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 transport schema requires both `code` and `description`, but the tool
description and both SDK instruction flavors described the call as passing a
program. `description` was reachable only through the parameter schema, so a
model following the prose emitted `{code}` alone and lost the whole written
program to an INVALID_ARGS rejection.
The length and format guidance stays in RUN_CODE_DESCRIPTION_PARAM_DESCRIPTION
alone, so the schema and the prompt cannot drift.
Fixes#2426
The executor collapse landed without telling the model it exists. Every
tool contributes its own guidance section naming its tool, none of them
qualify how that tool is reached, and they all render before the SDK
(orders 100-199 against SDK_SECTION_ORDER 150), so the prompt said "Use
the read tool" eleven times and never said only run_code is callable.
A real session shows the consequence: the model emitted a native call,
read `unknown tool "read"` for a tool the same prompt declares, and
concluded the deployment was inconsistent rather than routing through
run_code.
The registry now contributes `tools:code-only` at order 99 -- ahead of
the guidance band -- stating the rule, registered wherever `tools:sdk`
is and rendering empty outside an effective `code`. `both` renders it
empty because its native calls do execute, which is also why
both-mode-turn no longer shares code-mode-turn's expected prompt. The
denial itself now names the route back, since a bare UNKNOWN_TOOL for a
declared tool is what misled the model.
LegGasai noted that filtering prompt sections by tool:* prefix is a poor
heuristic: it conflates section naming convention with presentation
semantics and would incorrectly drop tool:structured_output. The executor
collapse already enforces the boundary — a model-direct native call is
rejected as UNKNOWN_TOOL regardless of what the prompt says, so filtering
the prompt adds no security and only risks losing useful guidance.
The tool:read/tool:bash/etc sections describe capability usage patterns
that apply to both native and code presentations, and keeping them does
not reopen the native direct-call path because the executor blocks it.
Yichen Jiang