Moves chord2 (contracts, objects, cells, roots, plugins, hot reload, memory/IPC/WebSocket links)
from p2 into packages/chord, with its tests (now on vitest) and docs. The root export is the new
runtime plus the shared base (Context, JsonValue, JsonRepresentation, Draft, copyJson, isJsonValue,
replicatedState and its types). ./node now holds the host manager, child control, process locks,
and nodeLoader. ./context and ./delta are unchanged; ./bundler is removed.
replicatedState stays a standalone module with the same exports. It no longer exposes the exact-op
subscription the removed service provider used, so listener failures of mutable states are always
reported asynchronously. The package depends on typebox instead of esbuild. The wire format stays JSON.
Deletes packages/client, packages/protocol (CBOR codec), packages/server, coding-agent's
src/experimental (except source-resolver.ts, which source runs still preload), src/cli/experimental,
src/client, the example plugin, and their tests. coding-agent no longer depends on @earendil-works/chord.
The installer lock drops esbuild, so its platform-package sanity check goes too.
A wait is rejected when it closes a cycle in the wait-for graph: waits of
tasks that are not abort-marked, and holds of live ordinary owned work. The
rejection names the cycle. Replaces the owner-chain rule, so a background task
may wait on its conversation's owner. Approach and tests from #10533 by
@odysseus0.
fixes#10411
On each OpenRouter refresh, combine the catalog (built-in models plus a newer pi.dev catalog) with GET /models/user for the user's key: keep only the listed chat models and take their context size, max output and prices from it. Image and classifier models are kept. The combined list is saved and replaces the OpenRouter catalog, including offline.
- ai: add personalizeOpenRouterCatalog and RefreshModelsContext.baseUrl, so the request follows a base URL configured in models.json (regional endpoints)
- ai: share OpenRouter price conversion with the generator; negative prices (router models such as openrouter/auto) become 0
- coding-agent: add withPersonalizedCatalog next to withRemoteCatalog
- docs: OpenRouter section in providers.md
fixes#10353
- ToolRegistration.deferred: not offered to the model until modelTools names it
- modelTools gains { remove?, add? }; control.addTools writes names to add, so
loaded tools follow the others in load order and a load only appends
- Agent.deferred lists deferred tools not loaded yet
- @earendil-works/pi-durable/tool-search: the coding agent's tool_search (BM25)
- the codemode section leaves deferred tools out
joinAdjacentText checked endsWith on the growing joined string, which
makes V8 flatten the concatenation on every item. Collect each text run
in an array, check the separator against the last non-empty part, and
join once. Same fix in the durable code mode copy.
closes#10678
Runs started by pi.sendMessage(..., { triggerTurn: true }) skipped before_agent_start. After their first tool call, the next-turn refresh fell back to the base prompt options and patched away the sections handlers had added, changing the system prompt mid-run and breaking the provider prompt cache.
These runs now go through the same setup as prompt(): compaction check, before_agent_start, tool loadout, nextTurn messages, handler messages, and the system patch. BeforeAgentStartEvent is a union on messageType ("user" or "custom"); custom events carry the starting message.
The session counts as busy from the start of setup, so messages sent while handlers run go into the starting run instead of starting a second one, and an abort requested meanwhile aborts the run once it starts.
closes#10267, closes#5581, closes#10639, closes#10289
partly fixes#10606
pi-codemode no longer exports MCP_TYPESCRIPT_PREAMBLE or mcpStructuredContentSchema(); output schemas render as their own types, as in the coding agent.
Tools that reach other tools declare it with reachesTools(input); MCP uses it to wait for the servers a call needs and to decide which codemode or tool_search tool to activate, instead of parsing codemode scripts and comparing parameter schemas. Namespaces get aliases, replacing codemode's MCP-specific __ suffix rule.
pi-codemode no longer knows MCP: MCP_TYPESCRIPT_PREAMBLE and mcpStructuredContentSchema() are removed, and the MCP result schema describes itself, with the server's local $refs relocated so they resolve as before.
- structuredOutput is always the tool's own value; a schema tool's error text
without one becomes a tool_error diagnostic on nested results
- code mode: a schema tool's error result with its value resolves (bash nonzero
exit), schema-less tools resolve with their text, as in the coding agent
- codemode prompt section with tool declarations by namespace within an inline
budget; searchTools/describeTool/describeNamespace from the start data
- result layout of the coding agent: header, text markers, console block,
truncation within max_output_tokens and the tool's output limits, spill and
image files in a recorded output directory (a rerun overwrites them)
- models.*: catalog globals; classify/generateImages as hidden internal tools
(nested calls with usage), at most four at once
- Lark grammar of the script source as constrainedSampling
A host that closes its CodemodeSandbox before the Harness, as on a graceful
shutdown, made the running code mode call store "Script error (aborted):
Sandbox closed" as its result, so the script never resumed. A run the sandbox
aborts without the call asking for it now stores nothing and ends with the
call's invocation; the next process replays the script.
details.calls (CodemodeToolDetails): each tool call the script made, in call
order, with status ok/error/cancelled, duration, and cut arguments and error;
calls still running or not yet launched when the script ended are cancelled.
Written once with the result: live progress is pi.live.nestedTools.
bench:codemode compares a code mode script with a plain tool making the same
nested calls: 1,000 parallel calls 0.82 vs 0.78 ms/call, sequential 1.52 vs
1.10 ms/call (two more commits per call: the journal entry and the count).
A codemode tool whose scripts call the conversation's tools as nested calls and
survive a restart. The script runs deterministically (pi-codemode start()); each
result is journaled before it is delivered. A rerun starts the script again from
the stored start data, delivers the journaled results in order (nested calls
reattach by key c<id>), checks every recorded call count, holds calls the
journal lacks until replay ends and checks them against earlier attempts' keys,
then continues live. A call launches only once the count of the drain that made
it is committed, so a rerun can prove which calls are new. Divergence, a changed
module version or sandbox fingerprint, and a passed deadline end the call with a
stored error; a decided outcome is returned without a VM. store()/load() keep
values per conversation, rewindable and forked as of the fork's entry.
pi-codemode: CodemodeRun.finished, set when the outcome is decided.
Property comments now carry constraints the TypeScript type cannot
express (integer, bounds, lengths, pattern, format, default, ...) as
@ tags, read from nullable anyOf branches too. Types over the size
limit are rendered again without tags before becoming unknown.
closes#10707
For durable code mode: a host can rerun a script after a restart and get the
same calls with the same arguments. start() runs the script with a virtual
clock (startedAt, then each result's time, never backwards), randomness from a
seed, UTC, no WeakRef/FinalizationRegistry, and the run's own tools and
globals; every call goes to onCall with the VM's call id, and the host answers
with run.deliver(seq, now, id, outcome). The VM applies results strictly in
seq order whatever the transport does, reports its fingerprint (engine
versions, prelude hash, limits) before the script runs, the calls issued after
each result (onIdle), and the result it was processing when the script ended
(result.during). Rejections can carry properties for the script's Error.
execute() is unchanged.
Preserve terminal reply strings until their terminator while retaining bounded recovery for ambiguous legacy Alt-key input.
Related: #10256, #10657. Missing OSC prefixes and fragmented DA1 replies remain separate issues.
The delta tracker planned every structural array change with a walk over all
base elements; inserts and removals that keep the remaining elements in order
now get the same plan from the changed ranges alone. applyImmutable built an
appended array as a copy that the insert then grew, a second copy; an append now
builds it with one concat. Ops, values, immutability and errors are unchanged.
32,000 one-item appends, one commit each: 11.3 s -> 1.4 s
(npm run bench:document-append in packages/durable).
Programs calling a tool with structuredOutputSchema got isError and no text when
its error result omitted structuredOutput; they now get its bounded output,
collapsed like a schema-less tool's, so a code mode script sees the error text.
A thrown error's message stays in its tool_error diagnostic.
Radius omits providerThinkingLevel from pi-messages done events, so replayed
signed thinking has no effort marker and Anthropic drops it (block_binding
drop_block). Compares follow-up input tokens with and without the thinking
block against an Anthropic control.
The process group in front of the pty when it is not the program's own (a job its shell runs), with the leader's name on macOS and Linux. Exposed as RemoteTerminal.foreground().
The daemon gets pty, ptyWrite and ptyResize: raw terminal output as events, input in arrival order, SIGWINCH on resize, and the process group killed on cancel or a lost connection. openTerminal() is the client, localDaemon() the daemon for this machine. Unix only; Windows answers ENOTSUP.
A generation waiting on a blocking threshold compaction sent its request
whatever the outcome. When summarization failed, the context was still above
the blocking threshold: a session at 996K of a 1M window got an empty length
answer on every turn and nothing showed the error. The generation now fails
with model_error (Compaction failed: <error>) when its compaction failed,
faulted, was orphaned, or was aborted directly; a compaction that found
nothing to compact or was declined still lets the request go ahead.
pi.live.compactionFailure keeps the newest failure (task, reason, error)
until a compaction produces a summary. The agent events' snapshot carries it,
and compaction_end carries error for a failed compaction.
Replayed thinking is often not part of the live context (for example,
radius drops it, and Anthropic drops blocks that do not match their
request), but serializeConversation included all of it. A Pi Mobile
session at ~996K of 1M context tokens produced 1.38M-token summarization
requests that failed on every attempt. Providers only return summaries of
the reasoning anyway; visible text, tool calls, and tool results record
what happened.
* feat(coding-agent): use pi.dev configuration schemas
* refactor(coding-agent): separate schema URL from output path
* docs(coding-agent): document published schema URLs
- Input submissions take entry: { kind, data } (InputEntry): placed as that
kind with that data and the content as user message, instead of pi.user.
Kinds must be non-empty strings outside pi.*. Queued inputs carry it in
their pi.inbox item.
- Writes with only kind and data (no model, head, edits) are appended and
settled done at once, busy or not; only a queued reset holds them back.
- Experimental hosts render such inputs as user messages and title by them.
closes#10407
createNodePhotonImages() now runs Photon in a worker thread, so preparing a large photo (a second or two) no longer blocks the event loop: in the test a 6-megapixel image blocked the loop over 200 ms inline and under 100 ms with the worker. The worker keeps the process alive only while it prepares an image, exits after 30 s without images to free Photon's WebAssembly memory, and a worker that cannot start falls back to the calling thread. Options: { worker: false, idleMs }.
The Session kept every document it loaded or created in memory for the life of the process: every conversation and
session document ever touched, built-in and third-party. Now a cached document that no watch or document state
observes leaves the cache once it was not read or written for `documentRetentionMs` (HarnessSettings and
createSession options; default ten minutes, 0 drops it right after use, Infinity keeps documents; anything that is not
a number of milliseconds means the default), and loads from Storage again on its next access. An observer pins the
incarnation it observes and releases it when that incarnation retires.
Eviction runs on the mutation line, after every commit outcome (published, no-write, rejected, failed settlement) and
from one unreferenced timer for the next expiry, so a transaction never loses a tracker it holds. Where timers cannot
be unreferenced only commits evict. Cache bookkeeping is contained: a throwing clock or setting never changes an
operation's outcome. A hot document moves in the eviction order at most once per granule (a tenth of the retention,
up to 1 s), keeping the fan-out bench at HEAD's speed.
`npm test` runs the durable suite a second time with every unattached document evicted right after use.
5,000 conversations kept 26,000 documents and about 3.8 KB each; with a 50 ms retention, after waiting, none
(bench:document-cache).
The scheduler kept every conversation edge it saw and, for conversation owners, chain loads, and owners of known
tasks, the ended tasks above them, for the life of the process. Now an owner edge or an ended task's ownership is kept
only while something is listed below it in #below: only the chains above live work are known, whole. Unneeded nodes
are queued and swept at the end of each publication, or by a sweep job on the Session line after operations that
publish nothing (rejected commits, reads), never during an operation, so a commit that creates work below a chain its
own callback loaded keeps it. Later work in an old conversation loads its chain from storage again.
Removes #conversationOwners, #pinned, #knownChildren, and #forget. A queued input triggers the withdrawal scan only
while some task has cancellation intent, and chains are loaded only for queued inputs.
10,000 ended subagent cycles under a live parent: 10,000 settled tasks and 10,001 edges before, 0 and 1 now
(bench:scheduler-memory); a later message into an old subagent conversation reads 7 records instead of 5.
Session, Harness, and runtime reads such as snapshot(), documentState(), watchDoc(), and getTask() may queue on the mutation line a commit callback holds; on a cache miss they never resolve. Documented on Session.commit, TaskRuntime.commit, and in the spec's pitfalls.
Node 24/26 --watch posts watch:import/watch:require messages on the
worker channel (nodejs/node#65044), which failed the run as a broken
bridge.
closes#10725
NestedCallsDoc (one index per caller, a key per nested call, copied on every admission) and NestedResultDoc are
replaced by NestedCallDoc (`pi.tool.nested-call`): a task-scoped family of the caller keyed by the call's key,
holding { taskId, result? }. Admission creates the member; the nested call's terminal commit adds its result;
executeTool() reads it by key. No commit rewrites another call's document.
TaskRuntime.ownedTasks(context) returns copies of the live tasks a task owns directly, whichever invocation
created them, from the scheduler's index. A tool call settling aborts the live nested calls among them, which
the removed index used to list. A task memo marks a tool that made a nested call in any attempt, so settling a
leaf still skips scanning pi.live.
8000 parallel nested calls: 9.2 s -> 2.7 s, linear (bench:scheduler-fanout).
Every reconcile pass walked up from every live task to the root to find cancelling owners, and owned live work,
idle scopes, chain loads, reservation, finalization, and failFast checks each scanned all live tasks. A chain of n
held tasks cost about n^3 steps: 500 took 12 s, 20,000 never finished.
The ownership semantics are unchanged; in-memory indexes, rebuilt at open, hold each pass's candidates:
- #below: the ownership tree downward around live work; owned live work, idle scopes, and conversation aborts walk
down from one node. Ended tasks stay while walks pass through them (conversation owners, chain loads, owners of
known tasks), so loaded chains stay loaded.
- #intent: cascades walk down from tasks with cancellation intent, not entering another such task, so the nearest
cancelling owner still decides the reason; upward checks return at once when it is empty.
- #unloaded: only tasks whose chain is not loaded are loaded, and count as inside every scope until then.
- #finalizeChecks: finalize is a worklist in rounds from the nearest live holder of ended work.
- #runnable and #waiters: reservation and failFast candidates.
No walk recurses.
Chain of 500 held tasks: 12.3 s -> 50 ms; 20,000: 1.5 s. 8000 parallel nested calls: 18.8 s -> 9.2 s
(bench:scheduler-fanout); the rest is document copying of the nested-call index.
Without metadata there is no issuer to compare an RFC 9207 iss parameter with, so the check was skipped and a malicious MCP server could receive a code from another authorization server. Reject the response before the code exchange instead.
Reported by Alexios Stefanopoulos (@AizenvoltPrime).
Split the host into transport-independent execution logic and a VM driver, so the VM can run in a worker thread (CodemodeSandbox, unchanged), on the calling thread (InlineCodemodeSandbox), or in another isolate over RPC (RemoteCodemodeSandbox).
Add the /portable entry without Node imports and the /cloudflare entry with CodemodeSandboxDurableObject, meant to be deployed as its own Worker so scripts never block the caller's thread. Scripts that compute without awaiting a tool are bounded by an interrupt budget, now also polled while draining promise jobs. examples/cloudflare shows the setup.
read returns an image file as one image block (programs get an ImageContent) plus an info diagnostic,
instead of an unsupported_image error. Limits are the conversation model's inputLimits.images.resize,
by default 2000x2000 pixels and 4.5 MB of base64.
- Without a processor: PNG, JPEG, GIF, and WebP within the limits pass through undecoded; the byte limit
is checked before reading the file, dimensions from the header. Larger images and BMP are error results.
- With createCodingTools({ images }): the processor orients by EXIF, converts BMP to PNG, and shrinks to
fit. ImageProcessor.prepare is async.
- @earendil-works/pi-durable/images: createPhotonImages(wasm) on @silvia-odwyer/photon 0.3.3 (web build);
/images/node and /images/cloudflare load the wasm. The root and ./tools entries never reach it (enforced
in check-entry-graphs).
- An image must be read whole: one that changes while read is read again, unlike a growing text file.
- A model without image input gets a diagnostic that it sees a placeholder.
Nested tool calls: ToolExecutionApi.executeTool() runs another tool as its own pi.tool task owned by the caller,
keyed for replay-safe reattachment. Results for programs live in the caller's NestedResultDoc until it settles;
pi.live.nestedTools shows each call with the arguments it runs with and a summary. TaskOptions.abandonOnRestart
aborts work its creator awaited only in memory before it runs again after a restart.
Structured output: structuredOutputSchema and structuredOutput, validated; a tool without a schema gives programs its
bounded output (one text as a string, one image as itself, else the list). ToolExecutionResult.content is now output.
callers and modelTools decide who may call a tool.
Storage errors: any error a Storage method throws fails the Session, which closes itself; later calls and waits reject
with SessionFailed, watches end with session_failed, and Session.closed settles. Nothing retries in the Harness.
StorageRejected is removed; StorageRequestError marks invalid reads that fail only their call. Listener and onReport
errors are contained and reported.
Also updates scripts/issues/triage.ts and the vacation experiment for the content -> output rename.
The nested worktree dedup compared the context file's realpath with the
main repo's file, so a worktree AGENTS.md symlinked to the main repo's
copy was skipped along with it. Compare the walked directory's realpath
plus the file name instead.
fixes#10681
Remove hardcoded Opus 5.5, Sonnet 5.5, and Haiku 5.5 fallbacks now that models.dev lists them. Fix the Sonnet 5.5 test to expect the official cache read price of 0.1.
Long Mistral streams (e.g. extended thinking) were aborted once timeoutMs
elapsed, even while data was still arriving. The timeout now covers only
waiting for response headers; body stalls are left to the HTTP client idle
timeout.
fixes#10609
Editor and Input wrap their drawn cursor in fake cursor markers. The TUI
resolves them to reverse video, or drops the focused one when the hardware
cursor is shown, so terminals that extend edge cell colors into the window
padding no longer stretch the drawn cursor at the first or last column.
Also keep CURSOR_MARKER when an overlay ends directly left of the cursor.
