Distillation

Date: Sep 14, 2026

• 🟡 (23:26) In commit dc661f30cb2055ad6248fdc59b14998d834497be, packages/core/src/session/execution/local.ts:68 delegates local execution through SessionRunner.Service.use((runner) => runner.run({ sessionID, force })).
• 🟡 (23:26) Provider-history lowering code maps completed tool states to ToolResultPart.make; when tool.provider?.executed === true and tool.state.result !== undefined, it reuses tool.state.result, otherwise it calls ToolOutput.toResultValue({ structured: tool.state.structured, content: tool.state.content }).
• 🟡 (23:26) Provider-history lowering has a TODO to materialize remote and managed URIs before lowering because ToolOutput.toResultValue rejects unresolved URIs rather than treating them as media bytes.
• 🟡 (23:26) Error tool states are lowered with resultType: "error"; provider-executed tools can reuse tool.state.result, while other tools produce { error: tool.state.error, content: tool.state.content, structured: tool.state.structured }.
• 🟡 (23:26) Assistant-history lowering considers a prior message to use the same model only when both message.model.providerID matches model.provider and message.model.id matches model.id; provider metadata is reused only for same-model messages without message.error.
• 🟡 (23:26) Same-model reasoning remains a reasoning content part and can retain providerMetadata; cross-model reasoning is converted to a text part when nonempty and omitted when empty.
• 🟡 (23:26) Provider-executed tool history is emitted inline as [call, result] when a result exists, or [call] otherwise; non-provider-executed tool calls stay in assistant content while their completed/error results are emitted afterward as separate Message.tool messages.
• 🟡 (23:26) Assistant-history lowering filters empty text parts and filters empty reasoning unless it has nonempty provider metadata; if no meaningful assistant content remains, it returns only separate tool-result messages.
• 🟡 (23:26) toLLMMessage omits agent-switched and model-switched records; maps user messages to text plus file media and conditionally adds agents metadata; maps synthetic messages to user-role messages; maps system through Message.system; and maps shell to user content formatted as Shell command: ${message.command}\n\n${message.output}.
• 🟡 (23:26) Compaction records are lowered to user-role messages wrapped in <conversation-checkpoint>, containing <summary>${message.summary}</summary> and <recent-context>${message.recent}</recent-context>, with an instruction to treat the contents as historical context rather than new instructions.
• 🟡 (23:26) toLLMMessages(messages, model) translates projected V2 session history into canonical @opencode-ai/llm context using messages.flatMap((message) => toLLMMessage(message, model)).
• 🟡 (23:26) In packages/opencode/src/server/routes/instance/httpapi/handlers/sync.ts, SyncHttpApi.start invokes workspace.startWorkspaceSyncing((yield* InstanceState.context).project.id), ignores its result, and forks it into the current Scope.
• 🟡 (23:26) SyncHttpApi.replay reconstructs EventV2.SerializedEvent[] from the request’s id, aggregateID, seq, type, and copied data; logs the source session, event count, first/last sequence, and directory; obtains ownerID from InstanceState.workspaceID; then calls events.replayAll(payload, { ownerID, strictOwner: true }) and returns { sessionID: source }.
• 🟡 (23:26) SyncHttpApi.steal rejects a missing workspaceID with HttpApiError.BadRequest, otherwise calls session.setWorkspace({ sessionID: ctx.payload.sessionID, workspaceID }), logs "sync session stolen", and returns the session ID.
• 🟡 (23:26) SyncHttpApi.history queries EventTable, excluding each aggregate’s events through its client-known sequence with not(or(...and(eq(EventTable.aggregate_id, id), lte(EventTable.seq, seq)))); it returns full history when no exclusions are supplied, orders by asc(EventTable.seq), executes .all(), and terminates database defects via Effect.orDie.
• 🟡 (23:26) The sync handler group registers four operations in order: start, replay, steal, and history.
• 🟡 (23:26) The experimental sync API uses root /sync; sync.start starts workspace sync loops for workspaces in the current project with active sessions, sync.replay validates and replays a complete sync event history, sync.steal moves a session to the current workspace through the sync event system, and sync.history.list returns events newer than each supplied aggregate sequence while returning full history for aggregates absent from the input.
• 🟡 (23:26) Other event-replay call sites found in commit dc661f30cb2055ad6248fdc59b14998d834497be include packages/opencode/src/control-plane/workspace.ts:351, packages/opencode/src/control-plane/workspace.ts:402 using events.replay(payload.syncEvent, { publish: true, ownerID: space.id }), and packages/opencode/src/server/routes/instance/httpapi/handlers/sync.ts:51 using strict owner validation.