[Bug]: client config export sets maxTokens to the 32000 schema stand-in for every model, truncating max-effort thinking turns (and over-reporting grok-4.20)

[thisisjun786]

Client or integration

OMO (senpi) through ocx integration client enable --client omo. The same export path serves Pi and OMP.

Area

Client config export (src/clients/config-export.ts, config-export/model-metadata.ts)

Summary

Every model in the exported providers.opencodex block gets maxTokens: 32000. The value comes from outputBudgetFor(context) = min(SCHEMA_REQUIRED_OUTPUT_BUDGET, context). The constant's own comment says it is "a ceiling for schema validity, NOT a claim about any specific model's true maximum".

Clients treat maxTokens as the real output ceiling and send it as max_tokens on every request. That causes two failures:

1. Too low for thinking models. Claude Opus 5.5 at effort: max on a ~220k-token context spent the whole 32,000-token budget on adaptive thinking. The turn ended stop_reason: max_tokens with no text and no tool call (usage.jsonl: status 200, outputTokens 32000, "Output reached the requested token limit"). The client got an empty assistant message after 319 s. The real Opus 5.5 limit is 128K, and ocx's own model-seeds.ts says so.
2. Too high for some models. xai/grok-4.20-0309-non-reasoning is exported at 32,000, but its limit is 30,000 (ocx generated/model-metadata.ts).

ocx already has authoritative per-model output limits for most exported models. They are in generated/model-metadata.ts (third column) and the Anthropic registry seeds:

exported id	exported	ocx metadata
anthropic/claude-opus-5-5, claude-fable-5-1, claude-sonnet-5	32000	128000
anthropic/claude-haiku-4-5	32000	64000
gpt-5.6-{luna,sol,terra}, gpt-6-{luna,sol} (+ --fast)	32000	128000
google/gemini-3.8-flash	32000	65536
ollama-cloud/glm-5.3, glm-5.3-flash	32000	131072
kimi/k3[1m] (kimi-k3)	32000	131072
xai/grok-4.7	32000	500000
xai/grok-4.20-0309-non-reasoning	32000	30000

Expected

• When the catalog knows a model's output limit, the export uses it, clamped to the exported context as it is today.
• The 32K stand-in applies only to models with no known limit.
• A limit known to be lower than 32K is never raised to 32K.

Reproduction

1. ocx integration client enable --client omo
2. Inspect ~/.omo/agent/models.json → providers.opencodex.models[*].maxTokens: all 31 entries are 32000.
3. Run a long max-effort Opus 5.5 turn. It ends max_tokens with empty output.

Version

opencodex 2.65.0

Operating system

Linux x64

Provider and model

anthropic (OAuth) claude-opus-5-5; affects every exported model

Checks

• I searched existing issues and documentation. Related: [Bug]: Anthropic adaptive thinking is requested with default display (omitted), so long Opus max-effort thinking reaches Chat clients as minutes of heartbeat-only silence #5824 (thinking display / silent long thinking). This issue is the separate output-budget half.
• I removed secrets, tokens, account details, request credentials, and personal data.

[devin-ai-integration]

Triage: priority: P2 — client config export writes the 32000 schema stand-in as every model's maxTokens; client-side workaround.

Criteria (P2): Medium: provider/client-specific bug with a workaround, bounded enhancement tied to a tracked issue, perf, or CI reliability.

No open pull request addresses this yet.

[lidge-jun]

Fixed on dev by #5891, carried in #5901 (merge dd1e327). It ships with the next release.