Replies captured on 2026-09-02 from OpenRouter, models.dev, and DeepSeek
replay through the discovery parser, and OpenRouter's nested
top_provider.max_completion_tokens now feeds the output-token cap.
Model requests hand the configured baseURL to pi-ai unchanged. Only the
discovery listing URL accepts the Anthropic API root with or without a
trailing /v1, because gateway documentation publishes both spellings.
The new keyless sdk-spawn-node scenario drives the platform shell tool
through a command starting with node and requires the machine's own Node
version in the tool result with no PKG_EXECPATH in the child environment,
so a pkg upgrade that re-records the child-process patch cannot silently
restore the hijack. Runs on every target through --scenario all.
@yao-pkg/pkg's SEA bootstrap rewrites child_process commands named node --
including the string after a -c flag, exactly the Bash tool's bash -c form --
to the executable itself and stamps PKG_EXECPATH into every child
environment, so a model-issued 'node --version' silently booted the dsh CLI
instead of the machine's Node. Pin the packager as an exact root
devDependency invoked through pnpm exec and patch out the single
patchChildProcess call from the SEA bootstrap bundle; packaged children now
resolve node through PATH like any other process. The third-party notices
drop the build-time tools section: the packager is now a declared, patched
devDependency, so the manifest and patch tiers disclose it.
Prefer gh v2.99.0's repeatable --attach flag for publishing demonstration
GIFs: it uploads the artifact and rewrites the body's local-path
reference in place, keeping media out of git history without an orphan
assets branch. The assets-branch workflow remains the fallback when the
GIF exceeds 10 MB, gh is older than v2.99.0, or the repository is not on
github.com.
Review follow-ups: the Agent Note triplet moves to implemented/ rewritten as
shipped state (Decision/Consequences/Testing, present tense), cross-linked
both ways with the 2026-07-28 storage recovery proposal whose projcache
reset/destroy path it supersedes (that proposal stays live for authoritative
and whole-medium damage). The fixtures spec header and the note state the
fixture provenance as recorded facts of the released builds instead of
citing local tooling, and the spec JSDoc points at the note's final home.
The proposed Agent Note records the three shipped on-disk generations of
session_projcache, the read-compat and backup-and-skip decisions, the
upgrade matrix, and the rejected alternatives. The package README documents
the upgrade guarantees and requires every future schema or domain-version
change to land with archived fixtures and tests proving its upgrade story.
The storage subsystem page and the generated cordis catalog pick up the new
DomainSpec fields.
The session_projcache domain declares compatibleVersions: [3, 4] and
invalidRecords: 'backup-and-skip'. The two lineage identity fields become
optional — records admitted from older versions predate them, and the single
reader (identityMatches) interprets absence as the unseeded lineage: exact
for unseeded sessions, while a seeded caller fails the match and refolds
cold, so the lineage binding keeps its protection. Upgraded homes therefore
boot and serve their cached listing titles immediately, including homes
whose new tree already holds current-stamped documents without lineage
fields, and a record failing validation anyway is backed up and skipped
instead of refusing the plugin tree.
tests/fixtures/ archives the real on-disk media of every shipped generation
(v3 whole-unit file, v4 and v5 per-record documents, and the lineage-less
current-stamped shape); fixtures.spec.ts proves each recovers through the
real storage stack, rewrites to the current format on the next live write,
and that a hopeless record is salvaged without costing the boot.
A DomainSpec may declare compatibleVersions: older domain versions whose
stored records the current record schemas still accept. The json backend's
per-record reads admit documents stamped with a declared version (writes
always stamp the current one), and the legacy whole-unit bootstrap migrates
only a file whose stored version is in the accepted set — previously it
migrated any version and stamped the records current, turning a discardable
stale cache into invalid-record failures that refused the whole domain at
open and permanently poisoned the new tree on first boot.
A DomainSpec may also declare invalidRecords: 'backup-and-skip' for domains
whose records are disposable derived data: a stored record failing its zod
schema is moved aside through the new optional KvUnit.backupRecord
(<key>.json.bak.<YYYYMMDDHHmm> under the json backend), logged with its
cause, and skipped, instead of rejecting the open. The default stays
fail-loud, and so do backends without backupRecord.
Refusing a response whose tool call never receives an identity needed a new
failure code, a change to the default retryable set, and a `[DONE]` gate that
overrode the finish reason a provider had already sent — turning a safe
`max-tokens` truncation into up to five retries. The lenient wire it guarded
against is hypothetical: no report describes a stream that omits identity
entirely, and the pre-existing test for it is labelled as such.
Only `acceptIdentity` and the widened wire types remain. They close the
reported erasure and cannot reach a worse outcome than the previous
assignment, because the set of inputs that assign only narrows.
`LlmRuntime.adapterStream` normalizes a thrown `LlmError` into the same error
`finish` the loop routes to `agent/request-error`, so throwing would reach
retry too. The Note claimed otherwise. Yielding is chosen because it reports
the attempt's billed usage first and matches the neighbouring `EMPTY_RESPONSE`
refusal.
The rejection comment repeats the corrected durability wording, and the
assembler's delta-only fallback carries a TODO for the empty name it still
invents for adapters that never close a tool-call block.
The `MALFORMED_TOOL_CALL` JSDoc claimed nothing durable is written, but the
loop appends an `assistant/chunk` for every yielded chunk; only the assistant
message and tool result are withheld. The bounded-recovery Note still listed
a five-code transient set, and neither Note linked the other.
A keyless `malformed-tool-call-retry` scenario now records the refusal, the
retry, and the absence of a `tool/call` for the failed attempt. A translator
case pins that an already closable block also withholds its `block-end`.
Adding `MALFORMED_TOOL_CALL` to the default retryable set changes two
expected outputs the session snapshot lane does not own: the shipped Web
composition's inline snapshot, and the canonical packed layout of the
refreshed `empty-response-retry` fixture.
fz