pi-ai infers a request's shape from the provider id and baseURL, and for
an endpoint it does not recognize it answers as though it were OpenAI
itself. A hand-declared route is by construction such an endpoint, so a
model declaring reasoningEfforts sent its system prompt as the developer
role with no configuration able to say otherwise — a gateway rejecting
that role could not be connected at all. Writing the switch anyway
validated, persisted, and was then dropped, so the misconfiguration
looked applied.
Three drift gates classify all thirty upstream compat fields as offered
or withheld, keyed by `keyof` so a pi-ai upgrade fails the build until
the new field is classified. Twenty are offered: what a private URL
cannot imply. The rest stay withheld because pi-ai's installed catalog
sets them for a named vendor.
Protocol applicability is now per field rather than per block, so
supportsDeveloperRole reaches an openai-responses route and the
anthropic-messages switches reach theirs. A compat key no protocol
declares, or one a gate withholds, is refused where it is written.
Fixes#2646
Refs #1976
Every image in history is base64-inlined into every request, so a long
session's request body grows until a gateway request-size cap rejects it
with 413 and every retry resends the same oversized body, permanently
wedging the session. Each provider route now carries maxRequestImageBytes
(default 24MiB): when the accumulated base64 image payload exceeds it,
the oldest images are replaced by a fixed model-facing placeholder until
the request fits, so the newest images survive and the session keeps
completing requests. 413 and request-body-cap wording now classify as
INVALID_REQUEST instead of the generic PI_AI_ERROR.
Fixes#2644
An image with a side above the deployed routes' 2000px many-image bound
could be durably committed by read_image, ride every later request, and
permanently fail the session with provider 400s. Admission now enforces a
configurable maxImageDimension (default 2000) during the full decode, so
read_image surfaces a recoverable tool error naming the limit instead of
poisoning durable history; the Web composer gets dedicated copy for the
new IMAGE_DIMENSION_TOO_LARGE reason.
Fixes#2626
master introduced AttachmentStore.saveImages as the batch admission
(count/aggregate-byte/media-type limits, validate-all-before-save,
ordered commit). admitEncodedImages narrows to the shared wire entry:
canonical-base64 enforcement plus delegation to saveImages, keeping one
home for batch policy while both wire endpoints (prompt RPC and the
command executor) still call one function. Test doubles gain saveImages;
batch-limit error texts follow saveImages' wording.
A claimed slash command consumed only the text half of the composer
submission: /goal with reference images executed, cleared the draft, and
silently stranded the images in the rail. Model-visible attachment intent
had no route through the command plane.
The submission envelope is now modeled end to end. CommandDefinition
input.images declares acceptance; the declaration rides the descriptor to
every client, onto the minted CommandClaim, and into the input machine's
claim snapshot. commands.execute carries the submission's base64 images
and enforces the declaration in the executor: non-declaring commands, a
missing attachment store, and exceeded batch limits settle as logged
error results before the handler runs. Admission reuses the attachment
package's new admitEncodedImages, extracted from api-proxy's prompt path
so both wire endpoints share one limits/validation/commit sequence.
Producers own model visibility: /goal submits one user followup (image
blocks + a fixed reference line) after a successful create/edit so goal
rounds read the images from session history; /plan folds them into its
steered message. Grammar misfits (/goal pause, bare /plan, /plan off)
return direct errors and the composer keeps the images.
On the client, enter adjudication carries a SubmitEnvelope and every
command route that cannot consume images throws a localized refusal that
renders as one composer notice with draft and images retained; the
claimed pre-gate applies the same copy. An accepting claim serializes the
draft images, forwards them to commands.execute, and clears plus releases
them only on a success outcome.
The assembled web test roster gains the ui-input-trigger and ui-commands
plugins, mirroring the shipped composition, so slash submissions exercise
the command plane; a new keyless snapshot pins the refusal banner and the
accepting /goal flow over the built client graph.
Replace the legacy reference.* API Proxy domain with @Remote methods on the
owning services, following the typert gateway design master adopted on
2026-08-02 (message-feedback and plugin-inventory precedents):
- FileReferenceService and SessionReferenceResolver extend TypertRemoteService;
fileReferences/list and sessionReferenceResolver/candidates are unary Remote
methods cancelled through the reserved trailing signal, and the candidates
face attaches each candidate's canonical mention under the configured limit
- move the wire types to type-only ./types subpaths (FileReferenceCandidate,
SessionReferenceMentionCandidate) and export ./typert plus ./remote artifacts
- mount both contributions in the api-remotes client assembly; ui-reference
consumes ctx.remote instead of connection.api.references and registers zh/en
locale dictionaries for its sections and labels
- delete the reference.* routes, schemas, map rows, client stubs, and fixtures;
the connection fixture serves the Remote endpoints instead
- release deliverPrompt admission listeners when the agent is disposed with the
prepared prompt still pending, and cover the reference-* RpcError codes in
the schema spec
- add the missing tsconfig paths for the /grammar and /types subpaths (clean-
tree vitest could not resolve @deepseek-ai/dsh-file-reference/grammar)
- regenerate the cordis catalog, capability seams, and event matrix; update the
owning bilingual READMEs, Agent Notes, and the reference-composer golden
Resolutions:
- docs/module-graph.md, docs/config-catalog.md: generated files. Regenerated
with gen-module-graph and gen-config-catalog on the merged tree, then carried
the new package's entries into the Chinese sides and re-recorded both
pairings. Each side now differs from master by exactly the
code-runtime-python rows.
- scripts/verify-package-readme-model-experience.ts, tsconfig.host.json: master
renamed packages/bash -> packages/shell, packages/pty -> packages/terminal,
code-runtime-worker -> code-runtime-worker-thread and agent-tool-mode ->
agent-tool-presentation. Kept master's names and re-added this branch's
code-runtime-python entry.
Adapted the package to conventions master introduced while the branch was open:
version 0.1.0-rc.6 with publishConfig.access "public" (the release-member rule
check-workspace-constraints now enforces), the invariants project reference
moved to packages/runtime-diagnostics/invariants, and the README companion link
retargeted to code-runtime-worker-thread.
The Python lane never compared what the minimal composition shows the
model: the mock model only asserted system-role messages, and the
advanced snapshot tokenizes the assembled system prompt and tool
schemas.
The sdk-minimal scenario now records model-visible.json — every model
request's advertised tool schemas verbatim and its message list, with
system and user text kept and assistant/tool payloads reduced to call
identity so the expected output replays on macOS and Linux. It excludes
the dynamic runtime-context snapshot, which the same composition emits
on macOS and not on Linux (#2488).
AGENTS.md and the testing policy name both SDKs as independent
projections of the agent loop, session lifecycle, and SessionEventMap.
A max-tokens response that included a tool call persisted assembler-transformed
content next to replay metadata projected from the untransformed native message,
so the next request died in history reconstruction with INVALID_REPLAY_STATE and
the session stayed permanently stuck.
Write side: the finish chunk's replayState becomes a typed ReplayEnvelope —
opaque response-level metadata plus optional per-block entries aligned with the
emitted block sequence. BlockAssembler computes one keep/drop decision for
blocks and entries together, so stored metadata always describes stored content
and retained blocks keep their signatures. pi-ai splits its state into a
version-2 response half and per-block signature entries.
Read side: durable content is authoritative. toPiAssistant degrades any
unusable state — foreign kind, other versions (including the flat v1 form
already on disk), malformed metadata, or content/block mismatches — to the
existing provider-neutral conversion with an onReplayDegrade diagnostic instead
of failing the request, which un-bricks sessions poisoned before this change.
Covered by assembler and replay unit tests, an agent-loop continuation
regression, keyless real-composition continuation tests (native pruned-envelope
replay and legacy flat-state degrade), and the authored keyless snapshot
scenario max-tokens-continue through the assembled ACP app.
tool-bash-persistent overwrote the backend's PS1, so terminal-bash prompt
readiness never matched and every send degraded to the 3.5s silence tier
(idleSilenceMs + handoffGraceMs) under production defaults.
The controlled PROMPT_COMMAND now re-asserts PS1 before every prompt, so an
in-shell override never survives to the next prompt. The tool initializes
with stty -echo alone and detects the no-end-marker fallback through the
seam's stdin_read wait reason instead of matching its own prompt text.
Tool calls drop from 7180/3560/3566 ms to 355/88/91 ms (spawn+init+echo,
echo, pwd; darwin, production defaults). The loader composition suite now
pins the fast path by pushing idleSilenceMs beyond the send bound, and a
real-PTY case proves PS1 self-healing.
Fixes#2585
