[PierrotAWB]

Issue for this PR

Issues are disabled on this fork; motivation below. Stacked on #25 (phase telemetry); re-target after it merges.

Type of change

• Refactor / code improvement

What does this PR do?

runLoop reloaded and re-hydrated the entire session history from SQLite on every step (filterCompactedEffect inside the while loop) — O(history × steps) I/O per turn, the largest per-step prep cost on long sessions. The runner lock makes a turn the session's single writer, so the loop can reuse what it already loaded.

Now the first iteration loads history as before; each later iteration fetches only rows with id greater than the last-seen id (new MessageV2.after, one indexed query — ids are monotonic per session) and appends. Two correctness guards, both traced against real behaviors:

• A queued follow-up prompt writes its user message mid-turn (createUserMessage runs before the runner parks); today's per-step reload picks it up, and the id-tail query preserves exactly that.
• If the fresh tail contains any compaction part or summary assistant, the retained-window computation (filterCompacted) is no longer append-stable, so the loop recomputes from scratch.

The cache is captured after SessionReminders.apply so reminder transforms persist into it, and it lives only for one runLoop invocation — cross-turn behavior is untouched.

How did you verify your code works?

The existing multi-step, subtask, question, and compaction suites all depend on correct cross-step history threading and are the golden tests here: prompt.test.ts + compaction.test.ts + message-v2.test.ts 148/148, full packages/opencode suite 3129 pass with the only failures being the pre-existing Truncate date bug (fixed in #24, not in this branch's ancestry) and a pty test that passes 9/9 in isolation (load-contention flake). bun run typecheck clean, bun run lint 0 errors. Fleet-wide before/after comes free from #25's history_ms field.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

🤖 Generated with Claude Code

---

Summary by cubic

Loads session history once per turn and appends newly created messages by id on subsequent steps, replacing per-step full reloads. This removes O(history × steps) I/O in long turns while preserving correctness and cross-turn behavior.

• Adds MessageV2.after(sessionID, id) to fetch rows with id > last seen (ids are monotonic per session).
• Caches msgs across steps in the turn runLoop; if fresh rows include any compaction part or an assistant summary, recomputes the retained window from scratch.
• Captures the cache after SessionReminders.apply, so reminder transforms persist; cache lives for a single runLoop.
• Preserves mid-turn queued user messages: the id-tail fetch picks them up.
• No schema or API changes; no migration required. Monitor history_ms to validate the perf win.

^{Written for commit 2137c4c. Summary will update on new commits.}

[github-actions]

Hey! Your PR title perf(opencode): load session history once per turn, append the id-tail per step doesn't follow conventional commit format.

Please update it to start with one of:

• feat: or feat(scope): new feature
• fix: or fix(scope): bug fix
• docs: or docs(scope): documentation changes
• chore: or chore(scope): maintenance tasks
• refactor: or refactor(scope): code refactoring
• test: or test(scope): adding or updating tests

Where scope is the package name (e.g., app, desktop, opencode).

See CONTRIBUTING.md for details.

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

[cubic-dev-ai]

1 issue found across 2 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="packages/opencode/src/session/prompt.ts">

<violation number="1" location="packages/opencode/src/session/prompt.ts:1218">
P2: On multi-step plan/build turns, `cachedMsgs` retains each iteration's in-memory reminder and plugin mutations, so synthetic prompts or transformed parts accumulate across model requests. Keep the cached history immutable, or run per-step transforms on a clone and merge only persisted changes.</violation>
</file>

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[cubic-dev-ai]

P2: On multi-step plan/build turns, cachedMsgs retains each iteration's in-memory reminder and plugin mutations, so synthetic prompts or transformed parts accumulate across model requests. Keep the cached history immutable, or run per-step transforms on a clone and merge only persisted changes.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At packages/opencode/src/session/prompt.ts, line 1218:

<comment>On multi-step plan/build turns, `cachedMsgs` retains each iteration's in-memory reminder and plugin mutations, so synthetic prompts or transformed parts accumulate across model requests. Keep the cached history immutable, or run per-step transforms on a clone and merge only persisted changes.</comment>

<file context>
@@ -1197,6 +1215,7 @@ const layer = Layer.effect(
             Effect.provideService(FSUtil.Service, fsys),
             Effect.provideService(Session.Service, sessions),
           )
+          cachedMsgs = msgs
 
           const msg: SessionV1.Assistant = {
</file context>

[PierrotAWB]

Consolidated into #29 (one commit per concern there; review history for the prep_ms retry semantics is on this PR's threads).