fix: include cached input tokens in OpenAI usage

[nmzpy]

Cognition's cloud API reports cached prompt tokens separately from uncached prompt tokens. OpenAI-compatible "prompt_tokens" should be the full input count (cached + uncached), so fold "cachedInputTokens" back in.

This prevents fully prompt-cached requests (large system prompts + tool schemas) from appearing to consume only ~1 input token in opencode.

Changes:

• Added "buildOpenAIUsage" helper that returns prompt_tokens, completion_tokens, total_tokens, and prompt_tokens_details.cached_tokens.
• Used the helper in both streaming and non-streaming completion paths.
• Added unit tests for cached, uncached, and empty usage.

Summary by CodeRabbit

• New Features

  • Token usage reporting now includes cached input tokens.
  • Streaming and non-streaming responses expose cached token counts through prompt_tokens_details.cached_tokens.
  • Prompt and total token counts now accurately reflect cached input usage.

• Tests

  • Added coverage for cached, uncached, and missing token usage data.

[coderabbitai]

📝 Walkthrough

Walkthrough

Cached input tokens are now included in OpenAI-compatible usage accounting. Streaming and non-streaming responses use a shared exported helper that reports cached prompt tokens and updated totals, with unit tests covering cached, absent, and empty usage inputs.

Changes

Usage Accounting

Layer / File(s)	Summary
OpenAI usage builder src/plugin.ts	Adds buildOpenAIUsage to combine prompt and cached input tokens, calculate totals, and expose prompt_tokens_details.cached_tokens.
Streaming and non-streaming integration src/plugin.ts, tests/unit/usage.test.ts	Both usage paths capture cached input tokens and build responses through the shared helper; tests cover cached, missing, and empty token values.

Estimated code review effort: 2 (Simple) | ~10 minutes

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly matches the main change: including cached input tokens in OpenAI usage reporting.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[coderabbitai]

🧹 Nitpick comments (1)

src/plugin.ts (1)

421-421: 🎯 Functional Correctness | 🔵 Trivial | ⚡ Quick win

Add coverage for both response paths.

The current tests exercise buildOpenAIUsage directly, but not the streaming SSE or non-streaming response wiring. Add focused tests that emit cachedInputTokens and assert the serialized usage in each path; otherwise either capture/serialization path could regress while the helper tests still pass.

Also applies to: 599-599

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@src/plugin.ts` at line 421, The tests cover buildOpenAIUsage directly but not
its integration with streaming SSE and non-streaming responses. Add focused
tests for both response paths, emitting cachedInputTokens and asserting the
serialized usage payload, including the wiring around buildOpenAIUsage in each
path.

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Nitpick comments:
In `@src/plugin.ts`:
- Line 421: The tests cover buildOpenAIUsage directly but not its integration
with streaming SSE and non-streaming responses. Add focused tests for both
response paths, emitting cachedInputTokens and asserting the serialized usage
payload, including the wiring around buildOpenAIUsage in each path.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: 9f326a29-a259-4294-acac-a4a21ded7c88

📥 Commits

Reviewing files that changed from the base of the PR and between c331611 and c21ba5c.

📒 Files selected for processing (2)

• src/plugin.ts
• tests/unit/usage.test.ts