[patrykkopycinski]

Live 400 on the claude/claude-opus-5 hop of best-reasoning-paid:
[400]: _omnirouteSkipContextRelay: Extra inputs are not permitted

contextHandoff.ts stamps _omnirouteSkipContextRelay / _omnirouteInternalRequest
onto internal handoff-summary bodies so the context-relay injection in chat.ts
skips its own request. stripInternalBodyFields() -- the single pre-serialization
chokepoint in base.ts -- was a hand-maintained allowlist of 5 unrelated markers,
so those keys survived into the serialized payload and strict Anthropic-compatible
gateways rejected the whole request.

Strip by prefix instead of per-key: every _omniroute* marker is internal by
construction and consumed by routing before dispatch, so no future marker can
become the next upstream 400. Client fields that merely start with _ are
untouched.

Tests: helper-level cases plus an end-to-end DefaultExecutor.execute() test that
asserts on the actual serialized upstream bytes. Mutation-verified: reverting the
fix turns 4 of 7 red, including the e2e case.

---

Complements #12735: that PR deletes the two currently-known handoff flags key-by-key; this one strips the whole _omniroute* prefix so future internal markers can't become the next upstream 400 (e.g. _omnirouteResponsesStore, already on this base, is caught here).

[diegosouzapw]

Verified the bug independently on the current tip and your prefix-based approach is the
right shape for this chokepoint — it also automatically covers _omnirouteResponsesStore
and any future _omniroute* marker, which the old allowlist would have missed again. Ran
your test file (7/7 pass, including the e2e byte-level assertion). Since you already flagged
this complements #12735: we're planning to land this one for the shared
cliFingerprints.ts chokepoint and ask #12735's author to rebase down to just their
4-executor defense-in-depth calls (dario/gitlab/glm/ninerouter bypass this chokepoint
entirely). One ask before merge: please add a changelog.d/fixes/ fragment referencing
#12729.

[patrykkopycinski]

Fragment added in fbb690ba9: changelog.d/fixes/13355-strip-internal-omniroute-markers.md, referencing #12729. Passes npm run check:changelog-integrity locally — the only fragment that gate flags is fixes/reset-aware-model-family.md (missing leading - , landed via #12637), i.e. pre-existing base debt untouched by this PR.

[patrykkopycinski]

CI attribution for the red fast-path on fbb690ba9 (and on be31d7233 before it) — pre-existing base heat, not this diff:

• fbb690ba9 adds only changelog.d/fixes/13355-*.md. The only CI consumer of changelog.d/ is the changelog-integrity gate, which passes on fbb690ba9 (it failed on be31d7233 — the fragment fixed that). Unit shards never read changelog.d/.
• be31d7233 (code only, no fragment) already failed Fast Quality Gates + unit shards on Sep 11 — with a different failing-test set (chatcore-hierarchical-admission, injection-guard-scan-bound-3932, glm-sse-transform-arity, tunnel-routes-error-sanitization) than today’s (paid-model-target-routes-6540 ×3, 200 !== 400). Same tree, disjoint failures ⇒ runner/base flakiness.
• Fast Quality Gates fails inside check-forgotten-sibling-tests.mjs (Invalid string length while tracing agyModels.ts → shared.ts → ~every call-log test) — the sibling-graph blowup also aborted the be31 run’s step-summary upload (1065k > 1024k limit). The last green release/v3.8.51 run predates the current base tip.

Vitest fast-path, Docs Gates, ESLint, Merge integrity, semgrep all pass on fbb690ba9. Happy to re-trigger if you want a fresh roll of the shards — repo precedent 226e6ee54.

[patrykkopycinski]

CI attribution (cross-posted to #13355 #13448 #13617 #13627 #13359): the red Fast Quality Gates on these heads are pre-existing base debt, visible on the base branch's own runs with no PR diff applied:

• mutation-test-coverage: run 34970965604 (base head 1e1af28) fails "2 covering unit test(s) across 2 module(s) are missing from stryker.conf.json" (incl. tests/unit/noauth-model-lockout.test.ts) — same signature on our PR runs.
• secrets ratchet: "REGRESSÃO — 1 secret findings > baseline 0" reproduces across PRs with disjoint diffs.
• forgotten-sibling-tests: "Invalid string length" (RangeError, string overflow) crashes the checker on the agyModels→shared→tests graph — hit 13355 and 13627 hours apart on unrelated diffs.
  Base tip FQG itself: last completed run 34970965604 = failure. Happy to PR the stryker testFiles additions if useful.