[Bug] openai-codex applies one flat 272k context window to every model, including 1M-context models

[LCTechnoConception]

Version: v0.7.4 (af0b8e00b9f704e834787fd321065ca78281f2aa)
Provider: openai-codex (ChatGPT OAuth)

Summary

openai-codex reports contextWindow: 272000 for 12 of its 13 models (the exception is
gpt-5.3-codex-spark at 128000). That constant is applied uniformly, including to models
whose real window is ~4x larger on every other provider that serves them. gpt-5.6-sol is
the clearest case.

Where it comes from

packages/ai/scripts/generate-models.ts:1948-1953:

// OpenAI Codex (ChatGPT OAuth) models
// NOTE: These are not fetched from models.dev; we keep a small, explicit list to avoid aliases.
// Context window is based on observed server limits (400s above ~272k), not marketing numbers.
const CODEX_BASE_URL = "https://chatgpt.com/backend-api";
const CODEX_CONTEXT = 272000;
const CODEX_MAX_TOKENS = 128000;

Twelve of the thirteen Codex entries take contextWindow: CODEX_CONTEXT verbatim; only
gpt-5.3-codex-spark carries its own value. No entry scales with the model's actual window.

Two dates suggest the value may not have been re-checked for newer models:

• CODEX_CONTEXT = 272000 was introduced 2026-01-07 in d893ba7f
  ("fix(ai): clean up openai-codex models and token limits").
• gpt-5.6-sol was added to the Codex list on 2026-07-09 in 4a0b3de7
  ("add gpt-5.6 model catalogs", add gpt-5.6 model catalogs #361) and inherited the existing constant.

The discrepancy

gpt-5.6-sol across providers, read from a freshly built v0.7.4 catalog:

provider	contextWindow
azure-openai-responses	1,050,000
cloudflare-ai-gateway	1,050,000
github-copilot	1,050,000
openai	1,050,000
opencode	1,050,000
openai-codex	272,000

For gpt-5.1 the constant is consistent with that model's published limits
(400,000 total - 128,000 output = 272,000). For a 1,050,000-context model the
same subtraction would give 922,000, so the flat constant is not explained by
output reservation for the 5.6 generation.

Why it matters

shouldCompact() (packages/coding-agent/src/core/compaction/compaction.ts:206) compares
usage against contextWindow - reserveTokens, and contextWindow is read verbatim from the
registry. If the constant understates the real limit, sessions compact far earlier than
necessary and lose context that the provider would have accepted.

Measurement

A differential test on v0.7.4, same payload and provider, run seconds apart:

model	contextWindow in registry	result
gpt-5.6-luna	272,000 (stock)	blocked client-side before sending: "Auto-compaction skipped: Session is too short to compact"
gpt-5.6-sol	1,050,000 (via models.json override)	request accepted by Codex, normal answer returned

Method: a ~1.2 MB text payload (~127k words) attached to a single print-mode prompt
(-p --no-session --provider openai-codex --model <id> @payload.txt "Reply with exactly one word: ACK"), with gpt-5.6-sol overridden via models.json modelOverrides:

{ "providers": { "openai-codex": { "modelOverrides": {
  "gpt-5.6-sol": { "contextWindow": 1050000 } } } } }

The Luna run establishes that the payload genuinely exceeds 272k rather than being
truncated somewhere upstream: it triggered the context-overflow path and never reached
the API. The Sol run then shows the Codex backend accepting that same payload.

Caveat: I did not capture exact input token counts, so this demonstrates "accepted well
above 272k", not the precise ceiling. I have not tried to find where the real limit sits.

Reproduction

prime-agent model list | grep openai-codex
# every model reports 272K

Possible direction

If the limit does differ per model, the Codex list could carry a per-model context value
(defaulting to CODEX_CONTEXT where unmeasured) rather than a single shared constant,
so newly added models do not silently inherit a figure measured for an older generation.

Happy to run further measurements if that would help.

[aniruddhaadak80]

Verified still present on main @ e319a66d7 (current HEAD). All 12 non-spark entries in the Codex block (packages/ai/scripts/generate-models.ts:1951-2110) take CODEX_CONTEXT = 272000 verbatim; only gpt-5.3-codex-spark carries its own value (128000).

Extended the cross-provider comparison with the current generated catalog (packages/ai/src/models.generated.ts), including Bedrock:

model	openai-codex	openai	azure-openai-responses	amazon-bedrock	github-copilot	cloudflare-ai-gateway	opencode
gpt-5.1	272,000	400,000	400,000	-	-	400,000	400,000
gpt-5.2	272,000	400,000	400,000	-	400,000	400,000	400,000
gpt-5.4	272,000	272,000	272,000	-	1,050,000	1,050,000	272,000
gpt-5.5	272,000	272,000	272,000	272,000	1,050,000	1,050,000	1,050,000
gpt-5.6-luna	272,000	1,050,000	1,050,000	272,000*	1,050,000	1,050,000	1,050,000
gpt-5.6-sol	272,000	1,050,000	1,050,000	1,050,000	1,050,000	1,050,000	1,050,000
gpt-5.6-terra	272,000	1,050,000	1,050,000	1,050,000	1,050,000	1,050,000	1,050,000

* Bedrock lists openai.gpt-5.6-luna at 272,000 but sol/terra at 1,050,000 — even upstream sources disagree on luna.

Two refinements to the original report:

1. The flat constant is not obviously wrong for every entry: OpenAI's own catalog lists gpt-5.4 and gpt-5.5 at exactly 272,000, matching CODEX_CONTEXT. The discrepancy concentrates in the gpt-5.6 family (sol/terra/luna at 1,050,000 on six providers), which was added later (add gpt-5.6 model catalogs #361) and inherited the constant without rechecking. The 400,000 vs 272,000 gap on gpt-5.1/5.2 is plausibly a genuine ChatGPT-OAuth backend limit, consistent with the block's comment about observed 400s above ~272k.

2. Concrete runtime impact beyond display: shouldCompact() (packages/coding-agent/src/core/compaction/compaction.ts:206) and isContextOverflow() both consume model.contextWindow, so a 5.6-family session compacts roughly 3.9x earlier than necessary (at ~250k tokens instead of ~1M), discarding context that would have fit.

Fix direction, per the repo's rules:

• The change belongs in packages/ai/scripts/generate-models.ts (the catalog is generated; models.generated.ts is never hand-edited).
• Per-model contextWindow overrides sourced from the openai provider entries where IDs match, keeping CODEX_CONTEXT as the fallback plus the observed-limit caveat comment for entries without a cross-provider reference.
• Regenerate models.generated.ts via the generator script and add a packages/ai/.changes/ fragment.

One open question for maintainers before bumping values: whether the chatgpt.com/backend-api endpoint actually accepts >272k input tokens for the 5.6 family, or whether it caps lower than API access like it appears to for 5.1/5.2. If it genuinely caps lower, keeping conservative values is correct and this should stay closed as working-as-intended for those entries. Happy to prepare the patch once that's confirmed.