane home that carries headers goes unchecked on Python 3.9.
• The builder's read-back refusal (a table that does not parse back) has no test.
• Headers stop at 20129. The effect of the lost Codex identity headers on 20128's cache hits is not measured.
• A lane home under the temporary directory makes Codex 0.157.1 skip its apply_patch and sandbox PATH aliases (codex-rs/arg0/src/lib.rs L344-357). The parity run printed that warning. On Linux the sandbox falls back to the codex executable (L281-291), and the effect on shell-invoked apply_patch is not measured. This is a follow-up; it is out of scope for this round.
• CI: validate-macos failed once on an unrelated codex_lane interruption-timing test, and the rerun passed. It is recorded as a flake to watch.
• Harness acceptance on 20129 is still open (Status below).

Status

• Built by: GPT-6 (gpt-6-astra, effort max), in two rounds, plus this coordinator repair round after review-439. The coordinator verified and committed.
• Next, after merge:
  1. Restage the gateway lane home, which still holds the old two-slash model.
  2. Run a canary builder job on 20129 against a 20128 control, plus a multi-turn cache check. Record outbound effort from 20128's pipelinePayloads.providerRequest.
  3. Send the switch-on UTC time to peers for the evidence window.

🤖 Generated with Claude Code