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.
The shipped bundles keep ctx.sessionQuery mounted but set the new
session-query-sqlite `openAt: never` phase: searchSessions/searchEvents
fail with the typed SESSION_QUERY_SEARCH_DISABLED code before any request
normalization, node:sqlite is never imported or opened, and no source
observation or reconciliation runs. Every inherited exact read, filter,
and trace — session export descendants, subagent-fork Workspace
inheritance, title reads — keeps working, and the Web sidebar search
degrades to its designed local title/workspace matching. Enabling content
search is a one-line openAt override in a later patch layer; the web e2e
scaffold keeps it enabled as the assembled opt-in coverage.
Card registration order is stable only for cards one package installs
together: apply order between packages is unconstrained, so several
external cards can still reorder between boots. The note and README said
otherwise. Recorded alongside it: the redactor returns a secret reachable
only through a union, intersection, or transform verbatim, and serving
every registered namespace widens that gap to third-party schemas; and
the headline capability still lacks an assembled-composition test.
publish() now keeps its snapshot reference when neither the loaded flag
nor the dispatched namespaces moved, so an unrelated settings commit no
longer re-renders the card list.
The Agent Note claimed the plugin inventory page already exposed every
mounted plugin's effective configuration to the same browser, so the
removed allowlist blocked nothing a caller could not already read. That
is false: PluginInventoryEntry carries entryId, moduleName, enabled, and
fiberPhase, and the page's configuration row renders an enabled tag, not
a stored value. The allowlist did keep an unlisted namespace's resolved,
base, and user values off the wire; the note now says so and rests the
decision on the loopback pin, secret redaction, the user's own document,
and the fact that permission and agent-presets were already served.
The post-await disposal check reads through an opaque method, mirroring
the settings seam's isStopped(): control flow narrowed the field to false
across the await, so the lint gate saw the guard as dead.
A plugin that registered a settings namespace could not reach the browser
configuration page: the api-proxy filtered every read and gated every
write through two hardcoded namespace lists, and the plugin configuration
section rendered an unordered list of cards carrying an opaque id rather
than the namespace they edit. Both gates lived in this repository, so a
user-authored plugin was configurable only by hand-editing settings.yaml.
The proxy now serves whatever ctx.settings.describe() returns and adds no
boundary of its own; a name no registration answers folds into the seam's
own settings-rejected, and the settings-not-exposed code retires. The
settings seam is untouched: which client may read a namespace, and which
page renders it, are facts about consumers.
settings.plugin.item becomes a keyed slot whose key is the namespace a
card edits, following tool.call.toolview. The section reads describe once
and dispatches the intersection of the slot ledger and the served set, so
a namespace another surface owns renders nothing without declaring
anything, and a card for an uncomposed plugin is never dispatched.
The declared-provider-catalog note still recorded modalities as closed
because "nothing reads them: context.ts keeps only text blocks" — the
same reasoning this change disproves. Rewrite that fact in place, keep
pricing closed for its own still-current reason, and cross-link both
directions.
The new resolver cases never crossed the config schema, so neither
acceptance path this change adds had an executable gate: config.spec.ts
now refuses an unknown modality at both levels, pins the empty route
list to the namespace validator the settings seam actually runs, and
pins the `[]` materialization the inheritance rule depends on. A
catalog.spec.ts case re-asserts every rung through a written settings
section and `ctx.llm`, which is what the note's Testing section claimed.
Also export PiAiModality from the package entry — it types two public
config fields and consumers could not name it — record the unverified
declaration as a Known Limitation, and stop sharing one fallback array
across a route's models.
A model the installed pi-ai catalog does not describe was reported as
text-only with no way to say otherwise, so a vision model added through
the custom-provider form was refused at every image admission point.
The justification in the source described the DeepSeek chat-completions
serializer, which does reject image blocks; the pi-ai request converter
and every wire protocol it speaks carry images.
Modalities now resolve entry `input` -> installed catalog entry -> route
`defaultInput`, the chain the two capacity fallbacks already use, so the
route value is a fallback and never narrows a catalog model. Its default
stays `[text]`: nothing can interrogate a gateway for its modalities, and
over-claiming admits an image the provider rejects mid-turn, after prompt
admission has already committed the message.
parseCmdline(ctx, program): void only adapts commander control flow to
the launcher: it parses the immutable cmdlineArgs snapshot and turns
help, version, parse errors, and action rejections into a ctx.appExit
request. App validation and the ctx.provide of the app-owned service
live in the program's own synchronous .action(), which commander runs
inside parse — program.error(...) there shares the exit path with a
grammar rejection. Deletes the CmdlinePlan export, its unread ctx
parameter, the type-unsound (() => ({}) as T) default, and the
T | undefined return with its per-caller publish guard.
