Conversation
- GLM 5.3 Flash accepts images and GLM 5.3 does not, yet both carry vendor zhipu, so a vendor-level set cannot express the difference - modelSupportsImages replaces vendorSupportsImages and checks the vendor set or an explicit image-capable slug set - the five supportsImages call sites (Pi adapter, CLI direct and Tinfoil bindings) switch to it
- GLM 5.3 ships with reasoningEffort 'max'; piThinkingLevels stopped at xhigh, so a 'max' profile silently fell back to medium - the test pins that 'max' now derives 'max'
- pi-ai 0.80.7 has no glm-5.3, so glm-5-3 and glm-5-3-flash resolve compatibility through the glm-5.2 entry, which already carries the right thinkingFormat (zai), zaiToolStream and supportsReasoningEffort - the thinking map is overridden for those two slugs: Pi's glm-5.2 map sends low as 'high', so low stays 'low', medium/high map to 'high', and max maps to 'max' - the glm-5-2 alias stays for rows not yet migrated - tests assert the outgoing reasoning_effort for both new slugs - swap the aliases for the real catalog ids once Pi is upgraded (THU-867)
- Tinfoil removed glm-5-2 from its catalog, so GLM 5.2 becomes GLM 5.3 (glm-5-3) and DeepSeek V4 Flash becomes GLM 5.3 Flash (glm-5-3-flash, vendor zhipu) on the same row ids; defaults version 6 drives the OTA - GLM 5.3 Flash is the new defaultModelId for new accounts, /config, and the CLI - Profiles ship reasoningEffort max (GLM 5.3) and low (Flash); no profile data migrates because modelId is unchanged - Eval slug names and compile-only symbol references follow the rename
- reconciliation never touches a row whose hash diverged, so users who edited GLM 5.2 or DeepSeek V4 Flash stayed on a dead slug; the Opus repair from THU-843 only covered its own row - the Opus migration becomes a table of lineages — Opus (sonnet-4.5, opus-4.8), GLM (glm-5-1, glm-5-2), Flash (deepseek-v4-flash) — iterated by the same normalize/upgrade logic, preserving a user-chosen name - reconcile-defaults calls the generic version
- the app only stores selected_model when the user picks one, so a new account fell back to the system models ordered by name — after the rename that would be GLM 5.3 at max effort - getSelectedModelQuery now orders the defaultModelId row first, so new accounts land on GLM 5.3 Flash while an explicit selection still wins
- The proxy header check, the no-header fallback and receipt verification all resolve slugs from defaultModels, so renaming the defaults would 400 every client whose row is not renamed yet: edited rows, old builds, and receipts issued up to two hours before the deploy. - The two retired slugs join the same map as legacy identities; their old price rows stay in place. - The no-header fallback imports the current default GLM slug from shared instead of a hand-written 'glm-5-2'.
- inference_prices is keyed on the exact (provider, model) with no fallback, so a slug without a row answers 503 INFERENCE_PRICE_UNAVAILABLE the moment the new defaults go live. - Migration 0029 seeds glm-5-3 at 1500/5250 and glm-5-3-flash at 300/700. These are the quota prices of glm-5-2 and deepseek-v4-flash on purpose, not Tinfoil's list prices (1800/5750, 400/1250), so the per-token quota weight stays the same across the rename. The old rows stay in place for the legacy slugs. - The journal entry (idx 29) is included: Drizzle only discovers migrations through the journal. - Run this migration in production before deploying the backend.
Semgrep Security ScanNo security issues found. |
ital0
marked this pull request as ready for review
September 6, 2026 18:31
|
Preview environment destroyed 🧹 Stack |
PR Metrics
Updated Mon, 07 Sep 2026 16:04:15 GMT · run #2971 |
- The lineage migration reused a row's id but only moved model+name, leaving vendor/description stale. The Flash lineage changes vendor deepseek→zhipu, so an edited DeepSeek Flash row kept vendor 'deepseek' after becoming glm-5-3-flash, and resolveConfidentialModelCompatibility then looked up deepseek/glm-5.2 (undefined) and threw compatibility-missing on every send. - Reconcile already carries these server-owned metadata fields for intact rows, but edited rows only pass through the migration, so normalizeModelDefault/upgradeModelDefaults now also migrate vendor and description to the target. - The lineage tests now seed the legacy vendor (deepseek) and an old description so the regression is covered. - Found by thunder-deep-review; GLM and Opus lineages were unaffected (no vendor change).
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
There was a problem hiding this comment.
🔭 thunder-deep-review (advisory)
Complements the other bots — surfaces only what they did not flag. Never approves, never requests changes, never gates merge.
head: f952e6becbc4 · mode: deep · deferred 0 item(s) already reported by other bots (best-effort dedup)
…t-to-53-and-replace-deepseek-v4-flash
- the JSDoc above defaultModelGlm53Flash still said confidential Flash "ships under a fresh id"; that was true when #1264 minted 01a06dd7 instead of reusing the retired direct Flash row, but GLM 5.3 Flash now reuses that same id in place, so the sentence read as if this PR created a new id - the frozen provider/isConfidential rationale already lives in the reconciliation code, so remove the comment rather than reword it
ital0
changed the title
ital0
deleted the
italomenezes/thu-860-bump-tinfoil-glm-default-to-53-and-replace-deepseek-v4-flash
branch
September 10, 2026 13:57
This branch was successfully deployed
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
1 participant
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.
Summary
Tinfoil removed
glm-5-2from its catalog (the enclave still answers, on borrowed time). This PR moves both confidential defaults to the GLM 5.3 family on the same row ids, so threads, profiles and selections keep working and the change ships OTA through/config:019e7580-…glm-5-2glm-5-3·reasoningEffort: 'max'01a06dd7-…deepseek-v4-flash· vendordeepseekglm-5-3-flash· vendorzhipu·reasoningEffort: 'low'· image inputdefaultModelsVersion5 → 6. GLM 5.3 Flash becomes the default for new accounts,/configand the CLI (getSelectedModelQuerynow prefersdefaultModelIdbefore the alphabetical fallback). Effort levels follow the decision recorded on the ticket (max / low).What else had to change
zhipu, so the vendor set could not express "Flash yes, GLM 5.3 no".modelSupportsImageschecks the vendor set or an explicit slug set; five call sites switch to it. Composer behaviour for non-image models is unchanged (THU-844).glm-5-3/glm-5-3-flashalias onto Pi'sglm-5.2catalog entry, with the thinking map overridden solowstayslow(Pi's glm-5.2 map sends it ashigh). The adapter now accepts themaxthinking level. Theglm-5-2alias stays for rows not yet migrated. Removing the bridge is tracked in THU-867 (Pi 0.85.x migration; installable with the 7-day quarantine intact from 2026-09-12).upgrade-opus-default.tsbecame a lineage table (upgrade-model-defaults.ts): Opus (sonnet-4.5,opus-4.8), GLM (glm-5-1,glm-5-2), Flash (deepseek-v4-flash). A user-chosen name is preserved.glm-5-2anddeepseek-v4-flashjoin the same map that backs the proxy header check, the no-header fallback and receipt verification, so clients whose rows have not been renamed yet (edited rows, old builds, receipts issued before the deploy) keep working. The no-header fallback now imports the current default GLM slug instead of a hand-written string.0029seeds('tinfoil','glm-5-3',1500,5250)and('tinfoil','glm-5-3-flash',300,700)— the glm-5-2 and deepseek-v4-flash prices, not Tinfoil's list prices (1800/5750, 400/1250) — so the per-token quota weight does not change. Old rows stay.cost_nano_usdtherefore stops being the invoice for these two slugs.Compatibility notes
Managed model "glm-5-3(-flash)" has no Pi compatibility metadataafter the OTA rename (no request leaves the device); Opus keeps working. Builds 0.1.124–0.1.127 use the old AI SDK path and keep working, at the enclave's default effort.glm-5-2/deepseek-v4-flashkeep working as long as Tinfoil serves those slugs.Operations (before merge/deploy)
INFERENCE_PRICE_UNAVAILABLE.completion_tokensper message forglm-5-3vsglm-5-2ininference_usage; ifmaxinflates responses, lower the quota price ofglm-5-3with a one-line migration.