fix(agents): classify Zhipu GLM overload as overloaded for failover by 0xghost42 · Pull Request #93241 · openclaw/openclaw

0xghost42 · 2026-06-15T09:59:19Z

Summary

Model fallback (agents.defaults.model.fallbacks) did not recover correctly when Zhipu (GLM) returned its overload error. Zhipu signals overload as HTTP 200 with the error in the body:

[1305][该模型当前访问量过大，请您稍后再试]

The Anthropic-compatible client surfaces that body as a thrown error, which runWithModelFallback normalizes via coerceToFailoverError → isOverloadedErrorMessage. There was no pattern for this phrase, so the overload was not recognized as a failover reason.

This adds the distinctive overload phrase 访问量过大 ("traffic too high") to ERROR_PATTERNS.overloaded, next to the existing Chinese overload markers (服务过载, 当前负载过高). The error now classifies as overloaded — the failover reason the issue asks for — so the configured fallback is reached with the correct, overload-specific recovery semantics.

Closes #93211

Verification

Real behavior proof

Behavior addressed: Zhipu/GLM overload body [1305][该模型当前访问量过大，请您稍后再试] was not recognized as overloaded, so failover did not key on the provider-overload contract.
Real environment tested: drove the real runWithModelFallback orchestrator (the function the issue names as mishandling the error) in this checkout, before and after the patch, injecting the exact reported error at the throw point the Anthropic-compatible client surfaces it. No live Zhipu endpoint was available (no API key), so the [1305] body is reproduced at that runtime boundary rather than over the wire.
Exact steps or command run after this patch:

# orchestrator-level (real runWithModelFallback):
node scripts/run-vitest.mjs src/agents/model-fallback.test.ts -t "Zhipu"
# matcher-level:
node scripts/run-vitest.mjs src/agents/embedded-agent-helpers/failover-matches.test.ts

Evidence after fix: ran the orchestrator test with the overload pattern removed, then restored:

# BEFORE (pattern removed) — runWithModelFallback with the GLM body:
result.attempts[0].reason === "unknown"     # not recognized as overload

# AFTER (this patch):
result.attempts[0].reason === "overloaded"  # recognized; recovery keyed on overload
run called twice; fallback run === ["anthropic", "claude-haiku-3-5"]

Observed result after fix: with the patch the GLM overload error resolves to overloaded inside runWithModelFallback and the configured fallback model is attempted. For full honesty about current main: an unrecognized error already falls back generically with reason unknown when candidates remain, so the symptom from the 2026.6.5 report (no fallback at all) is partly masked on main; this patch supplies the correct classification (overloaded), which is what the issue requests and what carries the proper provider-overload retry/cooldown handling rather than the generic unknown path.
What was not tested: end-to-end delivery against the live Zhipu Anthropic-compatible endpoint (https://open.bigmodel.cn/api/anthropic) — no credentials. The orchestrator classification + fallback decision is the failover decision point, exercised above.

Coverage: failover-matches.test.ts → 25 passed (GLM body pinned as overloaded, not rate-limit/auth); model-fallback.test.ts → 80 passed (new orchestrator case asserts the GLM overload reaches the configured fallback keyed on overloaded).

clawsweeper · 2026-06-15T10:00:57Z

Codex review: needs real behavior proof before merge. Reviewed June 20, 2026, 11:47 PM ET / 03:47 UTC.

Summary
The PR adds 访问量过大 to the shared overload matcher and adds matcher plus runWithModelFallback regression tests for the reported Zhipu GLM 1305 body.

PR surface: Source +1, Tests +31. Total +32 across 3 files.

Reproducibility: no. high-confidence live reproduction is present. Source inspection shows current main and v2026.6.9 lack the reported overload phrase, and the PR tests exercise the injected thrown-error boundary, but no real Z.AI endpoint or redacted OpenClaw fallback run was provided.

Review metrics: 1 noteworthy metric.

Shared failover matcher change: 1 generic overload phrase added. The matcher is shared across providers, so this small source change can alter fallback reason and cooldown behavior beyond the reported Z.AI route.

Root-cause cluster
Relationship: fixed_by_candidate
Canonical: #93211
Summary: This PR is the active candidate fix for the canonical Zhipu GLM 1305 overload classification issue; sibling PRs targeted the same root cause but are closed unmerged.

Members:

canonical: Model fallback not triggered for Zhipu (GLM) error code 1305 — overloaded pattern mismatch #93211 - The PR body uses closing syntax for this open issue and implements the reported missing overload phrase classification.
superseded: fix(failover): classify Zhipu (GLM) error [1305] as overloaded #93212 - This closed sibling targeted the same issue but kept a broader shared numeric matcher and was closed unmerged.
superseded: fix(agents): classify Zhipu GLM error 1305 / 访问量过大 as overloaded for model fallback #93669 - This closed sibling targeted the same payload and was closed unmerged after maintainer feedback about ownership boundary and coverage.

Proposal only: this assessment does not dispatch repair, suppress jobs, mutate sibling items, close, or merge anything.

Merge readiness
Overall: 🧂 unranked krab
Proof: 🧂 unranked krab
Patch quality: 🐚 platinum hermit
Result: blocked until real behavior proof from a real setup is added.

Overall follows the weaker of proof and patch quality, so missing proof can cap an otherwise strong patch.

Rank-up moves:

[P2] Add redacted terminal output, copied live output, logs, or a recording showing the reported Z.AI/OpenClaw overload body reaches the configured fallback after this patch; redact IP addresses, API keys, phone numbers, and non-public endpoints before posting.

Proof guidance:

[P1] Needs real behavior proof before merge: The PR supplies structured injected-error Vitest proof, but no redacted terminal output, logs, live output, linked artifact, or recording from a real Z.AI/OpenClaw fallback run. After adding proof, update the PR body; ClawSweeper should re-review automatically. If it does not, the PR author or someone with repository write access can comment @clawsweeper re-review.

Risk before merge

[P2] The PR still needs after-fix proof from a real setup; the current evidence is injected-error Vitest/orchestrator proof, not redacted terminal output, logs, live output, or a recording from a Z.AI/OpenClaw fallback run.
[P2] Adding a shared overload phrase changes fallback reason, cooldown, and provider/model routing behavior for any provider that emits the same Chinese high-traffic wording, not only Zhipu GLM.

Maintainer options:

Require Redacted Real Fallback Proof (recommended)
Ask for terminal output, copied live output, logs, or a recording showing the reported Z.AI/OpenClaw overload body reaches the configured fallback after this patch.
Accept Provider-Neutral Phrase Matching
Maintainers may intentionally accept that this Chinese high-traffic phrase should classify as overloaded for all providers that emit it.
Move Classification To Z.AI Hook
If shared phrase matching is too broad, move the recognition into the Z.AI provider classifyFailoverReason hook and keep regression coverage on that boundary.

Next step before merge

[P1] Human review remains needed for the real behavior proof gate and for maintainer acceptance of provider-neutral overload phrase matching.

Security
Cleared: The diff changes TypeScript classifier text and tests only; it adds no dependency, workflow, secret, package, or code-execution surface.

Review details

Best possible solution:

Land one canonical classifier fix once maintainers accept provider-neutral phrase matching and the PR has redacted real fallback proof or an explicit proof waiver; keep the linked issue open until merge.

Do we have a high-confidence way to reproduce the issue?

No high-confidence live reproduction is present. Source inspection shows current main and v2026.6.9 lack the reported overload phrase, and the PR tests exercise the injected thrown-error boundary, but no real Z.AI endpoint or redacted OpenClaw fallback run was provided.

Is this the best way to solve the issue?

Likely yes, subject to maintainer acceptance and proof: the latest head avoids the rejected bare numeric matcher and uses the existing shared Chinese overload bucket. If maintainers want provider-specific scoping, the safer alternative is a Z.AI provider classifyFailoverReason hook.

AGENTS.md: found and applied where relevant.

Codex review notes: model internal, reasoning high; reviewed against 9750d887f502.

Label changes

Label justifications:

P2: The PR targets a normal provider/model fallback bug with limited but real Zhipu GLM user impact.
merge-risk: 🚨 compatibility: Changing shared failover classification can alter existing users' fallback and cooldown behavior during provider errors.
merge-risk: 🚨 auth-provider: The change affects provider/model fallback routing and auth-profile cooldown handling for configured model chains.
rating: 🧂 unranked krab: Overall readiness is 🧂 unranked krab; proof is 🧂 unranked krab and patch quality is 🐚 platinum hermit.
status: 📣 needs proof: The PR needs real behavior proof before ClawSweeper can clear the contributor ask. Needs real behavior proof before merge: The PR supplies structured injected-error Vitest proof, but no redacted terminal output, logs, live output, linked artifact, or recording from a real Z.AI/OpenClaw fallback run. After adding proof, update the PR body; ClawSweeper should re-review automatically. If it does not, the PR author or someone with repository write access can comment @clawsweeper re-review.

Evidence reviewed

PR surface:

Source +1, Tests +31. Total +32 across 3 files.

View PR surface stats

Area	Files	Added	Net
Source	1	1	+1
Tests	2	31	+31
Docs	0	0	0
Config	0	0	0
Generated	0	0	0
Other	0	0	0
Total	3	32	+32

What I checked:

Repository policy read: Root AGENTS.md and scoped src/agents/AGENTS.md were read fully; provider/model fallback behavior is compatibility-sensitive, and the scoped agent guide applied without conflicting instructions. (AGENTS.md:1, 9750d887f502)
Current main lacks the requested phrase: The current overload matcher includes Chinese markers 服务过载 and 当前负载过高, but not 访问量过大, so main does not already implement the central change. (src/agents/embedded-agent-helpers/failover-matches.ts:97, 9750d887f502)
PR head adds the phrase only: The latest PR head adds 访问量过大 to the shared Chinese provider overload bucket and does not add the rejected bare numeric [1305] shared matcher. (src/agents/embedded-agent-helpers/failover-matches.ts:110, db79e94c89dc)
Classification path: classifyFailoverClassificationFromMessage returns overloaded when isOverloadedErrorMessage matches, so the added phrase changes failover reason classification. (src/agents/embedded-agent-helpers/errors.ts:1033, 9750d887f502)
Fallback behavior contract: The model failover docs say overloaded errors move through the overloaded bucket and can advance to configured fallbacks after profile rotation, making this classification user-visible. Public docs: docs/concepts/model-failover.md. (docs/concepts/model-failover.md:276, 9750d887f502)
PR regression coverage: The PR adds an injected-error runWithModelFallback test asserting the reported body falls back and records reason overloaded, plus matcher tests that keep it out of rate-limit/auth buckets. (src/agents/model-fallback.test.ts:749, db79e94c89dc)

Likely related people:

vincentkoc: Current blame attributes the shared overload matcher to Vincent Koc, and a repository-member comment on the closed sibling PR gave ownership-boundary guidance for this exact root cause. (role: recent area contributor and reviewer; confidence: high; commits: 3968fea383b2; files: src/agents/embedded-agent-helpers/failover-matches.ts, src/plugins/provider-runtime.ts, extensions/zai/index.ts)
steipete: Recent history added GLM-5.2 support in the bundled Z.AI provider, and the latest PR head commit keeps this overload matching provider-neutral. (role: recent Z.AI and PR boundary contributor; confidence: high; commits: 735f59af7366, db79e94c89dc; files: extensions/zai/index.ts, src/agents/embedded-agent-helpers/failover-matches.ts, src/agents/model-fallback.test.ts)
Jason (Json): Commit history shows this author added structured provider error signals to hooks, which is the adjacent provider-owned classification boundary considered during review. (role: provider hook boundary author; confidence: medium; commits: 89975eea24a5; files: src/plugins/provider-runtime.ts, src/plugins/types.ts)
Mason Huang: Recent history includes provider quota/error classifier work in the same shared failover matcher area. (role: recent failover classifier contributor; confidence: medium; commits: 439dcbde3b1f; files: src/agents/embedded-agent-helpers/failover-matches.ts)

What the crustacean ranks mean

🦀 challenger crab: rare, exceptional readiness with strong proof, clean implementation, and convincing validation.
🦞 diamond lobster: very strong readiness with only minor maintainer review expected.
🐚 platinum hermit: good normal PR, likely mergeable with ordinary maintainer review.
🦐 gold shrimp: useful signal, but proof or patch confidence is still limited.
🦪 silver shellfish: thin signal; proof, validation, or implementation needs work.
🧂 unranked krab: not merge-ready because proof is missing/unusable or there are serious correctness or safety concerns.
🌊 off-meta tidepool: rating does not apply to this item.

Shiny media proof means a screenshot, video, or linked artifact directly shows the changed behavior. Runtime, network, CSP, and security claims still need visible diagnostics.

How this review workflow works

ClawSweeper keeps one durable marker-backed review comment per issue or PR.
Re-runs edit this comment so the latest verdict, findings, and automation markers stay together instead of adding duplicate bot comments.
A fresh review can be triggered by eligible @clawsweeper re-review comments, exact-item GitHub events, scheduled/background review runs, or manual workflow dispatch.
PR/issue authors and users with repository write access can comment @clawsweeper re-review or @clawsweeper re-run on an open PR or issue to request a fresh review only.
Maintainers can also comment @clawsweeper review to request a fresh review only.
Fresh-review commands do not start repair, autofix, rebase, CI repair, or automerge.
Maintainer-only repair and merge flows require explicit commands such as @clawsweeper autofix, @clawsweeper automerge, @clawsweeper fix ci, or @clawsweeper address review.
Maintainers can comment @clawsweeper explain to ask for more context, or @clawsweeper stop to stop active automation.

0xghost42 · 2026-06-16T11:21:01Z

Strengthened the proof to the orchestrator level. Added a runWithModelFallback test that drives the real fallback path with the reported GLM body [1305][该模型当前访问量过大，请您稍后再试]:

with the overload pattern removed, the run classifies the error as reason: "unknown"
with this patch, it classifies as reason: "overloaded" and the configured fallback model is attempted (run called twice; fallback run ["anthropic", "claude-haiku-3-5"])

Commands and before/after are in the updated PR description. model-fallback.test.ts → 80 passed, failover-matches.test.ts → 25 passed. No live Zhipu endpoint (no credentials), so the [1305] body is injected at the throw boundary the Anthropic-compatible client surfaces it; the orchestrator classification + fallback decision is the failover decision point.

On the same-issue overlap with #93212: this branch changes only the message-pattern list (ERROR_PATTERNS.overloaded) and leaves the shared PROVIDER_SPECIFIC_PATTERNS untouched, so it does not add a bare numeric [1305] matcher that could classify unrelated providers' payloads as overloaded. Happy to defer to whichever branch maintainers prefer; flagging the scope difference so the choice is easy.

@clawsweeper re-review

clawsweeper · 2026-06-16T11:21:03Z

🦞🧹
ClawSweeper re-review requested.

I asked ClawSweeper to review this item again.
Action: item re-review queued (workflow sweep.yml, event repository_dispatch).
Result: the existing ClawSweeper review comment will be edited in place when the review finishes.

Re-review progress:

State: Failed
Detail: The targeted re-review timed out or failed while waiting for exact-review Codex capacity. Retry after capacity recovers.
Run: https://github.com/openclaw/clawsweeper/actions/runs/27613955361
Updated: 2026-06-16T11:43:21.830Z

0xghost42 · 2026-06-17T06:20:37Z

@clawsweeper re-review

clawsweeper · 2026-06-17T06:20:41Z

🦞🧹
ClawSweeper re-review requested.

I asked ClawSweeper to review this item again.
Action: item re-review queued (workflow sweep.yml, event repository_dispatch).
Result: the existing ClawSweeper review comment will be edited in place when the review finishes.

Re-review progress:

State: Failed
Detail: The targeted re-review timed out or failed while waiting for exact-review Codex capacity. Retry after capacity recovers.
Run: https://github.com/openclaw/clawsweeper/actions/runs/27669955130
Updated: 2026-06-17T06:42:44.638Z

Zhipu (GLM) signals overload as HTTP 200 with body "[1305][该模型当前访问量过大，请您稍后再试]". The failover classifier had no pattern for it, so a configured model fallback never triggered and the run failed outright. Add the distinctive overload phrase to the overloaded message patterns, alongside the existing Chinese overload markers, so failover keys on it. Closes openclaw#93211

…Fallback Drive the real runWithModelFallback orchestrator with the reported Zhipu overload error [1305][该模型当前访问量过大，请您稍后再试]. Before the overload pattern the run classifies the error as "unknown"; with it the recovery is keyed on "overloaded" and falls back to the configured model. Covers openclaw#93211 at the orchestrator level alongside the matcher-level regression tests.

steipete · 2026-06-21T03:49:43Z

Merged via squash.

Prepared head SHA: db79e94c89dce69f3d894f51e694e31875043150
Merge commit: 6c1041339da6c1e0a66eca2f8127b8d752e4543f

Thanks @0xghost42!

@steipete

…penclaw#93241) Merged via squash. Prepared head SHA: db79e94 Co-authored-by: 0xghost42 <[email protected]> Co-authored-by: steipete <[email protected]> Reviewed-by: @steipete

openclaw-barnacle Bot added agents Agent runtime and tooling size: XS triage: mock-only-proof Candidate: PR proof only shows tests, mocks, snapshots, lint, typecheck, or CI. labels Jun 15, 2026

openclaw-barnacle Bot added proof: supplied External PR includes structured after-fix real behavior proof. and removed triage: mock-only-proof Candidate: PR proof only shows tests, mocks, snapshots, lint, typecheck, or CI. labels Jun 15, 2026

clawsweeper Bot mentioned this pull request Jun 15, 2026

fix(failover): classify Zhipu (GLM) error [1305] as overloaded #93212

Closed

4 tasks

github-actions Bot mentioned this pull request Jun 16, 2026

🦞 OpenClaw 生态日报 2026-06-16 96loveslife/big_model_radar#7

Open

clawsweeper Bot added status: 🔁 re-review loop A fresh ClawSweeper review was explicitly requested after the latest review. and removed status: 📣 needs proof The PR needs real behavior proof before ClawSweeper can clear the contributor ask. labels Jun 16, 2026

0xghost42 force-pushed the zhipu-overload-failover branch from ed503cd to e0eb3c4 Compare June 16, 2026 12:04

clawsweeper Bot mentioned this pull request Jun 20, 2026

Model fallback not triggered for Zhipu (GLM) error code 1305 — overloaded pattern mismatch #93211

Closed

steipete force-pushed the zhipu-overload-failover branch from 7794eee to 92efb24 Compare June 21, 2026 02:29

0xghost42 and others added 3 commits June 20, 2026 23:41

refactor(agents): keep overload matching provider-neutral

db79e94

steipete force-pushed the zhipu-overload-failover branch from 92efb24 to db79e94 Compare June 21, 2026 03:41

steipete merged commit 6c10413 into openclaw:main Jun 21, 2026
156 checks passed

clawsweeper Bot mentioned this pull request Jun 26, 2026

fix: classify HTTP 429 'temporarily overloaded' as overloaded, not rate_limit #97009

Closed

github-actions Bot mentioned this pull request Jun 29, 2026

chore(openclaw): bump pinned version v2026.5.28 → v2026.6.10 Evans-Software-Solutions-Limited/axel-saas#120

Open

4 tasks

clawsweeper Bot mentioned this pull request Jun 30, 2026

[Bug]: HTTP 429 + 'overloaded' body (e.g. z.ai code 1305) misclassified as rate_limit → false 'API rate limit reached' instead of overloaded #98101

Closed

Supreshone mentioned this pull request Jul 1, 2026

Model switch in same session breaks all tool calls (2026.6.10 regression) #98378

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Uh oh!

fix(agents): classify Zhipu GLM overload as overloaded for failover#93241

fix(agents): classify Zhipu GLM overload as overloaded for failover#93241
steipete merged 3 commits into
openclaw:mainfrom
0xghost42:zhipu-overload-failover

0xghost42 commented Jun 15, 2026 •

edited

Loading

Uh oh!

clawsweeper Bot commented Jun 15, 2026 •

edited

Loading

Uh oh!

0xghost42 commented Jun 16, 2026

Uh oh!

clawsweeper Bot commented Jun 16, 2026 •

edited

Loading

Uh oh!

0xghost42 commented Jun 17, 2026

Uh oh!

clawsweeper Bot commented Jun 17, 2026 •

edited

Loading

Uh oh!

Uh oh!

steipete commented Jun 21, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Uh oh!

Conversation

0xghost42 commented Jun 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Verification

Real behavior proof

Uh oh!

clawsweeper Bot commented Jun 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

0xghost42 commented Jun 16, 2026

Uh oh!

clawsweeper Bot commented Jun 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

0xghost42 commented Jun 17, 2026

Uh oh!

clawsweeper Bot commented Jun 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

steipete commented Jun 21, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

0xghost42 commented Jun 15, 2026 •

edited

Loading

clawsweeper Bot commented Jun 15, 2026 •

edited

Loading

clawsweeper Bot commented Jun 16, 2026 •

edited

Loading

clawsweeper Bot commented Jun 17, 2026 •

edited

Loading