Coder/deepseek-harness
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

round 1: project the web transcript from append-origin events

Browse Source 
Replace the surface-ordered fold with a log-ordered human transcript:
append-origin surface events at their own log positions plus one marker per
landed compaction checkpoint. Command folding, the tool-call index, and the
rev-keyed memo carry over unchanged.

Removes foldDegraded, the padding sentinels, baseSeq, and degradedSeqs() --
they existed only to satisfy the core fold's seq === index assertion. That
also closes the pagination hole A1 exposed: a page can carry a checkpoint
whose shadowed range fell outside the window, and nothing resolves
surfaceOp.start anymore.
This commit is contained in:
Hypatia May
2026-07-30 10:18:24 +08:00
parent c249ad8253
commit 8e5c792b50
45 changed files with 1385 additions and 602 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
packages/client/runtime/README.md
View File

@@ -16,9 +16,15 @@ SlotsService gives the renderer separate bare observables for `useSessions` and
`WorkspacesService.connectWorkspace(workspaceId)` resolves the session a New Session flow lands in: it reuses the workspace's existing blank session from the list mirror (`blank && cwd == workspace.path`) or calls `session.create({workspaceId})`, returning the session id for the caller to open. `SessionSummary.blank` mirrors the host's derived empty-log bit and only ever lowers on the client: seeded by `session.list` / the `host/session-added` frame, flipped false by the first ACCEPTED local `prompt()` (on the RPC success response — acceptance proves the user message is in the host log; a rejected first prompt keeps the session blank and reusable) and by any `running: true` status frame, re-aligned by every list re-pull. List surfaces hide blank rows; the store carries every row. `SessionsService.create` accepts an optional caller-preallocated SessionId and throws `SessionCreateError` (carrying `requestedSessionId`) on failure.
## The human transcript
`ConversationSnapshot.nodes` is the human transcript, not the model surface. `TranscriptAdapter` projects the raw window in log order — every append-origin surface event (`isAppendSurfaceEvent`) at its own log position, plus one `CompactionSummaryNode` marker per landed compaction checkpoint — and never consults surface order. A landed compaction therefore keeps the conversation it shadowed on the model side: the marker reports where the model stopped seeing that history instead of erasing it. Model-only replacement copies stay out: a pruned `tool/result` and a regenerated `assistant/message` rewrite one node for the model and mark no boundary. A checkpoint is a `user/message` carrying the compaction seam's plugin source that **replaced** a surface range; an appending plugin-sourced `user/message` is injected context, not a compaction. That source literal is restated locally because `dsh-compact` is unreachable from this program in both directions (the client purity gate rejects a value import; a type-only import collides the host `Context.sessions` merge) — `tests/compact-checkpoint-pin.spec.ts` is the drift trap.
Because the projection is log-ordered, the node array is seq-monotonic by construction: log-only `command/run` / `command/done` nodes splice in by seq, `Session` merges interrupted frozen nodes by their fractional seqs, and a window whose checkpoint cites a shadowed range outside it renders the marker with nothing logged. The marker's summary text comes from the checkpoint's `compact/summary` provenance; a window cut that left the provenance outside makes the row non-expandable rather than empty, and a later page that supplies it resolves the text. Performance contract: one append materializes one node, an event that changes no node keeps the previous array reference (a chunk storm costs nothing), and unchanged nodes keep their object identity.
## Code Mode sub-dispatch index
`ConversationSnapshot.codeDispatches` groups a `run_code` call's sub-dispatches under their parent callId, in start order, using the native call-block shapes: a `tool/code-dispatch-start` event lands the `RunningToolCall` form (rows derive the running ring from the shape) and its `tool/code-dispatch` settlement replaces it in place with the `ToolResultNode` form, `callTime` carrying the paired start's time. A settle whose start fell outside the replay window appends directly with `callTime: null` (duration unknown — never a fabricated zero). Live mux frames and history replay build the identical index; sub-calls never join the surface `nodes` flow; per-parent array and map references are memo-stable across unrelated snapshot swaps.
`ConversationSnapshot.codeDispatches` groups a `run_code` call's sub-dispatches under their parent callId, in start order, using the native call-block shapes: a `tool/code-dispatch-start` event lands the `RunningToolCall` form (rows derive the running ring from the shape) and its `tool/code-dispatch` settlement replaces it in place with the `ToolResultNode` form, `callTime` carrying the paired start's time. A settle whose start fell outside the replay window appends directly with `callTime: null` (duration unknown — never a fabricated zero). Live mux frames and history replay build the identical index; sub-calls never join the transcript `nodes` flow; per-parent array and map references are memo-stable across unrelated snapshot swaps.
## Session title projection