yhh
2a62e704d3
Session persistence moves from internal/runtime/agent to internal/state/sessionstore: the Session type, its checkpoint and event log, listing sidecars, branches, leases, recovery copies, migration, and the compaction and subagent records saved beside a transcript. The agent keeps the loop and imports the store; history, historycatalog, migration, guardian, instanceid and sessiontool now reach the store instead of the agent, which drops layer-order from 26 to 18. Tests that exercise only store internals moved with it; the rest stay in agent and qualify the moved names. DigestSessionMessages and CompactionStateSchemaV1 are exported for the agent tests that build legacy fixtures. Baseline: carried with -allow-widen. The moved files bring their essay, file-size, complexity and orphan debt unchanged under new paths, and boot.go, acp/service.go, serve/serve.go, agent/task.go and control/controller.go each gain the one import line the split costs.
workgroup assay
Two assays, one question each. Keeping them apart is the point: mixing their turns would put two sampling mechanisms in one denominator.
go run ./tools/workgroupassay # P2-2e-natural, the current one
go run ./tools/workgroupassay -root DIR
P2-2e-natural — is ordinary use worth folding?
The live assay. Its protocol is in protocol.go so changing it is a reviewable
diff.
- Sample: the first 20 eligible authored turns produced after the freeze timestamp by ordinary Studio use, in the order they happened. Not the last 20, not a chosen 20, not 24 because the answer landed near a line.
- Eligible: the turn carries a named
authoredTurn— which only the current schema emits, and which a synthetic continuation never gets — and its trajectory readscomplete, because a prefix cannot answer a question about a whole turn. - In: quiet turns with no tool call, one-call turns, executor-only turns, plan-and-execute turns, and whatever approvals and asks happen on their own. A turn with nothing to fold is evidence about coverage, and dropping it would answer a different question and flatter the feature.
- Out: turns written for this experiment, benchmark and fixture runs, and anything shaped to call more tools than the work needed. Enforced, not merely stated: a workspace under a temp root is where a driven run lives, so its sessions never enter the sample. Before that check the only thing keeping a scripted run out was that it happened before the freeze, which says nothing about the next one.
- No verdict before the sample is whole. Under 20 the tool prints the count and stops.
P2-2d-designed — how good can it be where it fits? (completed)
A written task list of 12 turns, run once against a real model. It is recorded
in protocol.go as a reference condition and is excluded from the natural
verdict. Raw trajectory, task list and result:
~/.reasonix/research/p2-2d-designed/.
It cost ¥2.13 and its most durable result was not a number: on its ninth turn
it broke an invariant ten fixtures had agreed held, because a Partial=true
dispatch is a transport statement and not a promise that a completion frame
follows. See internal/state/workgroup.
The verdict
NO-GO requires both: more than half of groups are a single call, and fewer than 40% of authored turns contain a group of three or more. Anything else is not an automatic go — it hands the decision to a human reading the distributions, the group shapes, and the coverage number below them.
Coverage — how many turns have no assistant-owned call at all — is reported and deliberately not gated. Adding it to the gate now would be redrawing the line after seeing a sample; leaving it out of the report would let a low singleton rate read as broad reach when a fold might touch half the turns.