fix(compaction): apply warm summaries anchored to the summarized prefix by code-yeongyu · Pull Request #853 · code-yeongyu/senpi

code-yeongyu · 2026-08-13T07:03:28Z

Problem

The idle warm-up exists so summarization happens while the user is not waiting. Its result was
pinned to expectedRevision, a counter agent-session bumps on every appended message. A
session parked in a cache-warm wait keeps appending — wait notices, monitor state, and finally the
user's own next prompt — so the warm summary was guaranteed stale exactly when the blocking route
needed it.

Field logs from a real session:

04:35:04  speculative_started  gen11        # warm-up begins
04:39:44  threshold_trigger                 # user sends a message
04:39:44  warm_consumed        gen11
04:39:44  speculative_stale    gen11        # warm summary thrown away
04:40:33  compaction committed 361068 -> 21553   # 49s fresh summarization, user waiting

Aggregate over that log: 415 warm-ups started, 36 consumed, 26 stale, 10 applied. Commit
3689633 previously made stale non-terminal (regenerate instead of "Compaction did not apply"),
which made the path correct but left the warm result unusable.

Fix

A summary describes the entries before its cut point, so growth after that cut cannot
invalidate it — only a rewrite of the summarized prefix can.
CompactionPreparation.firstKeptEntryId is that cut.

New warm-anchor.ts predicate: the anchor still exists, the prefix ids are unchanged, and no
newer compaction boundary landed behind it.
applyGeneratedCompaction consults it when the revision moved, instead of rejecting outright.
ApplyCompactionOptions gains expectedFirstKeptEntryId; the core compare-and-apply gate
re-validates the same anchor before mutating the transcript. expectedRevision is untouched for
back-compat.

Verification

RED → GREEN: warm-summary-anchor.test.ts failed expected 2 to be 1 (warm summary
discarded, a second summarization billed), then passed at one summarization.
Safety case included: a warm summary whose prefix was rewritten by a landed compaction boundary is
still rejected and regenerated.
Compaction suite 342/342 (was 341; +2 new cases, 0 deleted). Root npm run check green.
Real-CLI QA (warm-summary-anchor-qa.mjs, RPC + mock provider): 8/8, compaction log records
speculative_applied with no speculative_stale, one summarization, second reply delivered,
real auth untouched.

One pre-existing test asserted the defect

speculative-compaction.test.ts had "skips applyCompaction when message revision changes", whose
fixture bumps the revision without mutating the session at all — provably an append-only case.
That assertion now expresses the intact-prefix contract, and a new case covers the rewritten-prefix
rejection it was standing in for.

Scope limit (measured, not assumed)

A file probe on the real CLI shows a second path that this PR does not change:

IDLE_TRIGGER
CORE_ROUTE_ENTER reason=threshold hasJob=true   # core route wins the race
BLOCKING_ROUTE_ENTER hasJob=false              # warm job already invalidated

When the core route (session_before_compact) wins, it invalidates the warm job and regenerates.
Handing the warm summary to that route needs core admission changes (a first attempt produced a
retry storm), so it is deliberately left to a follow-up rather than bolted on here.

Plan: .omo/plans/warm-summary-anchor.md

Summary by cubic

Applies idle warm-up summaries when the summarized prefix is unchanged, instead of discarding them because the message revision advanced. Previously tied to expectedRevision (stale on any append); now validated by a content anchor and still rejected when the anchor is missing or a compaction boundary rewrote the prefix.

Core: AgentSession.applyCompaction accepts expectedWarmAnchor and revalidates via a shared warm-anchor validator; expectedRevision remains supported.
Extension: speculative.ts builds a WarmAnchorSnapshot from preparation.firstKeptEntryId; uses expectedWarmAnchor when the revision changed and the anchor is valid, otherwise expectedRevision.
Predicate: compare the latest compaction boundary by entry id rather than array position to keep valid warm summaries in already-compacted sessions.
Types: ApplyCompactionOptions adds optional expectedWarmAnchor; backward compatible.
Tests/QA: updated speculative-compaction.test.ts, new warm-summary-anchor.test.ts, and a real CLI scenario warm-summary-anchor-qa.mjs that records speculative_applied and no speculative_stale.

Migration and rollout

No action required for callers using expectedRevision.
Extensions may pass expectedWarmAnchor to reuse warm summaries after idle appends.

^{Written for commit 3265c8e. Summary will update on new commits.}

The idle warm-up moves summarization off the user's critical path, but its result was pinned to `expectedRevision`, a counter that increments on every appended message. A session parked in a cache-warm wait appends wait notices, monitor state, and finally the user's own prompt, so the warm summary was guaranteed stale exactly when the blocking route needed it: field logs show 415 warm-ups started, 36 consumed and only 10 applied, the rest paying a second full summarization for work already done. A summary describes the entries before its cut point, so growth after that cut cannot invalidate it. Validate `preparation.firstKeptEntryId` against the current branch instead, and admit the warm result while the summarized prefix is unchanged; a landed compaction boundary or a missing anchor still rejects. `speculative-compaction.test.ts` asserted the defect directly - its fixture bumps the revision without mutating the session at all - so that case now asserts the intact-prefix contract and a new case covers the rewritten prefix.

Adds an RPC scenario that parks a session over the compaction threshold but under the hard limit, so `before_agent_start` takes the same `threshold_trigger` -> `applyBlockingCompaction` path as the field logs, then asserts the compaction debug log records `speculative_applied` with no `speculative_stale` and that the blocking path bills no second summarization. Requests are classified by the exact compaction instruction text rather than loose word matching, and the sandbox compaction log is captured as evidence before teardown.

code-yeongyu · 2026-08-13T07:19:30Z

CI status: the `Test (coding-agent)` failures are pre-existing on `main`, not from this PR

main is currently red on its own merge commit. Run 31676716029 (merge of #845, no code from this branch) fails all three coding-agent shards on the same files this PR's run reports:

Shard	`main` run 31676716029	this PR
1/3	`claude-sdk-oauth-diagnostic-render`, `regressions/6784-claude-sdk-oauth-default-lane`	same cluster
2/3	`claude-sdk-oauth-observability`, `claude-sdk-oauth-stream`	same cluster
3/3	`claude-sdk-oauth-auth-lane`, `regressions/494-claude-sdk-oauth-installed-sdk-hook-stop`	`claude-sdk-oauth-auth-lane`

Every failure is in the claude-sdk-oauth cluster. No compaction test fails. Nothing in this diff touches that subsystem, and no commit between this branch's base (1b7d72e19) and current main touches claude-sdk-oauth source or tests.

Local A/B (same machine, `CI=1`, `--shard=3/3`)

	files	tests failed
this branch	303	14
pristine `origin/main`	302	14

Identical failure count and identical failing file; the totals differ by exactly this PR's +1 test file / +3 passing tests. The targeted files also pass 14/14 on both branch and main locally, and 56/56 under CI=1 with the new test files co-resident.

What this PR's own gates say

Compaction suite 342/342 (341 before; +2 new cases, 0 deleted)
Root npm run check — green
Static checks, Terminal tools (mac/ubuntu/windows), Inspector handoff (mac/ubuntu/windows), Test (workspaces + scripts), Changelog gate, GitGuardian — all pass
Real-CLI QA warm-summary-anchor-qa.mjs — 8/8, compaction log records speculative_applied with no speculative_stale, one summarization, real auth sha256 unchanged
cubic skipped: monthly quota exhausted (80,038/80,000 lines, resumes 1 Sep) — the only permitted skip reason

The claude-sdk-oauth breakage on main is a separate defect and deserves its own fix rather than being bundled here.

The core admission gate rejected a warm summary whenever a compaction record sat later in the branch than the anchor. Compaction records are appended after the entries they summarize, so a valid next-generation anchor routinely precedes the boundary it updates: every warm summary in an already-compacted session was discarded and re-billed, which is the steady state for the long sessions this feature exists to serve. Compare the latest compaction entry by ID against the snapshot instead. Simply dropping the positional rule would have admitted a summary built on one branch's boundary onto a sibling branch, so the snapshot now carries the anchor, its prefix entry ids, and the boundary id, and both gates run the one validator in core/compaction/warm-anchor.ts rather than two rules that drifted apart. The negative test now also asserts a fresh summary is generated and applied, so it cannot pass by the route giving up; its fixture keeps a message past the new boundary because a compaction record at the branch tip makes prepareCompaction bail.

# Conflicts: # packages/coding-agent/CHANGELOG.md

code-yeongyu added 2 commits August 13, 2026 15:48

code-yeongyu added 2 commits August 13, 2026 16:33

Merge remote-tracking branch 'origin/main' into fix/warm-summary-anchor

3265c8e

# Conflicts: # packages/coding-agent/CHANGELOG.md

code-yeongyu merged commit 36b15c5 into main Aug 13, 2026
15 checks passed

code-yeongyu deleted the fix/warm-summary-anchor branch August 13, 2026 07:59

This was referenced Aug 13, 2026

fix(release): merge concurrent main updates #858

Merged

fix(compaction): let the core route claim the idle warm summary #862

Open

Navigation Menu

fix(compaction): apply warm summaries anchored to the summarized prefix - #853

fix(compaction): apply warm summaries anchored to the summarized prefix#853
code-yeongyu merged 4 commits into
mainfrom
fix/warm-summary-anchor

code-yeongyu commented Aug 13, 2026 •

edited by cubic-dev-ai Bot

Loading

Uh oh!

code-yeongyu commented Aug 13, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

code-yeongyu commented Aug 13, 2026 • edited by cubic-dev-ai Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Problem

Fix

Verification

One pre-existing test asserted the defect

Scope limit (measured, not assumed)

Summary by cubic

Uh oh!

code-yeongyu commented Aug 13, 2026

CI status: the Test (coding-agent) failures are pre-existing on main, not from this PR

Local A/B (same machine, CI=1, --shard=3/3)

What this PR's own gates say

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

code-yeongyu commented Aug 13, 2026 •

edited by cubic-dev-ai Bot

Loading

CI status: the `Test (coding-agent)` failures are pre-existing on `main`, not from this PR

Local A/B (same machine, `CI=1`, `--shard=3/3`)