Third application of the reference mechanism: session turns whose
system prompt + tools are byte-identical to the context predecessor's
materialized snapshot persist { agentId, model, inheritedFrom } instead
of ~70KB per turn (measured: turn 2 of a session is now ~1.1KB total).
Inheritance is decided at createTurn by equality against the
materialized predecessor; the model stays concrete, and on
materialization the inherited record's own agentId/model win — a rule
the new test matrix caught as a real bug (the chain base's model was
overriding a mid-session model switch). Reducer invariants: inherited
snapshots must reference the context predecessor; tool identity arrives
via invocation events.
Test matrix per review: prompt-diff -> full snapshot, tools-diff ->
full snapshot, model-switch -> inherits with concrete model, multi-hop
chains materialize, standalone turns never inherit, unreadable
predecessor falls back to full, cyclic inheritance is corruption,
sessions denormalize the model from inherited snapshots.
The inspector now handles sessions too (auto-detected): overview with
per-turn status/size/input preview, --turns to cascade full turn
inspection. Documented in a new repo-root AGENTS.md (storage layout,
reference model, inspector usage, invariants) with a CLAUDE.md pointer.
Breaking for dev turn files: wipe ~/.rowboat/storage.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Two debuggability follow-ups from reviewing real turn files:
- Request refs are now compact strings — 'context' | 'input' |
'assistant:<index>' | 'toolResult:<toolCallId>' — so a raw JSONL line
reads naturally: {"messages": ["assistant:0", "toolResult:toolu_…"]}.
Same exact-match reducer invariants, via parseRequestRef.
- npm run inspect-turn -- <turnId|path> [callIndex] [--full]: prints,
per model call, the EXACT provider payload (resolved system prompt,
tool list, wire-form messages with user-message context woven in, and
the response/failure) — rebuilt by the same composer the loop sends
through. This is the missing viewer for the derived-not-duplicated
request design: the file stores facts once; the inspector shows what
the model actually received.
Breaking for turn files written since the previous commit (dev only):
wipe ~/.rowboat/storage.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
model_call_requested carried three duplications that dominated turn-file
size (measured on a real 647KB two-step turn): the system prompt (28KB)
and the 38-tool snapshot (40KB) repeated per call despite being byte-
identical to turn_created.agent.resolved, and messages re-inlining the
whole current-turn transcript (230KB re-stating a 207KB tool result).
ModelRequest is now a list of references into the turn's own events —
call 0: [{context}?, {input}]; call N: [{assistant: N-1}, ...that
batch's toolResults in source order]; re-issue after an interruption:
[]. Every referenced byte exists exactly once in the file; the measured
turn drops to ~285KB and each further model call costs ~200 bytes. The
reducer's ordering invariant got stricter: reference lists are matched
exactly against the transcript.
Debuggability is now byte-for-byte at the wire level: ResolvedModel
gains encodeMessages (the structural->wire conversion — user-message
context weaving, attachment rendering, tool-result enveloping, i.e.
convertFromMessages), and composeModelRequest rebuilds the exact
provider payload (resolved system prompt + wrapped tools + materialized
prefix + resolved refs, encoded). The loop transmits exactly the
composer's output, so the durable file plus the composer reproduce what
the model received — pinned by a property test asserting composed ==
sent for every call, and a 2KB size guard on request events. A bridge
test demonstrates the woven wire form (no raw userMessageContext).
Breaking for existing dev turn files (same schemaVersion, pre-release):
wipe ~/.rowboat/storage.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Five bridges adapt existing app code to the runtime seams, each unit-
tested against injected fakes:
- RealAgentResolver: loadAgent + dynamic builders -> immutable
ResolvedAgent. Model precedence override > agent > app default; tool
descriptors from BuiltinTools (zod -> JSON schema) and MCP
attachments (toolId schemes builtin:/mcp:server:tool); ask-human as
the async requiresHuman tool. System prompt composed via
composeSystemInstructions — extracted verbatim from streamAgent so
old and new runtimes share one implementation — driven by the new
opaque RequestedAgent.overrides.composition (session-sticky inputs
preserve provider prefix caching; per-message context stays on
UserMessage.userMessageContext as today).
- RealModelRegistry: models.json -> live AI SDK model; one streamText
step (stopWhen: stepCountIs(1)) normalized into deltas, durable step
events, and the completed assistant message.
- RealToolRegistry: execTool dispatch (builtins + MCP) with per-call
abortRegistry bracketing and tool-output-stream -> durable progress.
- RealPermissionChecker: getToolPermissionMetadata rules (command
allowlist, workspace file boundaries); session grants deferred.
- RealPermissionClassifier: classifyToolPermissions with conversation
context now threaded through the classifier batch.
Interface refinements (doc updated): ToolExecutionContext gains
turnId/toolCallId; classifier takes a batch with turnId + messages;
TurnRuntime dep renamed bus -> lifecycleBus and SessionsImpl bus ->
sessionBus for strict-PROXY DI. Container registers the whole stack
(FS repos rooted at WorkDir/storage); the legacy agentRuntime
registration became lazy to survive the new import-order cycle.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The TurnRuntime per turn-runtime-design.md: constructor-injected,
container-ignorant, no per-turn state — every advanceTurn reconstructs
from the JSONL log, so normal execution and crash recovery share one
path.
- runtime.ts: createTurn/advanceTurn/getTurn; durable-barrier appends
before side effects; sequential sync tools with progress; async tool
exposure + durable suspension; permission pipeline (checker →
optional classifier batch → human/deny fallback, fail-closed on
checker errors); cancellation with synthetic results; recovery that
re-issues interrupted model calls and continues after interrupted
sync tools; model-call-limit exhaustion with machine-readable code.
- fs-repo.ts: date-partitioned append-only JSONL, strict line
validation, path-traversal rejection, per-turn in-process locking.
- stream.ts: hot execution stream — buffers before the consumer
attaches, outcome independent of consumption, close-drops-events.
- context-resolver.ts: chain-walking materialization of context
references with cycle detection.
- Seam interfaces for stage-4 bridges: model/tool registries,
permission checker/classifier, agent resolver, lifecycle bus, clock.
- shared reducer: track classification failures durably so failed
classifications are never re-run.
67 new tests: all nine §26 end-to-end scenarios (mocked deps,
recovery boundaries seeded directly) plus repo/stream/resolver suites.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Durable contracts for the new turn/session runtime, shared by core and
renderer:
- turns.ts: zod schemas for all 16 durable turn events, TurnContext
(previousTurnId ref | inline messages), ModelRequest with contextRef
and current-turn-only messages, ephemeral delta types, reduceTurn
enforcing every spec invariant, and pure derivations
(deriveTurnStatus, turnTranscript, outstanding work).
- sessions.ts: session_created/turn_appended/title_changed schemas,
reduceSession, and the SessionIndexEntry projection.
- vitest wiring for packages/shared (config, build-excluded tests);
92 tests covering happy paths, recovery-shaped histories, and one
named test per corruption invariant.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
