Distillation

Date: Sep 14, 2026

• 🟡 (08:28) User surfaced Session API semantics: supplying an absent Session ID creates the Session; sessions.prompt({ id?, sessionID, prompt, delivery?, resume? }) returns the same admission receipt for exact ID reuse, and omitted or true resume schedules execution after admission; sessions.interrupt(sessionID) clears any coalesced follow-up wake registered with the current coordinator, while an idle or missing Session is a no-op.
• 🟡 (08:28) User surfaced that Session execution routing begins solely from the Session ID through SessionExecution.resume(sessionID) to SessionRunner.run({ sessionID, force? }).
• 🟡 (08:28) User surfaced Session runtime scoping: SessionExecution and read-side SessionStore are process-global; SessionRunner, catalog, model resolver, tool registry, permission state, and filesystem are cached per Location; no layer takes a Session ID; omitted Location.workspaceID means implicit-local placement, while explicit workspace identity is reserved for future placement semantics.
• 🟡 (08:28) User surfaced local runner behavior: one explicit llm.stream(request) is issued per Step; each complete local tool call is durably projected before its structured child execution starts eagerly; every started tool fiber is awaited after provider-stream closure; projected history is reloaded once before continuation.
• 🟡 (08:28) User surfaced Step allowance behavior: promoting any new user input resets the selected agent’s configured Step allowance, while multiple steers promoted at one safe boundary reset it only once.
• 🟡 (08:28) User surfaced that tool settlement events carry the owning assistant message ID because provider-local call IDs can repeat across Steps.
• 🔴 (08:28) User stated that before assembling a provider request, the runner must durably fail every local tool still projected as running from a previous process with the exact error Tool execution interrupted; abandoned side effects are never silently replayed.
• 🟡 (08:28) User surfaced hosted-tool projection semantics: call-side and settlement-side provider metadata are preserved separately so settlement or interruption recovery cannot erase continuation identifiers.
• 🟡 (08:28) User surfaced model-switch replay semantics: provider-native reasoning and provider metadata replay only when the historical assistant model matches the selected continuation model; after switching models, visible reasoning remains ordinary assistant text and provider-native metadata is omitted.
• 🟡 (08:28) User surfaced current Context Epoch follow-ups: add durable post-crash continuation recovery for promoted or provider-dispatched work; consider watcher-backed per-file caching only if measurements show that direct safe-boundary observation is too expensive.
• 🟡 (08:28) User surfaced pre-Step context budgeting: before each Step, the runner estimates the complete model-visible request and compares it against the selected model’s context window minus absolute reserved headroom. The reserve is max(requested/model output allowance, configured compaction.buffer). If the request exceeds that budget and older complete turns are available, the runner compacts before executing the pending Step.
• 🟡 (08:28) User surfaced compaction representation: the full transcript remains durable, but its active model representation is replaced by one hidden checkpoint containing a structured rolling summary plus token-bounded serialized recent context.
• 🔴 (08:28) User stated provider-native assistant, reasoning, and tool messages must never survive across the compaction boundary, preventing signature and encrypted-reasoning failures when the earlier prefix changes.
• 🟡 (08:28) User surfaced repeated-compaction behavior: each later compaction updates the prior structured summary with newly compacted messages; the runner then reloads projected history and executes the original pending Step.
• 🟡 (08:28) User surfaced overflow recovery semantics: if a provider rejects a request for context overflow before durable assistant output or tool execution, the runner attempts exactly one overflow-triggered compaction even when local estimation did not predict pressure; a completed checkpoint rebuilds the same logical Step with one remaining physical attempt.
• 🔴 (08:28) User stated a second context overflow, unavailable compaction, or overflow after durable output becomes an ordinary terminal failure; recovery never loops or replays partial side effects.
• 🟡 (08:28) User surfaced that deterministic pruning of old tool results remains a separate follow-up from overflow-triggered compaction.
• 🟡 (08:28) User surfaced V1 runtime-context parity status: per-Step request assembly for placement, selected model, chronological history, and canonical lowering is complete with no remaining V2 work; configured local/glob and remote-URL durable context sources are missing and need independent sources with explicit precedence, unavailable, and removal semantics; policy-filtered built-in, MCP, plugin, and structured-output tools are partial and still require materializing definitions for the effective agent and request.