An awaited refresh could resolve before its availability pass saved. Passes
ran concurrently, and only the most recently started pass could save its
result; a superseded pass returned without saving and refresh() still
resolved. At startup, the awaited refresh could lose to a fire-and-forget
refresh started by an extension provider registration, so initial model
selection read a snapshot without that provider and fell back to another
provider's default. A login also cancelled any full pass that was running.
Availability passes now run one at a time. A request made while a pass is
waiting to start joins it, so a caller always gets a pass that saw its
changes. Each pass has a 10 s timeout, so a hung auth check or credential
store no longer blocks the caller forever (#7301). A failed or timed-out
provider check keeps that provider's previous state and is reported in
getError().
The snapshot drops configuredProviders and all; hasConfiguredAuth reads
the auth map.
Refs #8810, #9884
registerNativeProvider() did not update the auth snapshot, so the provider
stayed unconfigured until an async refresh saved a snapshot that included it.
At startup that refresh can resolve before its result is saved, so initial
model selection saw the provider as unconfigured and fell back to another
provider or warned that no models are available.
Share registerProvider's provisional entry through markProvisionallyConfigured()
so native providers with a stored credential or a models.json apiKey are
configured immediately.
fixes#9962
Generation offers active registered tools through positional system entries and
hands a tool-calling answer to parallel or sequential pi.tool tasks and a
pi.post-tools task that applies addTools/terminate and continues the run. Tool
tasks validate, run beforeTool/afterTool, record intent with the replay policy,
publish bounded, sanitized, adaptively throttled progress in pi.live.tools, and
write results with a rendered diagnostics block. Hook dispatch with conversation
and subtree scopes, prepareArguments, HarnessOptions.env, and read/bash/edit/write
tools in pi-durable/tools. The env shell now streams raw output and spills past
thresholds; its bounded capture and publisher are removed.
Durable generation imports retry helpers from pi-ai/utils/retry; examples use
pi-ai/models and pi-ai/providers/faux. Entry graph budgets and an import
footprint test guard the models entry.
The lazy OAuth entry list in the bundle build was missing openai-chatgpt,
so /login with OpenAI failed with a missing module error. Also fail the
build if any importOAuthModule() target in load.ts has no lazy entry.
Add gpt-6.1-sol to the OpenAI, Azure OpenAI Responses, and OpenAI Codex
providers with official pricing ($2 input, $0.10 cached input, $2.50
cache writes, $10 output, long-context tier above 272k). The model
rejects reasoning.effort "none", so off is unsupported and clamps to
the lowest effort.
Addresses GHSA-82fw-gwwq-j7x9 (vitest mocker path traversal) and
GHSA-3wwx-pv8p-q78v (undici 6.x permessage-deflate DoS). Both are
dev/example-only; the shipped coding-agent shrinkwrap is unchanged.
Structured bash results now hold the full output up to 1 MiB instead of the
model-facing 2000 lines or 50KB, keeping the first and last 512 KiB of longer
output, and add truncated and full_output_path.
Codemode sums the usage of models.classify() calls into its tool result and shows each call's cost. Usage of tools called through ctx.executeTool() is added to the calling tool's result usage instead of being dropped.
System One services (TypeSafe, OpenRouter, OpenCode Zen, Cloudflare Workers AI) return usage.input_tokens/output_tokens. ClassifierResult now carries usage priced from the model catalog like chat usage.
Vercel evaluation models are generated from the AI Gateway catalog and
routed through its TypeSafe-compatible /typesafe/v1/systemone endpoint.
OpenCode Zen's jev-1.13 and jev-1.13-free are routed through
/zen/v1/systemone.
Tools without a custom call renderer now show their arguments as key=value
pairs when collapsed and one key: value line per argument when expanded.
MCP tools get a server/tool title and results that collapse to 5 lines.
Submissions with request-ID deduplication and handles, the pi.live run
control and presentation document, the built-in pi.generation task
(system prompt preparation, streamed durable partials, retries, deferred
polling, abort), typed entry tokens, stream options and retry policy,
TaskRuntime reads, Tx.settleSubmission, registry conversation setups,
Storage.scanSubmissions, Harness.inspect, and byTaskId attribution.
Waiting calls enable scheduling. The scratch tour is split into
test/examples.
Entries of only +name/-name modify the inherited tool selection, so
"defaultTools": ["+codemode"] enables codemode without repeating the
defaults. Project entries layer on top of user settings. Document how to
enable codemode without MCP and how to use classifier models from it.
Anthropic, OpenAI Codex, OpenRouter, and Radius now use one loopback callback server in pi-ai. MCP sign-in renders the same browser page through a new renderPage option on the pi-mcp callback server.
Built-in extensions (mcp, llama.cpp, codemode, tool-search) are now extension resources like files: the package manager resolves them as builtin:<name>, the extensions setting disables them with -builtin:<name> (per project too), pi config lists them, and --no-extensions turns them off unless loaded with -e builtin:<name>. They load after project trust is resolved and after file and package extensions.
Built-in extensions and tools are named builtin:<name> in errors, diagnostics, RPC source info, and bug reports.
Refreshes hold a per-server lock file from reading the tokens to saving new ones, and use tokens another process already refreshed. Rotating refresh tokens (Cloudflare) were lost when two pi processes refreshed at once, leaving the server needing sign-in. Closing a connection waits for a running refresh so shutdown does not drop rotated tokens.
pi.live turn control and generation presentation, built-in entry kinds,
pi.generation checkpoint and phases, typed entry tokens, built-in tasks in
the registry, TaskRuntime reads, stream options and retry policy config,
Harness hook for scheduler-written outcomes, and createConversation without
input.
Adds fullscreenWheelScrollLines ("auto" or 1-100). Auto keeps one line
per event in local macOS terminals and scales with wheel velocity
elsewhere, capped at six lines per event.
closes#9758
This adds codemode and MCP support to pi. Codemode (packages/codemode) runs
model-written JavaScript in a QuickJS wasm VM inside a worker. From there the
script calls pi's tools as async functions, uses a per-session store, and can
read the model catalog and run classifiers. The codemode tool in coding-agent
is compatible with Codex's exec tool, and codemode.mode works like Codex's
tool modes (on or only). Large results are written to a JSON temp file. MCP
(packages/mcp) is a standalone client for stdio and streamable HTTP servers,
with OAuth.
The core only gets general mechanisms; codemode, tool_search and MCP are
built-in extensions that use them. Other extensions can replace these
built-ins. Tools declare an exposure: direct, model-only, codemode, deferred
or hidden. prepareLoadout() lets an orchestrating tool change which tool
declarations the model sees. ctx.executeTool() runs nested calls through
validation and the tool_call/tool_result hooks. Those calls emit events with
parentToolCallId and are recorded as bounded nestedCalls on the parent
result, which compaction and HTML export use. Tools can also return
structuredContent with an outputSchema, and can report errors with isError
instead of throwing. MCP servers come from mcp.json (global, or per project
after the project is trusted) or from pi.registerMcpServer(). They are
managed with /mcp and pi mcp list|login|logout, and their runtimes load on
first use.
Closes#10040
Theme previews re-render the whole transcript, and streaming re-renders
every frame, so per-line and per-session work dominated in long sessions.
- visibleWidth: fast path for ANSI-styled ASCII, allocation-free escape scanning
- Box: compare unpadded child lines and skip a second width measurement
- Markdown: reuse parsed tokens across invalidation
- footer: cache usage totals and context usage per session state
- bash results: cache the collapsed preview including the expand hint
- sanitizeBinaryOutput: single regex replace
Together no longer lists Kimi K2.6, which broke type checks against the hydrated catalog. Tests use Kimi K3 and the Together default model now points at it.
Adds defineTask, registry-resolved phase handlers, and a scheduler that
decides every task transition in one callback on the Session line:
reservation with migration, dependencies, precedence rules in a step
before each phase, handover to replaced definitions, direct-task abort
(mark, signal, join, fresh abort invocation), orphaning of blocked tasks,
open reconciliation, and close that joins invocations without writing
outcomes. runtime.commit() callbacks return the typed next state; Tx no
longer exposes setTask. Adds Harness.resume, getTask, waitForTask,
abortTask, and task-aware waitForIdle; RegistryReader.subscribe wakes
the scheduler.
This adds experimental support for virtual models. A virtual model is a catalog entry that an extension registers with pi.registerVirtualModel(). It does not talk to a provider itself; for every request it picks a physical model and thinking level. The motivation is to make routing policies pluggable: plan on a strong model and implement on a cheap one, pick a model with a classifier, or move to a larger window when the context grows. Today this needs model switching by hand or hacks in prepareRequest.
The selection stays virtual: model_change records it, resume restores it, and each response records the physical model and thinking level that produced it. Router state can be stored on the session branch, so it follows /tree and forks and survives compaction. Context limits, compaction and image sizing use the physical model a request was routed to. examples/extensions/jev-router.ts shows a full router that uses the Jev classifier to pick Sol or Terra for planning and hands off to Luna after the first edit. docs/virtual-models.md describes the API.
The agent runs every tool call in the final assistant message, including
calls whose output_item.done never arrived. Those can carry cut-off or
mixed-up arguments. With llama.cpp, which omits output_index, two parallel
calls became three (echo a, echo a, echo b), two of them with the same id.
processResponsesStream now fails the stream when a completed response
still has an unfinished tool call.
fixes#9974
Read Finder file URLs before icon image data on macOS. Quote paths in
bash mode and reject terminal control characters before rendering.
Fixes#9999
Co-authored-by: 1dustycy <onedustycy@gmail.com>
Implements Pico5 Package 13: Harness.open over a Session with an application-owned
registry (tools, tool wrappers, hooks, tasks, system prompt sections; batched
publication, stable keyed order), lazy root creation with init, atomic
create/fork with init, conversation-bound commits, fork-aware entries, model
context derivation, and the built-in ConversationConfig document.
Fixes cached documents skipping migration when accessed with another definition
version. Rewrites the Pico5 spec and handoff to the registry/system-prompt
design and removes the superseded live-registries proposal.
Storage reports replayed deltas after the selected base; Session tracks the
count across adopted writes so checkpoint predicates can bound replay.
Also adds internal reserved-root creation and conversation-bound commit
defaults for the upcoming Harness.
Mistral models with effort levels now use reasoning_effort based on
their thinking level map from models.dev instead of a hardcoded id list.
GLM 5.3 no longer gets prompt_mode, GLM 5.2 can use max, and off sends
the model's off value when it has one.
fixes#9678
Add the llama-cpp-classify classifier API. Each question becomes one chat prompt with single-token answer labels; llama-server's pre-sampling next-token log-probabilities give the softmax over those labels. Missing labels are retried with deeper readouts and then reported as errors.
David Brailovsky