fix(openai): align GPT-5.6 Codex context limits

[LionSR]

Summary

• keep the existing 272k ChatGPT-subscription context cap for GPT-5.5 and earlier
• use the Codex-reported 372k input plus 128k output budget for GPT-5.6
• drive runtime and model-picker context from the same capability profile
• add one focused GPT-5.6 profile assertion

Validation

• npm test -- --run src/test-kernel/model/ProviderCapabilities.vitest.ts src/test-kernel/agent/modelHandlers/CodexSubscriptionFallback.vitest.ts src/test-kernel/model/ComputeModelOptions.vitest.ts (29 tests)
• npm run typecheck
• npm run lint (0 errors; 47 pre-existing warnings)
• targeted Prettier and ESLint
• git diff --check

Reference: anomalyco/opencode#36248

Closes #8305

---

Note

Medium Risk
Changes pre-flight rejection and compaction math on the ChatGPT-subscription path where token counting is unavailable; wrong limits would block valid turns or allow backend failures, but scope is limited to Codex capability profiles and the OpenAI Responses handler.

Overview
ChatGPT-subscription (Codex) limits are split so GPT-5.5 and earlier keep a 400k total window with a 272k input cap, while GPT-5.6 uses 500k total and 372k input (per Codex-reported budgets). ProviderCapabilityProfile gains optional inputTokenLimit, resolved from codexBackendModelId when the model id starts with gpt-5.6.

ModelHandlerOpenAIResponse uses getEffectiveInputTokenLimit() for automatic compaction thresholds and, on routes without token counting, rejects requests locally when input estimate + safety buffer hits the input cap—before relying on shrinking max_output_tokens. The model picker still shows the profile contextWindow (now the larger default/GPT-5.6 totals).

Tests rename constants, assert GPT-5.6 limits, and drop the case that allowed a subscription request when only the output budget was trimmed past the old single ceiling.

^{Reviewed by Cursor Bugbot for commit 98536ad. Bugbot is set up for automated code reviews on this repo. Configure here.}

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit fa75494. Configure here.}

[github-actions]

{
  "body": "## TeXRA Code Review\n\n### Summary\n\nThis PR correctly aligns the Codex ChatGPT-subscription context budget for GPT-5.6 by introducing a model-family-specific dispatch in `resolveProviderCapabilities`. The changes are limited in scope, well-tested, and mathematically sound.\n\n### Findings\n\n**1. Arithmetic and dispatch logic verified (no issues)**\n\nThe new constant `CODEX_GPT56_SUBSCRIPTION_CONTEXT_WINDOW = 500_000` is documented as \"372k input plus 128k output\" (372,000 + 128,000 = 500,000 \checkmark). The dispatch function `codexSubscriptionContextWindow` uses `codexBackendModelId(model).startsWith('gpt-5.6')`, which correctly matches all GPT-5.6-family model IDs (e.g., `gpt-5.6-luna`) while leaving GPT-5.5 and earlier at the default 272,000. The `Math.min` guard in `resolveProviderCapabilities` ensures the Codex cap never exceeds the model's declared native `contextWindow`.\n\n**2. Constant rename is complete (no issues)**\n\n`CODEX_SUBSCRIPTION_CONTEXT_WINDOW` has been renamed to `CODEX_DEFAULT_SUBSCRIPTION_CONTEXT_WINDOW`. All references across the three test files (`CodexSubscriptionFallback.vitest.ts`, `ComputeModelOptions.vitest.ts`, `ProviderCapabilities.vitest.ts`) have been updated. A grep for the old name returns no matches.\n\n**3. Test coverage (no issues)**\n\n- `ProviderCapabilities.vitest.ts`: the existing GPT-5.5 test still asserts 272k; a new test asserts 500k for GPT-5.6.\n- `CodexSubscriptionFallback.vitest.ts`: all fallback-boundary tests use the renamed default constant; no new GPT-5.6-specific fallback tests are needed because the fallback logic is model-agnostic and the context cap is now sourced through `resolveProviderCapabilities`.\n- `ComputeModelOptions.vitest.ts`: the model-picker display test uses the renamed constant; the picker display for GPT-5.6 would automatically derive the correct 500k cap from `resolveProviderCapabilities` without additional test changes.\n\n**4. No regressions identified (no issues)**\n\nThe PR title reports 29 passing tests, 0 lint errors. The diff is self-contained and introduces no new abstraction layers or TypeScript anti-patterns.\n\n### Risk assessment\n\n- **Low risk.** The change is a narrowly scoped addition of a model-family dispatch with a new constant. No existing code path is removed or restructured.\n- **Future-proofing note:** The `startsWith('gpt-5.6')` dispatch will match any future GPT-5.6 variant (e.g., `gpt-5.6-mini`). If a future variant requires a different budget, the dispatch will need adjustment, but this is a straightforward extension point.\n",
  "comments": [],
  "thread_actions": []
}

---

Reviewed by TeXRA agent review with model deepseekproT.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: fa75494a61

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[claude]

Findings

Suggestions

• src/agent/modelHandlers/openai/modelHandlerOpenAIResponse.ts:637 — JSDoc says "context window" but the implementation now uses getEffectiveInputTokenLimit(). Update the comment to reference the input token limit.

Notes

• src/model/providerCapabilities.ts:57 — codexSubscriptionTokenLimits() is a single-caller private function (only resolveProviderCapabilities). It carries meaningful logic (model-family dispatch across two return fields), so it's defensible and not a trivial identity extraction. If this remains at one caller through future iterations, consider inlining.
• The removed test case 'sends subscription requests when only the stripped output budget exceeds the Codex cap' is correctly dropped: the scenario was specific to the old single 272k cap where output budget could be squeezed but still sent. With the new inputTokenLimit guard in applyTokenCountFailureFallback, the input-limit check catches this scenario before the output-budget reduction path, so the test is no longer meaningful.
• The constant rename from CODEX_SUBSCRIPTION_CONTEXT_WINDOW to the new split constants is complete — zero references to the old name remain in the tree.
• No platform decoupling, Zod, security, PocketFlow, error-handling, or webview findings. All changes are in VS Code-free zones (src/agent/modelHandlers/, src/model/) and test files.

Verified

• src/model/providerCapabilities.ts:1-141 — constant definitions, dispatch function, capability profile resolution; interface and export changes
• src/agent/modelHandlers/openai/modelHandlerOpenAIResponse.ts:289-294, 635-644, 1378-1390 — new getEffectiveInputTokenLimit(), getCompactionTokenThreshold change, applyTokenCountFailureFallback guard
• src/test-kernel/agent/modelHandlers/CodexSubscriptionFallback.vitest.ts:1-270 — constant import updates, test assertion updates, removed test case
• src/test-kernel/model/ComputeModelOptions.vitest.ts:13-250 — constant import update
• src/test-kernel/model/ProviderCapabilities.vitest.ts:7-89 — constant imports, GPT-5.5 assertion update, new GPT-5.6 test
• Confirmed attachContextWindowError exists at src/common/errors/sdkError/errorMetadata.ts:102
• Confirmed no remaining references to CODEX_SUBSCRIPTION_CONTEXT_WINDOW (without suffix) in src/

[github-actions]

{
  "body": "## TeXRA Code Review\n\n### Summary\n\nThis PR splits the ChatGPT-subscription (Codex) context budget into separate **context window** and **input token limit** fields, with model-specific constants for GPT-5.5/earlier (400k total / 272k input) and GPT-5.6 (500k total / 372k input). The arithmetic is internally consistent - the 128k output budget is correctly derived from the difference - and the clamping, fallback, and early-rejection logic is sound.\n\n### Finding 1 - PR description does not match code behavior (moderate)\n\nThe PR title says \"keep the existing 272k ChatGPT-subscription context cap for GPT-5.5 and earlier,\" and the CURSOR_SUMMARY says \"GPT-5.5 and earlier still use **272k** via `CODEX_DEFAULT_SUBSCRIPTION_CONTEXT_WINDOW`.\" But the code defines:\n\n```typescript\nexport const CODEX_DEFAULT_SUBSCRIPTION_CONTEXT_WINDOW = 400_000;  // was 272_000\nexport const CODEX_DEFAULT_SUBSCRIPTION_INPUT_LIMIT = 272_000;     // new\n```\n\nThe **context window cap** for pre-5.6 models has been raised from 272k to 400k. The 272k number now governs only the input side. This is a material behavioral change: a model with a large native window that previously had its effective context window clamped to 272k now gets 400k. The code and tests are self-consistent, but the PR description is wrong about what was preserved.\n\n### Finding 2 - Single-caller helper extraction (minor)\n\nThe new `codexSubscriptionTokenLimits` function in `src/model/providerCapabilities.ts` (lines 57-71) is called from exactly one site (`resolveProviderCapabilities`). The repository's anti-indirection rules (CLAUDE.md § \"Discouraged Factory Patterns\" → \"single-caller extractions are banned\") discourage this pattern. The function does contain meaningful logic (model-ID → budget-constant mapping), which is an explicit exception, but the extraction adds a private function where the same `if (codexBackendModelId(model).startsWith('gpt-5.6'))` test could be inlined directly in `resolveProviderCapabilities`. Consider inlining to reduce indirection.\n\n### Finding 3 - Removed test without a positive-case replacement (minor)\n\nThe test \"sends subscription requests when only the stripped output budget exceeds the Codex cap\" was removed (CodexSubscriptionFallback.vitest.ts, old lines 253-278). This test previously verified a boundary: input near the old 272k cap but with enough output budget to succeed. With the new split budget (400k total / 272k input limit), the same input level now leaves ~128k of output room, so the old assertion no longer applies. The removal is logically correct, but no corresponding test verifies the *positive* case - that a request with cumulative input near (but below) the input limit, where the output budget is ample, actually succeeds and sends the request. Without such a test, a future change could silently break this path.\n\n### Verified\n\n- **Constants**: 400k − 272k = 128k (default output budget) \checkmark; 500k − 372k = 128k (GPT-5.6 output budget) \checkmark.\n- **Clamping**: `contextWindow = min(tokenLimits.contextWindow, model.contextWindow)` and `inputTokenLimit = min(tokenLimits.inputTokenLimit, model.contextWindow)`. The invariant `resolved inputTokenLimit ≤ resolved contextWindow` holds for all model context windows because `tokenLimits.inputTokenLimit ≤ tokenLimits.contextWindow` and both are clamped to the same `model.contextWindow`.\n- **`getEffectiveInputTokenLimit` fallback**: returns `inputTokenLimit` from the active profile, else falls back to `getEffectiveContextWindow()` → correct for non-subscription paths.\n- **`applyTokenCountFailureFallback` early rejection**: `inputEstimate + buffer >= inputTokenLimit` correctly guards the input budget before computing `contextWindow − inputEstimate − buffer` for the output budget.\n- **`getCompactionTokenThreshold`**: now uses input limit rather than context window → semantically correct (compaction should trigger against the input budget).\n- **Test constants**: `ProviderCapabilities.vitest.ts` uses `gpt55Config.contextWindow = 1_050_000`, so the GPT-5.6 test correctly expects unclamped subscription constants. `CodexSubscriptionFallback.vitest.ts` correctly updates imports and expected values. `ComputeModelOptions.vitest.ts` correctly shows \"400K\" from the renamed constant.\n- No new ES2023+ regressions; no new pass-through layers beyond the single-caller helper noted above.\n- No pre-existing TeXRA threads to resolve.",
  "comments": [
    {
      "path": "src/model/providerCapabilities.ts",
      "line": 57,
      "start_line": 57,
      "side": "RIGHT",
      "body": "The `codexSubscriptionTokenLimits` helper is called from exactly one site (`resolveProviderCapabilities`). Per the repository anti-indirection rules (CLAUDE.md § \"Discouraged Factory Patterns\" / Abstraction-cost guardrails), single-caller extractions are discouraged. The `if (codexBackendModelId(model).startsWith('gpt-5.6'))` test and the two-constant selection could be inlined directly in `resolveProviderCapabilities` (5 lines of logic). Consider inlining unless you expect additional callers imminently."
    },
    {
      "path": "src/test-kernel/agent/modelHandlers/CodexSubscriptionFallback.vitest.ts",
      "line": 253,
      "start_line": 253,
      "side": "LEFT",
      "body": "This test verified that a request with input near the old 272k cap (but output budget still viable) would succeed. With the new split budget (400k context / 272k input), the same scenario now leaves ~128k of output room, so the old assertion is inapplicable. The removal is correct, but no replacement test covers the *positive* case: cumulative input near (but below) the input limit with ample output budget should successfully send the request. Without such a test, this path is untested."
    }
  ],
  "thread_actions": []
}

---

Reviewed by TeXRA agent review with model deepseekproT.