OmniRoute rebuild 2026-09-30: decision record, evidence artifacts, owner-record script, unit templates (records what runs; compression on 20128 not applied) - #530

Scope

• What this PR changes: records the 2026-09-30 rebuild of both OmniRoute gateways on the NativeStack workstation and touches no host: upstream release/v3.8.51 at 2f42a9ac1 (upstream tagged v3.8.51 later that day with the identical tree, and published [email protected] to npm at 02:54Z; the record says so) with the open upstream PR #13788 on both gateways and a local affinity patch on 20128, the announced Codex version 0.159.1 (gpt-6.1-sol answers 200 on both gateways), the token-save settings (a nine-step plan and one delta: on 20129 the master switch is on and ten of the twelve shipped engines run at upstream's per-engine settings, which is not upstream's shipped state), the read-only owner-record script for Gate A (tools/omniroute/gateway_record.py), the two values-free unit templates with a unit test, and the checks and probes behind each claim. The decision record states every departure from upstream's shipped defaults, the provenance of the operation scripts, the stored combos (the header allow-lossy selects a peer-owned combo; headroom still runs where a combo names it) and the residuals, notably that lite folds whitespace in code-bearing text and that neither the headerless lane nor the gpt6-safe-lossy header lane passes the user's condition on real Codex traffic, so compression on 20128 is not applied.
• Not in this PR: the omniroute row of manifests/stack.json and the component pin (a shared hot file; its text change is the separate lane:shared PR Stack manifest: the omniroute row describes the 2026-09-30 builds (held for the trading lane's acknowledgement) #531, held for the trading lane's acknowledgement). Until it merges, this record and docs/foundation-stack.md are the current statement of the running builds.
• Base commit: be91e1f8 (main, after SocratiCode 1.16.0: qualification receipt and dated decision update (the Linux pin stays 1.15.0) #523). Main has since moved to 8f7086a2 (PyJWT relock record: redis decode search (D1), timing range from the outputs (D2), honest output labels #529); a test merge of this head onto it is clean (git merge-tree, exit 0).
• Lane: lane:foundation. It changes no path owned by the trading lane and no shared hot file's text: manifests/evidence.json is only re-registered through the hot-file protocol (last commit), which docs/lanes.md says does not by itself make a PR shared.
• Owned paths touched: docs/decisions/, docs/foundation-stack.md, evidence/artifacts/omniroute-rebuild-20260930/, tools/omniroute/, adoption/templates/systemd/, tests/test_omniroute_gateway_unit.py, manifests/evidence.json.

SOTA sources

• OmniRoute release/v3.8.51 at 2f42a9ac19d1a247ec9ce5473b790843724b3061 (it carries #14812 8183e9b79, #14886 7f5de7ba6, #15119, #15128, #15145 and #15147); carried: #13788 as its two upstream commits 24bbadbad and 6c7990058 (stable patch-ids equal), and the local non-upstream affinity patch 045aa81f3 (references in its commit message: CLIProxyAPI v8.0.2 4a2c818, LiteLLM v1.102.1 d09bbae, sub2api v0.2.8 fd80b08). release/v3.8.52 (a1a2dce1a) has no gpt-6.1 entry either.
• Source read at 2f42a9ac1 for the settings claims: open-sse/services/compression/types.ts:421-443, lossyRequestPolicy.ts:19, planResolution.ts:56-57, strategySelector.ts:812-816, src/lib/db/compressionCombos.ts:33-39, docs/compression/COMPRESSION_GUIDE.md:226-239, engineCatalog.ts, lite.ts:50-53, bodyAdapter.ts:154-201, src/shared/constants/featureFlagDefinitions.ts:564-570, open-sse/config/codexClient.ts, src/app/api/providers/[id]/models/discovery/codex.ts:82.
• Upstream's own scripts on our builds: check:pack-artifact, check:pack-boot, test:unit (unchanged upstream checks, kept apart from our integration checks in the receipt).
• The live model catalog (cx/gpt-6.1-sol, context 872000) is the only source for gpt-6.1-sol: the static registries of both upstream branches lack it.
• Precedent: docs/decisions/2026-09-27-omniroute-account-pool.md and its evidence package.

Evidence-class table

Local commands run

$ python3 -B -m unittest tests.test_foundation_stack_pins tests.test_omniroute_gateway_unit tests.test_adoption_contract tests.test_freeze_snapshot tests.test_catalogs tests.test_stack_lifecycle
Ran 268 tests ... OK                                                                                exit 0
$ python3 scripts/validate.py
{"components": 69, "hashed_files": 8297, "profiles": 4, "receipts": 174, "status": "passed"}        exit 0
$ python3 scripts/evidence_manifest.py --check
{"files": 8297, "status": "passed"}                                                                  exit 0
$ python3 tools/omniroute/gateway_record.py          (read-only; compared with the published record)
31 of 32 paths equal, only taken_utc differs                                                         exit 0

Head 20799c05: the three timestamps that no command printed (installed check, upstream release check, full-suite reconciliation entry) were replaced by printed ones, and both checks' outputs are published as returned (checks/installed-check-20260930.json, checks/upstream-release-check-20260930.json). The 268 tests, validate.py and evidence_manifest.py --check above were re-run on that head.

The full unittest discovery runs in CI's validate check.

Decision record

docs/decisions/2026-09-30-omniroute-rebuild.md (alternatives considered, the comparison that would overturn each choice, rollback, limitations and residuals).

Host evidence

This PR adds artifacts under evidence/artifacts/, not evidence/hosts/.

• scripts/validate.py (exit 0 above) validates the registered artifacts.
• Independent review: two Opus evidence-reviewers read this work in full (the first: 1 high, 5 medium, lows; the second, on the fixes: 1 high, 3 medium, 11 low); every finding is fixed here.
• No platform_status change.

Residuals

• Upstream released 3.8.51 after the rebuild: the tag commit c1e30b767 has the same tree (0f58d8df...) as the build base 2f42a9ac1, so no rebuild is needed for the tree; #13788 is still open upstream ([defer]), so the carry stays. The repo's component pin (npm 3.8.50) is unchanged here and is handled in Stack manifest: the omniroute row describes the 2026-09-30 builds (held for the trading lane's acknowledgement) #531.
• The applier's SECRET_GET guard names fewer routes than session 88's read of the gateway's anonymous surface (/api/settings/cache-config, /api/sync/*, /api/cli-tools/keys); this plan never reads them (recorded as a residual; the surface itself is 88's decision with the Gate A owner).
• The applier that ran T01-T09 is not retained, the published scripts are a stopped workflow stage's later text, and T10 ran through its unreviewed edit; the read-backs are the evidence.
• lite folds whitespace in code-bearing text; the pi-practice session's real-traffic measurement (reported, not reproduced) and our offline probe of eight real request bodies both show the headerless lane compresses nothing on real Codex traffic, so compression on 20128 stays off and is an open user decision.
• headroom is off in the engines map but runs in three stored combos for a caller that names them; the header allow-lossy selects a peer-owned combo. Narrowing either is a gateway write for the Gate A owner's timing and the peers' consent.
• blockedProviders and noAuthFallbackDisabledProviders are not expressible in the apply contract (open user decision).
• No live rollback has been run; the store copy loses everything written since. Upstream's e2e and integration suites were not run; max effort clamps to xhigh for gpt-6.1-sol.

Checklist

• No workflows or Actions are changed.
• No secrets are printed, logged or committed.
• No new paid hosting, subscription or billing surface.
• Peer-owned untracked files and worktrees were preserved.

🤖 Generated with Claude Code