Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@packages/opencode/src/session/compaction.ts`:
- Around line 369-375: The guard currently prevents clearing a persisted tail
because it requires selected.tail_start_id to be truthy; remove that truthiness
check so we compare and persist changes even when selected.tail_start_id is
undefined. Update the condition around compactionPart and the call to
session.updatePart (referencing compactionPart, selected.tail_start_id, and
session.updatePart) to simply check compactionPart &&
compactionPart.tail_start_id !== selected.tail_start_id and then call
session.updatePart with tail_start_id: selected.tail_start_id (allowing
undefined to clear the stored boundary).

In `@packages/opencode/src/session/message-v2.ts`:
- Around line 934-951: The loop currently only assigns tailStartID from
compaction parts on messages that are marked completed, which lets an older
completed compaction override a newer pending one; change the logic in the
newest→oldest scan (the block that inspects compactionPart and completed) so
that the first compactionPart you encounter with a defined tail_start_id
immediately sets tailStartID (regardless of msg.info.complete status) and stops
the scan (break/continue as appropriate), so the newest seen tail_start_id wins;
keep the existing fallback that scans result.flatMap(...).find(...) but ensure
it only runs if tailStartID is still unset.

In `@packages/opencode/test/session/compaction.test.ts`:
- Around line 695-702: The new tests create ManagedRuntime instances via
runtime("continue", Plugin.defaultLayer, ProviderTest.fake(), cfg({ prune: true
})) but never call .dispose(), leaking scoped resources; wrap each runtime usage
in a try/finally and call rt.dispose() in the finally (or use a helper that
auto-disposes) for the occurrences around the SessionCompaction.Service.use
calls (also fix the other two spans noted near the 825–840 and 882–897 regions)
to ensure each runtime is disposed after the test completes.

In `@packages/opencode/test/session/messages-pagination.test.ts`:
- Around line 714-742: Replace the raw Jest test declaration and manual Effect
runtime plumbing with the Effect test harness: change the top-level test("keeps
retained tail messages after a compaction part with tail_start_id", ...) to use
testEffect(...) (or it.live(...) as used in this file) from test/lib/effect.ts,
and lift the async body into the Effect test body so you no longer call
Instance.provide manually; keep the same interactions with svc.create,
svc.updatePart, MessageV2.stream/filterCompacted and the same assertions
(expect(ids).toContain(retained) / not.toContain(old)), and remove the explicit
Instance.provide/Instance cleanup plumbing so the harness manages the Effect
runtime and lifecycle.
- Around line 714-742: The test currently only asserts the retained user message
ID is preserved, but must also assert that its associated assistant/tool reply
is retained to ensure full-turn preservation; change the addAssistant call that
follows the retained user to capture its returned ID (e.g., const
retainedAssistant = await addAssistant(session.id, retained)) and add an
assertion expect(ids).toContain(retainedAssistant). Apply the same change for
the other similar test block (the one also noted at 744-780) so both user and
their assistant/tool message IDs are asserted present when using
MessageV2.filterCompacted(MessageV2.stream(session.id)).

In `@packages/sdk/openapi.json`:
- Around line 8918-8921: Add a clear description for the public OpenAPI schema
field "tail_start_id" so generated SDK docs are self-explanatory; update the
schema for "tail_start_id" (the property with "type": "string" and "pattern":
"^msg.*") to include a concise description like "Optional message ID (must start
with 'msg') to begin tailing events from the specified message onward" and
mention its behavior (e.g., inclusive/exclusive) if applicable.
- Around line 11545-11546: Update the OpenAPI description for the
compaction.prune property to explicitly state the behavior when set to false:
clarify that true enables pruning of old tool outputs, false disables pruning
(no automatic pruning), and omission defaults to no automatic pruning; modify
the description string for the "compaction.prune" schema entry accordingly so
generated docs and UIs unambiguously show true = prune, false = do not prune,
omitted = no automatic pruning.

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@packages/opencode/src/session/overflow.ts`:
- Around line 18-19: The current overflow check computes usable =
(input.model.limit.input ?? context) - reserved which can be negative when
input.model.limit.input is 0, causing count >= usable to always be true; update
the calculation in overflow.ts to clamp usable to a minimum of 0 (e.g., use
Math.max(0, ...)) so usable = Math.max(0, (input.model.limit.input ?? context) -
reserved) and then perform the existing comparison with count, referencing the
variables usable, input.model.limit.input, context, reserved and the return that
checks count >= usable.

In `@packages/opencode/src/v2/session-event.ts`:
- Around line 399-406: Change Retried.create to accept an optional attempt so
legacy callers can omit it: update the input type from BaseInput & { attempt:
number; error: RetryError | string } to BaseInput & { attempt?: number; error:
RetryError | string }, and in the factory set attempt to a safe default (e.g.
attempt: input.attempt ?? 0) when constructing the new Retried instance; update
references in the Retried.create implementation accordingly.

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@packages/opencode/test/session/compaction.test.ts`:
- Around line 685-727: Replace the ad-hoc test harness in the "compacts old
completed tool output when prune is explicitly enabled" test (which uses
Instance.provide, runtime(...), rt.runPromise and manual disposal) with the
existing testEffect/it.live harness: convert the test(...) into it.live(...)
from the file's testEffect(env) setup, use provideTmpdirInstance(...) instead of
Instance.provide(...) to supply the tmpdir and Instance, and run
SessionCompaction.Service.use((compact) => compact.prune({ sessionID: session.id
})) within the effect runtime provided by it.live (removing manual runtime,
cfg/prune wiring and explicit rt.dispose). Keep assertions the same and ensure
you still create the Part via svc.updatePart and then fetch svc.messages to
assert updated.state.time.compacted is set.

In `@packages/opencode/test/session/messages-pagination.test.ts`:
- Around line 718-842: The three new it.live tests use Effect.promise(async ()
=> ...) and call svc directly, which spawns separate runtimes; change each to
use Effect.gen(function* () { ... }) and yield the shared service from
SessionNs.Service (e.g., const svc = yield* SessionNs.Service) so the test
harness manages lifecycle; replace the async svc calls with yielded usage inside
the generator and remove the Promise wrapper, keeping the same test logic and
references to svc, MessageV2.filterCompacted,
addUser/addAssistant/addAssistant(summary), svc.updatePart and svc.remove.

---

Outside diff comments:
In `@packages/opencode/src/session/compaction.ts`:
- Around line 269-380: Move the logic that persists the tail_start_id boundary
to occur before processor.process() is called instead of after. Specifically,
before invoking processor.process(), check if compactionPart exists and if
compactionPart.tail_start_id differs from selected.tail_start_id, then call
session.updatePart() to persist the boundary. After this update, add revert
logic in the error handling section where result === "compact" to undo this
change by either clearing or reverting the tail_start_id on the compactionPart,
ensuring the storage boundary stays in sync with actual compaction progress.