Distillation

Date: Sep 15, 2026

• 🔴 [requested-bug-fix] (20:25) User asked to implement the confirmed replay defect in /home/byk/Code/opencode: ordinary prompt execution intent currently exists only in EventV2 local commit hooks, which never replay, so rebuilding projections changes an unpromoted default-resume prompt from wake_pending=true to false; exact retries upgrading an original resume:false prompt also require durable replayable intent.
• 🔴 [enforced-workflow] (20:25) User directed the agent to read current repository files and not use lore/recall.
• 🔴 (20:25) User required normal prompt execution intent to be persisted in durable events/projections and never inferred from an unpromoted input because resume:false must remain admit-only.
• 🔴 (20:25) User required initial PromptAdmitted to carry enough canonical intent to replay wake_pending exactly.
• 🔴 (20:25) User required an exact retry that upgrades an admitted resume:false prompt to execution intent to create a durable, replayable, idempotent event/state transition.
• 🔴 (20:25) User required conflicting retries to never mutate state or emit wake intent.
• 🔴 (20:25) User required admission and wake intent to remain atomic. User advised avoiding a second event if it creates a crash gap; if a distinct durable PromptExecutionRequested event is added for retry upgrades, it must be projected transactionally and duplicate exact upgrades must be idempotent.
• 🔴 (20:25) User proposed requestExecution in PromptAdmitted for initial admission plus a focused event for a later exact-retry upgrade.
• 🔴 (20:25) User required the Prompted projection to continue clearing wake_pending.
• 🔴 (20:25) User stated this is unpublished V2 recovery work, so backward compatibility for discarded intermediate events is unnecessary.
• 🔴 [requested-tests] (20:25) User requested deterministic tests for exactly 4 replay scenarios: 1. initial default-resume admission → lost advisory wake → projection rebuild → wake remains pending and recovery redrives; 2. initial resume:false → rebuild remains not pending; 3. exact retry upgrade → rebuild remains pending and recovery redrives; 4. conflicting retry → remains not pending before and after replay.
• 🔴 [requested-tests] (20:25) User required replay duplicate/idempotence behavior to be verified through real EventV2/projectors.
• 🔴 [enforced-workflow] (20:25) User required focused tests to run with --printConsoleTrace, plus Core and Schema typechecks, a migration check if schema changed, and git diff --check.
• 🔴 (20:25) User required the final report to include the design, files changed, exact commands and counts, and residual risk; the response must never be empty.
• 🔴 [preserve-files] (20:25) User required unrelated AGENTS.md, .lore.md, and perf files to be preserved.
• 🔴 [prohibited-operation] (20:25) User prohibited commits, builds, and service operations.
• 🟡 (20:25) Assistant reported that the required repo-setup skill was unavailable and said it would follow checked-in repository/package instructions and Effect-specific guidance directly.
• 🔴 (20:25) User-supplied Effect guidance required using current Effect v4 / effect-smol source rather than memory or Effect v2/v3 examples; if .opencode/references/effect-smol is missing, clone https://github.com/Effect-TS/effect-smol there, search it for exact APIs/examples/tests/naming patterns, and inspect local repository style first.
• 🔴 (20:25) User-supplied Effect guidance required package tests to run from package directories such as packages/opencode, never from the repository root.
• 🔴 (20:25) User-supplied Effect guidance favored Effect.gen(function* () { ... }), Effect.fn("Name") or Effect.fnUntraced(...), Effect Schema, branded IDs, Schema.TaggedErrorClass, thin HTTP handlers, explicit layer composition, Effect-aware platform abstractions, testEffect(...), it.live(...), scoped fixtures/finalizers, and prohibited introducing any, non-null assertions, unchecked casts, or old Effect APIs merely to satisfy types.
• 🟡 (20:26) Inspection found /home/byk/Code/opencode/packages/core/package.json defines package @opencode-ai/core version 1.18.30; relevant scripts are "test": "bun test --only-failures", "typecheck": "tsgo --noEmit", and "migration": "bun run script/migration.ts".
• 🟡 (20:26) Assistant selected the design of storing initial execution intent in PromptAdmitted and adding a durable, idempotent upgrade event for exact retries; both projectors would own wake_pending, while advisory wakes would remain outside persistence.
• 🟡 (20:26) Inspection of packages/core/src/session/input.ts found SessionInput.admit() currently publishes SessionEvent.PromptAdmitted without requestExecution, then conditionally invokes requestWake(db, input) through an EventV2 local commit hook; existing exact retries with requestExecution=true directly invoke requestWake().
• 🟡 (20:26) Inspection of packages/core/src/session/input.ts found requestWake() updates SessionInputTable.wake_pending=true only for the exact id/session_id while promoted_seq is null; projectAdmitted() currently inserts the admitted row without setting wake_pending, while projectPrompted() sets promoted_seq and clears wake_pending=false.
• 🟡 (20:27) Inspection of packages/schema/src/session-event.ts found Prompted (session.next.prompted) and PromptAdmitted (session.next.prompt.admitted) both use PromptFields; PromptFields currently contains timestamp, sessionID, messageID, prompt, and delivery, but no execution-intent field.
• 🟡 (20:27) Inspection found PromptAdmitted is included in both DurableDefinitions and Definitions in packages/schema/src/session-event.ts.
• 🟡 (20:27) Inspection of packages/core/src/session/projector.ts found the SessionEvent.PromptAdmitted projector passes admittedSeq, id, sessionID, prompt, delivery, and timeCreated to SessionInput.projectAdmitted() but no execution intent; the SessionEvent.Prompted projector delegates to SessionInput.projectPrompted() and then runs the general session projection.
• 🟡 (20:28) Existing tests in packages/core/test/session-prompt.test.ts already include "never arms durable wake for a conflicting execution retry", which admits "Original prompt" with resume:false, retries the same ID with conflicting text and resume:true, expects Session.PromptConflictError, verifies wake_pending:false, and verifies no wake before or after SessionRecovery.recover().
• 🟡 (20:28) Existing tests in packages/core/test/session-prompt.test.ts include concurrent exact retry coverage expecting one admitted row and exactly 1 PromptAdmitted v1 event, plus "reprojects pending inbox input without scheduling execution", which replays a resume:false admission through real EventV2 after deleting projections and expects no wake.