Repository navigation
[WRONG BRANCH] fix(errors): hop on Anthropic 400 "Third-party apps now draw from your extra usage" - #11
Merged
vadymhimself merged 1 commit intoNov 7, 2026
Conversation
…our extra usage" as insufficient_quota so combos hop Same spent-plan state as "out of extra usage" (#8), newer Anthropic wording. Without it the 400 classifies as invalid_request_error and a combo stops instead of hopping to the next tier. Co-Authored-By: Claude Code <noreply@anthropic.com>
|
✅ Deterministic PR hygiene checks passed. |
⏳ DRAFT
What to do
Its title has been prefixed with |
vadymhimself
marked this pull request as ready for review
November 7, 2026 20:30
vadymhimself
merged commit Nov 7, 2026
79faf06
into
release/2.65.0-vadymhimself.15
29 of 30 checks passed
vadymhimself
added a commit
that referenced
this pull request
Nov 14, 2026
) * fix(pool): rotate Anthropic pool accounts on out-of-extra-usage 400 Anthropic answers an account whose plan and extra usage are spent with HTTP 400 ("You're out of extra usage", "Third-party apps now draw from your extra usage"), not 429. Since #8/#11 src/lib/errors.ts classifies both as insufficient_quota, so a combo hops on them, but the Anthropic account pool still rotated on status 429 only (adapter-dispatch, adapter-continuation, and the web-search/image sidecar loops). One spent pool account therefore failed the request while the other pool accounts sat idle; on a hosted gateway, which has no combo, the caller got the error. Add src/oauth/anthropic-pool-rotation.ts: isPoolRotatableStatus() and peekAnthropicPoolRotation() decide the trigger ("429" or "quota-400"). A 400 counts only when the existing classifyError() says insufficient_quota, so no message wording is duplicated, and any other 400 (a malformed request) is passed through untouched. The 400 body is read with the bounded peek the key-pool 429 path already uses (readQuotaResetAt, which now also returns the peeked text), and the returned response replays it. All three rotation sites use the helper. The sidecar hook moves only the Anthropic pool on a quota-400; key pools and the generic OAuth roster still rotate on 429 alone. rotateAnthropicAccountOn429() takes the trigger. On a quota-400 it always cools the account (the 400 is quota evidence even if the unified status header still reads "allowed"), using the same Retry-After, then rejected-window reset, as a quota 429. With neither, it falls back to 10 minutes, the cooldown the combo layer gives the same insufficient_quota code, not the 60s rate-limit default that would send the next request back to the spent account a minute later. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(pool): anchor the quota-400 match to Anthropic's envelope and exact phrases Security review (Mars, RED R1): isQuotaExhaustedBadRequest ran the whole 400 message through classifyError, which matches quota words anywhere in the text. Anthropic's validation 400s echo client values back (tool ids, model names, prompt text), so one crafted request such as a tool_use id "toolu_insufficient_quota" could cool every pool account for 10 minutes. The 400 now counts only when the body is Anthropic's JSON error envelope (type "error", error.type "invalid_request_error") and error.message, after leading whitespace, starts (case-insensitive) with "You're out of extra usage" (straight or curly apostrophe) or "Third-party apps now draw from your extra usage". A non-JSON body, another error type, a missing envelope, or the wording anywhere but the start never rotates. classifyError is no longer used here; combo classification is unchanged. R2 (REAL-PAYLOAD-RATCHET): the third-party test payload is now the real recorded Anthropic 400 (staging gateway log line, 2026-11-11), copied byte-identical from AgentRadarLabs/agentradar-paperclip fixtures/external/anthropic-extra-usage-thirdparty-400.txt @ 615ec8db69 into fixtures/external/anthropic/ with a README saying where it came from. No real "You're out of extra usage" body has been recorded anywhere (paperclip marks its sample "derived"), so that wording stays a typed test string. New must-fail tests (5 fail on the previous commit): a 400 echoing toolu_insufficient_quota neither rotates nor changes on its way to the caller; the wording mid-message does not rotate; the right wording in a wrong envelope (rate_limit_error, api_error, missing type "error", non-string message, array body) does not rotate; raw non-JSON text does not rotate. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * chore(release): 2.65.0-vadymhimself.22 The fix commits on this branch inherited 2.65.0-vadymhimself.21, which names an already-published tag, so release-version-line failed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Gauge <gauge@agentradar-house.staging.agentradar.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
vadymhimself
pushed a commit
that referenced
this pull request
Nov 16, 2026
…ites (lidge-jun#8590) * devin: opt-in Cognition blocklist rewrite on history data surfaces The blocklist sanitizer only touched instruction surfaces (tool descriptions, the #2 system prompt); data fields stayed byte-exact, so one quoted trigger phrase anywhere in a long thread history made every later request fail with upstream permission_denied. New OPENCODEX_DEVIN_SANITIZE_DATA_SURFACES=1 opt-in extends the same rewrites to message text (#3), tool-call arguments (#6.3), image captions (#10.3) and replayed thinking (#11). Identifiers (tool name, tool_call_id) and opaque signature blobs stay byte-exact. Off by default: rewriting literal history changes what the model sees, so refusing stays the default failure. The permission_denied hint now names the opt-in. Tests pin both modes: data literal by default, rewritten when enabled. * devin: add Codex CLI boilerplate phrases to Cognition blocklist rewrites Two more live-isolated triggers in the Codex CLI developer boilerplate: the identity sentence "Codex refers to the open-source agentic coding interface" and the "Don not output ANSI escape codes directly — the CLI renderer applies them" bullet. Both are rewritten meaning-preservingly like the existing entries. Also adds the dump/bisect probe scripts used to isolate the triggers. * fix(devin): drop stale signatures after thinking sanitization Keep reasoning signatures only when the serialized thinking is unchanged. Preserve signature-only replay and default-off sanitization, and add wire, authentication and continuity regressions with matching adapter docs. * chore(devin): keep live credential probes out of the runtime PR
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fork #8 covers "You're out of extra usage" (src/lib/errors.ts:434). Anthropic now also answers 400 "Third-party apps now draw from your extra usage, not your plan limits", which still classifies as invalid_request_error, so comboFailureDecision returns stop (src/combos/failover.ts:786) and the House waterfall never reaches its GPT tier.
Change: one more text.includes in the insufficient_quota branch of classifyError; exact-text asserts in tests/server/error-fidelity.test.ts and tests/codex-integration/combos.test.ts.
Verified: bun test on both files 102 pass / 0 fail; with the src match reverted, the 2 new asserts fail (2 fail); tsc --noEmit clean.
package.json is already 2.65.0-vadymhimself.16, so merging this commit and tagging v2.65.0-vadymhimself.16 on it needs no bump.
Eval added/updated: N/A deterministic (unit test: tests/server/error-fidelity.test.ts, tests/codex-integration/combos.test.ts)
🤖 Generated with Claude Code