kit-provider-adapters

Kit provider adapters and harness drivers

The design is docs/PRD.md §6. Two kinds of adapter, two crates:

Kind	Crate	Interface	Example
Model provider (Kit's own agent)	crates/kit-providers	Provider trait: models(), stream(Request) -> EventStream	anthropic.rs, openai_compatible.rs
External agent driver (harness)	crates/kit-agents	Agent trait: probe(), spawn() → AgentHandle; events become KitEvent	claude.rs, codex.rs, grok.rs

Rules

1. Check the real docs, never guess a flag. Record the doc URL and the date in a comment above the invocation, and the CLI version you tested against. Upstream flags change: codex --full-auto and codex mcp-server were removed in 2026.
2. One file per provider or driver, ≤ 600 lines. Shared SSE parsing and retry live in one module.
3. Prompts never go on argv (Windows cmd injection and length limits). Use stdin, a prompt file the tool's published docs list, or the protocol (ACP).
4. No shells. Spawn the real executable; on Windows resolve .cmd shims to their entry point as the existing adapters do.
5. Auth stays where it is. Providers: env var → OS keychain (kit:<provider>) → key_command. Drivers: the tool's own login, checked with its status command (claude auth status, codex login status). Never read another tool's credential file. Never log a key; keys never enter a tool's environment.
6. Errors map to the closed set: Auth, RateLimited{retry_after}, ContextOverflow, Transient, BadRequest, Fatal. The user sees the class in words and the fix.
7. Cost: use the provider's reported cost when there is one (xAI ticks, Claude total_cost_usd, Grok's reported cost); otherwise compute from usage × prices.toml and mark it ≈. Never sum Claude total_cost_usd across resumes (it already includes earlier spend).
8. Context window comes from the provider, never a default guess.
9. Published docs beat repo docs. A flag found only in a tool's repo or user guide, not on its published CLI reference, is unverified: don't depend on it until a live check against the released CLI passes.

Tests (required)

• Recorded streams: capture real responses once (success, tool call, rate limit, context overflow, auth failure), strip keys and personal data, and commit them under the crate's fixtures folder (crates/kit-agents/src/fixtures/ for drivers). CI replays them; CI never calls a real API.
• Live tests go behind the live cargo feature and are run by hand with a budget.
• Drivers: replay a transcript through the parser, and use a fake binary (see crates/kit-cli/tests/land_cli.rs) for exit codes, kill, resume and a version whose stream changed shape, which must fail loudly with the version named.
• Run cargo test -p kitctl-providers -p kitctl-agents and the 3-OS CI before calling it done.