Changelog

Unreleased

Add repository commands for testing the current checkout either in a logged-out, automatically cleaned-up pi environment or with existing credentials and only Command Code models enabled.
Use the host-provided pi-ai and pi-coding-agent core packages instead of installing private runtime copies, including for local and out-of-store development checkouts.

Allow pi to start when model discovery is unavailable. The provider now caches the last successfully fetched model catalog so previously discovered Command Code models remain selectable offline; a first offline start without a cache keeps Command Code unavailable until /reload succeeds.

Fix Oh My Pi extension validation by avoiding the missing calculateCost export from OMP's legacy pi-ai shim.
Add a regression test that locks the local Command Code cost calculation to pi-ai's upstream calculateCost behavior.

Use the explicit $COMMANDCODE_API_KEY provider registration syntax expected by newer pi versions, removing the startup deprecation warning while keeping legacy placeholder compatibility.
Refresh development dependency lockfile entries to resolve npm audit findings for tsx/esbuild and protobufjs.

@plumj-am — fixed the pi provider apiKey deprecation warning.
@cad0p — reported retry/deprecation-related issues that helped validate the current behavior.
@bl4zee1g — reported provider availability concerns that prompted additional local/live validation.

Add retry mechanism for transient HTTP errors (429, 5xx) and stream-level errors, configurable via pi settings.json retry.provider fields (timeoutMs, maxRetries, maxRetryDelayMs). Supports exponential backoff with jitter and Retry-After header.

Add OMP (Oh My Pi) provider compatibility: support ~/.omp/agent/auth.json auth path, handle OMP's env-var-name-as-apiKey quirk, convert OMP system prompt arrays to text.
Close open thinking blocks before starting text or tool output to prevent event ordering issues when upstream omits reasoning-end.
Correct DeepSeek V4 Pro discount as permanent (no expiry), not time-limited.
Correct DeepSeek V4 Flash cache-read rate to $0.028/M and add xiaomi/mimo models to pricing table.
Upgrade pi dependencies from @mariozechner 0.72.0 to @earendil-works 0.75.5.
Move pi-coding-agent to optional peerDependencies.

Stream reasoning-delta events incrementally instead of buffering the full thinking block until reasoning-end. Emits thinking_start, thinking_delta, and thinking_end events as they arrive so the UI can show reasoning in real time.
Close open text blocks on reasoning-start and reasoning-delta so thinking and text never overlap in the output.
Add live display pricing (MODEL_COSTS) for known Command Code models. Cost falls back to zero for models not yet in the price table until the Provider API exposes pricing directly.
Fetch models from the Command Code Provider API at startup (inherited from upstream 0.1.1) and overlay the static cost table.

Align Command Code generate requests with CLI 0.27.2 headers and payload shape.
Support official Command Code CLI auth files using the command-code credential key.
Handle reasoning-start and ignore streamed tool-result events.
Cap generated max_tokens by the selected model and the Command Code output limit.