[diegosouzapw]

Root cause

AgentRouter routes GLM models through the generic DefaultExecutor (executor: "default"), which has no GLM-specific handling. When the connection's OpenAI-compatible alternate format is used and the client request already carries a Claude-style thinking: {type: "adaptive"} field (source format == target format == "openai", so translateRequest's hub-and-spoke conversion is skipped), that field survived stripUnsupportedParams untouched — STRIP_RULES in open-sse/translator/paramSupport.ts had no rule for provider agentrouter + glm-* — and reached AgentRouter's upstream GLM endpoint verbatim, which 400s with thinking.type "adaptive" is not supported by glm models; must be one of enabled, disabled.

The native GLM/ZAI executor (open-sse/executors/glm.ts:404-420) already performs exactly this adaptive → enabled remap for its own provider; AgentRouter just never reaches that code path since it uses the generic DefaultExecutor instead of GlmExecutor.

Fix

Extended the StripRule shape in open-sse/translator/paramSupport.ts with an optional mapThinkingType map (alongside the existing drop/clamp mechanisms), and added a rule scoped to provider: "agentrouter" + match: /glm-/i that remaps thinking.type from "adaptive" to "enabled", preserving any other thinking fields (e.g. budget_tokens) instead of dropping the whole field. Scoped to the whole glm-* family (not a single model id), since AgentRouter's static catalog carries no glm-* entries at all — passthroughModels: true lets arbitrary GLM ids through, so the fix must not be tied to one model.

Regression test

tests/unit/agentrouter-glm-thinking-adaptive-map.test.ts (new):

• agentrouter + glm-5.3 with thinking.type: "adaptive" → asserts thinking becomes {type: "enabled", budget_tokens: ...} (other fields preserved)
• agentrouter + claude-opus-5 (regression guard) → thinking.type stays "adaptive"
• non-agentrouter provider with a glm-* model id → thinking.type stays "adaptive" (native GLM executor handles its own mapping separately)
• agentrouter + glm-5.3 with no thinking field → no-op, nothing introduced

RED (before fix): AssertionError: adaptive must be mapped to enabled for agentrouter GLM models, preserving other thinking fields (actual {type: 'adaptive', budget_tokens: 1024} vs expected {type: 'enabled', budget_tokens: 1024}).

GREEN (after fix): all 4 new tests pass; npm run test:unit equivalent focused run — node --import tsx/esm --test tests/unit/agentrouter-glm-thinking-adaptive-map.test.ts → 4/4 pass.

Also updated the existing structural assertion in tests/unit/executors-strip-unsupported-params.test.ts ("STRIP_RULES is non-empty and every rule has a drop list or a clamp mechanism") to also accept the new mapThinkingType mechanism — it previously only recognized drop/clamp rules and would otherwise fail on the new rule shape. This is an alignment to the corrected contract (a StripRule may now sanitize by remapping a value, not only by dropping/clamping), not a weakening: the assertion still requires every rule to do something (drop, clamp, or map), just enumerates the new legitimate option.

Gates run

• node scripts/check/check-file-size.mjs — OK (no violation on touched files; the reported test-file-size violation on tests/unit/chatcore-translation-paths.test.ts is pre-existing/untouched by this PR)
• node scripts/check/check-complexity.mjs — OK (2857 violations vs baseline 3218)
• node scripts/check/check-cognitive-complexity.mjs — OK (1291 violations vs baseline 1437)
• npm run typecheck:core — exit 0
• npm run check:open-sse-typecheck — OK, 0 errors
• npx eslint --suppressions-location config/quality/eslint-suppressions.json <changed files> — clean
• Existing param-support tests re-run individually (DB-per-file to avoid SQLITE_BUSY contention): azure-max-output-clamp, chatcore-unsupported-params-strip, default-reasoning-effort-6879, executors-strip-unsupported-params, glm-executor-max-tokens-clamp-7364, nvidia-minimax-thinking-strip, nvidia-strip-prompt-cache-key, param-filters-apply, param-filters-db, zai-glm-max-tokens-clamp-7364 — all pass
• node scripts/check/check-changelog-integrity.mjs — OK

Live validation (Hard Rule #18)

Not performed against the production VPS in this pass — the fix is validated via TDD (a failing-then-passing regression test proving the exact reported 400 condition), which is the preferred validation path per the bug-fix protocol. The reported error string and root cause were confirmed by the prior triage's TDD probe and by the existing in-repo precedent (glm.ts's identical adaptive→enabled mapping for its own provider, and the identical NVIDIA NIM z-ai/glm-5.2 clash already fixed in paramSupport.ts).

⚠️ base-red inherited: #14004 — docs env/docs contract (fixed separately in #14022), chatHelpers file-size drift

Closes #13696

[diegosouzapw]

Validated in local merge-train /tmp/mt-train3b.log on 192.168.0.113 @ train tip 408e2128791696a18966681953136fc96aeb99b0 (32 PRs boarded): static gates green; full test:unit 40694 tests, 17 failing — every one reproduces on the pure release tip (base-red sweep list), zero new reds. Merged --admin per merge-gates §4/§7.