GLMCHAT V2.2.1 — TOGETHER PROVIDER CATALOGUE ALIGNMENT REPORT
Completed: 29 July 2026 12:33:39 UTC
Baseline: GLMCHAT_V2.2.1_SEVEN_DEFECTS_FIXED_HANDOVER_22072026192850.zip

REQUESTED WORK
Continue the autonomous app-completion workflow from the latest saved baseline, fix remaining locally verifiable gaps, preserve unrelated behaviour and package a verified handover.

CONFIRMED PROVIDER-COMPATIBILITY DEFECTS
1. The default text/coding identifier was zai-org/GLM-5.2, which was not present in the verified current Together serverless catalogue used for this audit.
2. GLM reasoning was sent through reasoning_effort although Together documents GLM as a hybrid reasoning model controlled through reasoning.enabled.
3. Preserved reasoning was not replayed under the documented reasoning_content field.
4. Text cached-input and Kokoro pricing metadata were stale.
5. Vision usage was calculated using the text-model price profile.
6. Streaming recognised only one reasoning delta field variant.
7. Cost preflight rejected saved off and auto reasoning modes even though the settings schema allowed them.

REMEDIATION
- Default text/coding model aligned to zai-org/GLM-5.1.
- Default context metadata aligned to 202,752 tokens.
- Text pricing metadata aligned to $1.40 input, $1.40 cached input and $4.40 output per million tokens.
- Qwen/Qwen3.5-9B vision pricing metadata added at $0.10 input/cached input and $0.15 output per million tokens.
- Kokoro TTS metadata aligned to $4.00 per million characters; Orpheus remains $15.00.
- GLM off/high/max modes use reasoning.enabled; auto leaves provider default behaviour unchanged.
- Preserved thinking is replayed using reasoning_content.
- Non-streaming and streaming response parsing accept both reasoning and reasoning_content field variants.
- Chat reservations, streamed usage and reconciled cost records use the actual selected text or vision model.
- Online, offline and synchronisation settings schemas include the three vision price fields.
- Migration 014 aligns only recognised legacy defaults and preserves administrator-customised values.
- Service-worker cache generation advanced to deliver the rebuilt provider-aligned frontend/runtime assets.

PERMANENT COVERAGE ADDED
- Node provider-catalogue alignment tests.
- PHP static provider payload, pricing and migration contracts.
- Python fresh-install, 13-to-14 upgrade, rollback and custom-value-preservation tests.
- Package validator checks for the 14-migration schema and aligned defaults.
- Target-host PHP integration assertions for GLM, cached input, vision and Kokoro cost calculations.

SCOPE PRESERVATION
- No route, feature, authentication model, user-data table, PWA scope, offline workflow or deployment architecture was removed.
- No dependency was upgraded.
- Existing administrator-customised provider settings are not overwritten by migration 014.
- Historical evidence files remain historical and were not rewritten merely to remove legacy model references.

SOURCE BASIS
- Together official serverless model catalogue and pricing, checked 29 July 2026.
- Together official reasoning overview for hybrid reasoning controls, checked 29 July 2026.
- Together official preserved-thinking guidance, checked 29 July 2026.

LIMITATIONS
- This work aligns source, defaults, migrations and cost governance to the verified catalogue; it does not prove that the user's Together account has live access.
- Live text, reasoning, streaming, cancellation, tools, vision, voice inventory and TTS remain target-system acceptance checks.
- PHP sqlite3, curl, mbstring and zip are unavailable in this sandbox, so the real PHP integration suite remains a target-host check.
- Production-data and physical Android/accessibility checks remain external.