fix(core): honor Codex input limits

[opencode-agent]

Summary

• preserve models.dev's separate input limit through the V2 LLM route
• overlay openai-codex model limits onto the existing OpenAI catalog only when ChatGPT OAuth is active
• compact against input - buffer and report TUI context usage against the effective input window
• leave direct OpenAI API-key model limits unchanged

This prevents GPT-5.6 Sol's ChatGPT subscription route from reaching output-budget exhaustion while the TUI still reports roughly 35% context usage. The canonical numbers live in anomalyco/models.dev#3754; OpenCode only selects and enforces the route-specific metadata.

@rekram1-node, the observed failure was repeated truncated patch JSON around 372K tokens. The provider returned max_output_tokens, V2 continued the malformed call, and manual compaction was the only escape. With the Codex input window applied, default auto-compaction begins at 252K (272K input minus the 20K buffer).

Verification

• packages/core: 38 focused tests pass across model resolution, OpenAI catalog routing, and compaction
• packages/tui: 4 context-usage tests pass
• packages/ai: GOMAXPROCS=2 bun typecheck
• package-wide Core/TUI typechecks remain blocked by existing generated SDK drift; no errors reference the changed limit or TUI usage code

Depends on anomalyco/models.dev#3754.

Fixes #38851
Related: #36766, #38976

Requested by: @kitlangton (Kit via Slack)

[magoz]

Production evidence strongly confirms this fix's diagnosis:

• OpenCode2 next-16302, openai/gpt-5.6-sol, ChatGPT browser OAuth
• read-only DB backup: 19 completed auto-compactions in one affected session
• 17/19 occurred at 33–36% displayed usage
• preceding provider usage consistently clustered at 370k–375k tokens

That is the 372k Codex input ceiling divided by the stale 1.05M public catalog window. #32119 alone would only change the display to ~40%; the credential-specific route limit is required.

One edge worth covering: OpenAIPlugin.load() currently calls connection.resolve() and catches refresh failure as undefined. For an expired OAuth credential, a transient refresh failure sets chatgpt = false; a later successful request-time refresh may not emit ConnectionUpdated, leaving public limits until another catalog reload. Reading stored credential/method type without refresh (or retaining ChatGPT route identity on refresh failure) would keep catalog metadata independent of token validity.

[magoz]

Rebased replacement: #39545. It preserves Kit’s commit authorship, resolves current v2 naming conflicts, removes the added TUI test assertion, and passes focused tests plus all 33 monorepo typechecks.