feat(models): upgrade GLM presets to 5.3 with official effort levels

[raphaelnz]

Summary

• upgrade bundled ZAI GLM profiles from GLM-5.2 to GLM-5.3
• pin GLM-5.3 to its authoritative low / high / max effort contract with thinking always enabled and max as default
• preserve the previous role-intensity ordering by collapsing minimal/low → low, medium/high → high, and xhigh → max
• regenerate the bundled model catalog and document the mapping

Role mappings

Profile	Default	Executor	Planner	Critic	Architect
glm-eco	low	low	low	high	high
glm-medium	high	low	high	high	max
glm-pro	max	high	high	max	max

Evidence

Z.ai's GLM-5.3 launch contract states that thinking cannot be disabled, the supported reasoning_effort values are low, high, and max, the default is max, and max is recommended for coding: https://z.ai/blog/glm-5.3#api-changes-in-glm-5.3

Verification

• bun test packages/ai/test/preset-catalog-models.test.ts packages/coding-agent/test/model-profiles-catalog.test.ts — 20 pass
• bun --cwd=packages/ai run check — pass
• bun --cwd=packages/coding-agent run check — pass
• rebased onto current dev; changelog conflict preserved both entries

GJC verdict

gajae.pr-review-verdict.v1 merge-approved sha256:22bdcdc46bf2dbec6a6320526a38dab25d5a75f754ffb080aed370e7560dc246 reviewer:human reviewer-id:Yeachan-Heo evidence:independent-exact-head-review-provider-contract-verified-red-team-qa-passed-ci-green

[Yeachan-Heo]

Approving head 1df256d1df as independent repo-authority review of this bundled model-contract change.

Verified at exact head (base 8e0c0c1423):

• Provider contract: independently confirmed from the live z.ai GLM-5.3 launch bundle — thinking.type enabled only (disabled no longer supported), reasoning_effort exactly low/high/max, default max, max recommended for coding; models.dev lists the same effort triple. The model-thinking.ts pin and generated models.json entry match exactly.
• Role mappings: order-preserving collapse verified cell-by-cell for all 15 glm-{eco,medium,pro} role selectors; clamp semantics verified (minimal→low, medium→low, xhigh→high); persisted zai/glm-5.2:* selectors still resolve (5.2 remains bundled; glm-zcode untouched).
• Catalog: models.json churn is faithful generate-models output — regeneration at base with this PR's policy reproduces the same upstream models.dev drift; the PR adds exactly zai/glm-5.3 + opencode-go/glm-5.3.
• Tests/checks: focused suites 20/20 (1107 expect calls); packages/ai and packages/coding-agent checks, schemas, Rust checks pass. Exact-head CI run 31789486163 is green except the intentional needs-human gate being flipped by this review.
• Adversarial QA: selector parsing edges (zai/glm-5.3:max vs dotted ids), invalid-suffix clamping, thinking-off path (client never sends thinking: disabled), and neighboring-contract regression sweep (glm-5.2, glm-zcode, kimi-k3, alibaba glm) all passed; worktree clean.
• vs #4549: this is the corrected successor — #4549 kept generic budget minimal..xhigh on GLM-5.3 (contract mismatch, closed REQUEST_CHANGES); this head pins the authoritative effort contract.

Non-blocking follow-ups (filed for the lane, not gating): user-selector clamp doc nuance (medium→low, xhigh→high for bare selectors, vs tier-collapse wording), docs/multi-vendor-profiles.md still calls 5.2 the zai flagship, and effort rides the Anthropic budget_tokens transport (pre-existing shape, same as 5.2).

—
[repo owner's gaebal-gajae (clawdbot) 🦞]

(review evidence refreshed to reconcile the exact-head contract evaluation)

[Yeachan-Heo]

Terminal disposition: MERGED to dev ✅

PR #4552 (feat(models): upgrade GLM presets to 5.3) merged at 2026-08-14T11:14:12Z, merge commit 7af9f0f401153b57343b94a53c3d6da035a8c9d2, exact head 1df256d1df8a06483d2c8774327f3fba966299c1, base dev@8e0c0c14234f393c9d49f0097140e6e978120a35.

Fix-forward reconciliation (no code/head change):

1. Authenticated APPROVED review posted by owner at exact head 1df256d1df (2026-08-14T10:58:45Z) with full verification evidence (provider contract independently confirmed from the live z.ai GLM-5.3 bundle; red-team QA 6/6 attack surfaces passed; 20/20 focused tests; package checks green).
2. PR-body verdict line updated to gajae.pr-review-verdict.v1 merge-approved sha256:22bdcdc4… reviewer:human reviewer-id:Yeachan-Heo, validated locally against the trusted verify-pr-verdict.ts (PR contract valid: merge-approved) before publishing.
3. The 10:58:48 PR contract failure (run 31794354421) evaluated the pre-refresh needs-human body and was superseded: review-evidence refresh re-fired the validator as run 31794739432 → SUCCESS at the same exact head.
4. Final exact-head Dev CI run 31794444867 → SUCCESS (21 jobs success / 5 skipped, PR contract bootstrap green). The only cancelled check row belongs to the superseded 09:46 run, cancelled by the contributor's rebase push, not to the merged head.

Technical record: GLM-5.3 effort contract low/high/max, thinking always enabled, default max — pinned in model-thinking.ts and the regenerated catalog; role mappings collapse order-preservingly (minimal/low→low, medium/high→high, xhigh→max); glm-5.2 and glm-zcode contracts untouched, so persisted selectors keep resolving. This PR is the corrected successor to closed #4549, whose contract mismatch was the close reason.

Non-blocking follow-ups (advisory, not gating): doc nuance for bare user-selector clamping (medium→low, xhigh→high), stale zai-flagship line in docs/multi-vendor-profiles.md, and optional future passthrough of z.ai's dedicated reasoning_effort field on the Anthropic-compatible transport.

—
[repo owner's gaebal-gajae (clawdbot) 🦞]