google/gemini-3.7-flash currently fails on the Provider API with a
zero-data-retention routing 404 for every request, which made the GOAT
live suite red regardless of provider changes. GPT-5.6 Luna is available
on every plan, accepts image input, and completes the vision request.
PR #69 added efforts for meta/muse-spark-* directly to the generated
catalog, which the drift check flags and the daily sync job reverts.
Keep src/commandcode-catalog.ts byte-identical to upstream and merge a
separate src/commandcode-catalog-overrides.ts over it at load time. A
test fails as soon as upstream publishes efforts for an overridden model
so the override gets removed instead of shadowing the CLI catalog.
Verified against the live endpoint: pi --thinking xhigh sends
reasoning_effort="xhigh" for meta/muse-spark-1.2-contributor and the
request succeeds; --thinking off sends none.
Closes#69
Every host start awaited a full catalog request before the provider
registered, costing one HTTPS round-trip (up to the discovery timeout on a
hanging connection) even when a cache written seconds earlier was on disk.
Register the cached catalog immediately and run the live refresh in the
background; the live result re-registers the provider when it arrives. A
first start without a cache still awaits the live catalog. The background
refresh is aborted on session_shutdown so print mode does not wait for it.
Closes#63
(cherry picked from commit 142191f9420fe90460cdc3cbae70e26d0c000860)
Declare the fixture's pi types inline so npm run typecheck passes without
the optional peer packages, matching the existing fixture convention.
(cherry picked from commit 0ba5827c58a0f6ccf54387454ae38d93bbead3d9)
pi routes the main chat through the registered provider, but sibling
extensions that call streamSimple from @earendil-works/pi-ai/compat with
the active Command Code model resolve model.api through the compat
api-registry, which only knows built-in APIs. On plain pi that failed
with "No API provider registered for api: commandcode-custom".
Register commandcode-custom there and delegate to the transport router.
The registry resolves no credentials for extension providers, so fall
back to the configured Command Code key when the caller passes none.
Closes#68
(cherry picked from commit 7e9659e672771c6a9223b95938e50fb2a051a0a7)
Verify display pricing against the official pricing page and the live
provider model list:
- add claude-fable-5-1 and deepseek/deepseek-v4-flash-fast
- remove minimax/minimax-m3-free and minimax/minimax-m2.7-free, which the
provider no longer lists
- end the claude-sonnet-5 introductory window; the listed rate is unchanged
- google/gemini-3.7-flash returns to list price after the promotion ended
The model catalog tests hard-coded specific model ids, the reasoning model
count, and output limits from command-code@1.32.2. Every upstream catalog
sync broke them, which made the daily catalog sync workflow fail before it
could open its PR.
Assert structural invariants over the generated catalog instead: image
models resolve to text+image, every effort entry has a reasoning flag,
reasoning without efforts yields an empty level map, output limits are
positive integers and clamp to the context length.
Closes#67
(cherry picked from commit 458e3a57892bb625a78fb5ded092627ba3cf2867)
Add the free minimax/minimax-m3-free and minimax/minimax-m2.7-free
promotional variants (free through September 5, 2026) and
tencent/hy4-preview, drop the retired stealth/ox-alpha, and refresh the
model-id and pricing snapshots to the current 62-model catalog verified
on 2026-08-28.
Qwen/Qwen3.8-Flash lists $0.16 input, $0.47 output, and $0.016 cache read
per million tokens, and z-ai/glm-5.3-flash lists $0.15, $0.50, and $0.03;
the official pricing page documents no cache-write rate for either model.
Synchronize the static model catalog with command-code@1.36.0, which adds
Qwen/Qwen3.8-Flash (reasoning efforts low, medium, xhigh) and
z-ai/glm-5.3-flash (low, high, max), the free minimax/minimax-m3-free
model, and the glm-5.3-flash output limit, and drops the retired
stealth/ox-alpha.
Align the developer-message fix with the Provider API transport router.
The fix stays scoped to the legacy /alpha/generate converter, and the
OMP advisory compat test now forces the upgrade_required fallback so the
advisory is asserted on the generate request body.
deepseek/deepseek-v4-flash-vision-exp is served by the Provider API but was
missing from the hardcoded MODEL_INPUT_MODALITIES allowlist, so pi rejected
any conversation containing an image block. Notably this also rejected images
returned by the read tool via toolResult:
Error: Selected Command Code model does not support image content in tool results
Adding the allowlist entry lets modelSupportsImageInput() return true and the
converters forward images using the current Command Code wire format.
Coverage:
- regression test for modelSupportsImageInput("deepseek/deepseek-v4-flash-vision-exp")
- end-to-end stream test for a tool-result image forwarded as a following user
image (the concrete read reproduction), not only a user-attached image
- end-to-end stream test asserting a text-only model still rejects tool-result
images before any network access
Refs: https://github.com/patlux/pi-commandcode-provider/issues/54
Fetches live account quota from Command Code alpha usage endpoints
(whoami, billing/credits, billing/subscriptions, usage/summary) and
renders a plain-text dashboard via ui.notify.
- Graceful degradation: optional endpoint transport/timeout/parse
failures degrade to null sections instead of aborting; only 401/403
are hard failures. 429 is transient, not fatal.
- Overall deadline bounds the whole command to QUOTA_TIMEOUT_MS
(per-request controllers are chained; post-deadline phases fail fast).
- OMP auth: filter unresolved $COMMANDCODE_API_KEY placeholder and fall
back to the host resolver via pickCommandCodeApiKey.
- ZDR privacy header respected on quota requests.
- Redaction reuses redactCommandCodeErrorText plus JSON-quoted
credential fields; outer-catch errors are redacted too.
- resetAt parsed from seconds, ms, numeric string, or ISO string.
- 21 hermetic unit tests wired into npm test (test:quota).
Add a fixture extension that injects an advisor-style custom message
on session start, and assert the advisory XML reaches params.messages
verbatim, in chronological position, and is not hoisted into
params.system.
Port the models phase to `omp models --json`. `--list-models` no
longer exists in current OMP (verified on 17.3.5), so the phase
failed before reaching print mode.
OMP converts custom and hook messages (advisor notes, todo reminders,
retry nudges) to role "developer" before calling the provider.
messagesToCC() only handled user, assistant, and toolResult, so those
messages were dropped before params.messages was sent to
/alpha/generate. Steering still interrupted pending tools, but the
model never saw the message content.
/alpha/generate has no developer role: the official command-code CLI
(0.32.3) only emits user, assistant, and tool messages plus a separate
params.system. Forward developer messages as user messages with
identical content in the same chronological position. Hoisting them
into params.system would turn a mid-conversation note into a global
top-priority instruction.
Patrick Wozniak