feat(cli): tole doctor full tier — chat, streaming and a real tool-call loop (#378, 2/3)

[ajianaz]

What

Part 2 of 3 of tole doctor: the full tier. Builds on the offline and catalog tiers merged in #388.

Why

#378. A full green is the machine-checkable gate for "ready to release" claims: the provider must actually chat, stream and complete a tool-call loop, not just answer /models.

Changes

• doctor.rs: full tier = three real calls through the real turn loop (run_turn + OpenAiProvider): a plain chat, a streamed chat, and a tool-call loop where the model must call a ReadOnly probe tool with the expected marker and the loop must settle on a final answer. Each failure is classified (chat / stream / tools) with a fix command; the model answering without calling the tool is reported as such.
• main.rs: --tier full.
• README, CONTRIBUTING (Release validation: tole doctor --tier full must exit 0), CHANGELOG.

Testing

• 3 new tests against a scripted local server that speaks plain JSON and SSE: well-behaved provider (green), model never calls the tool (blocker at tools, reached = catalog), broken completions (blocker at chat).
• fmt, clippy -D warnings, cargo test --workspace pass; cora review --staged clean.
• Release binary exercised earlier against a local mock (green run and no-tool-calling run, exit codes 0 and 1).
• Not done: a live run against a real provider (none configured here); feat(cli): provider doctor with tiered diagnostics + persisted coverage ledger #378 stays open for that and for the ledger (PR 3/3).

Refs #378

[github-actions]

🔍 Cora AI Code Review

✅ No issues found. Code looks good!

---

Review powered by cora-code · BYOK · MIT