Assert the empty catalog path recovers live models and writes a cache once discovery is available again, including the real pi extension entrypoint used by /reload.
Resolve the Command Code model cache through the host's getAgentDir helper so pi, OMP, and PI_CODING_AGENT_DIR use their own agent state directory instead of the official Command Code client directory.
Persist the last valid Command Code model catalog and use it when live model discovery fails. Keep first-time offline startup non-fatal, surface clear warnings, and cover cached model selection with unit and pi integration regression tests.
The model fetch at startup throws and blocks pi from starting when
there's no network. Catch the error and register the provider with
an empty model list instead.
Use the explicit $COMMANDCODE_API_KEY environment variable reference expected by newer pi versions while keeping compatibility with the legacy placeholder handling.
Align provider maxRetries default with pi (0), treat maxRetryDelayMs 0 as
no cap, keep per-attempt timeout through stream reads, and fix gitleaks
test fixtures.
Add retry mechanism driven by pi settings.json retry.provider config
(timeoutMs, maxRetries, maxRetryDelayMs).
HTTP-level retries handle 429/5xx with exponential backoff and jitter,
respecting Retry-After headers (seconds and HTTP-date formats).
Stream-level retries handle cases where the API returns 200 OK but
sends an error event in the stream body. Retries only when no content
has been emitted yet.
Per-attempt timeout via AbortController with automatic retry. Clean
abort propagation through the retry loop.
DeepSeek V4 Flash cache-read rate was /bin/bash.01/1M but docs list ~/bin/bash.028.
Added xiaomi/mimo-v2.5-pro and xiaomi/mimo-v2.5 to MODEL_COSTS so they
don't display as zero-cost models (pricing TBD, currently set to 0).
If upstream omits reasoning-end and starts emitting text-delta or
tool-call events, the thinking block remained open causing
thinking_start -> thinking_delta -> text_start -> ... -> thinking_end
ordering. Now endThinking() is called at the start of text-delta and
tool-call handlers so thinking_end always precedes text_start or
toolcall_start.
Adds 2 regression tests for the missing reasoning-end edge case.
Reduces node_modules from 312MB to 154MB. The import is type-only
(import type { ExtensionAPI }) so the 163MB package was installed
but never executed at runtime. End users get the lean install;
contributors still get it via devDependencies during development.
- Removed unnecessary blank lines and improved formatting for better readability in index.ts and converters.ts.
- Updated package-lock.json to downgrade several dependencies for compatibility, including @protobufjs/eventemitter, @protobufjs/fetch, and others.
- Enhanced error message in core.ts to clarify configuration options for the Command Code API key.
- Added a test to ensure the correct handling of environment variable values in test-stream.ts.
