[Bug] Compaction selects its own previous checkpoint as the whole span, so a long session never converges (59/79 attempts rejected, 44.6 min burned)

[gorban]

Summary

Automatic compaction can select the checkpoint written by the previous compaction as its entire shadowed span. Re-summarizing a summary produces a same-size summary, the shrink guard rejects it, and the agent/pre-step listener simply tries again on the next step. One session produced 59 rejected attempts in 44.6 minutes of summarization (48 of them inside a single turn), 5 failed manual /compact runs, and a session that never got below its threshold.

Environment

• dsh 0.1.6-alpha.2, commit ddefc45fbc7f8e46dd73185e68295696d1297887, --profile web
• routed model openrouter/deepseek/deepseek-v4.1-flash
• compaction defaults: thresholdRatio 0.8, retainRatio 0.16, compactionRetries 1, maxOverflowRetries 1
• session session-2496881f-8d36-4ce6-93d9-1954a31930c7 — 1,423 events, 5.28 h, exported log attached

Reproduction

1. Run any session long enough to compact once. The first compaction lands normally (in the log: shadowedRange 9–246, 38,883 tokens).
2. The replacement checkpoint now sits at the head of the surface: [system head, checkpoint, …retained tail].
3. Let compaction trigger again — by agent/pre-step pressure or by overflow recovery. selectCompactableRange always starts at the first non-system surface node, so the only eligible span is the checkpoint itself, and shadowedRange comes back as a one-node span ({start: 709, end: 709}).
4. The summarizer rewrites that checkpoint into another checkpoint of the same size. The guard in summarizeCompaction rejects it, compactIfNeeded throws, the agent/pre-step listener logs step compaction failed: …; continuing the turn, and the next step starts the cycle again.

Current behavior

Every rejected attempt is this guard, from packages/compaction/compaction-basic/src/region.ts:

summary is not smaller than the shadowed content (N estimated framed tokens >= M)

Metric	Value
compaction/start	79
failed compaction/end (with error)	59
succeeded	20
wall time in failed attempts	2,678 s = 44.6 min (median 38.9 s, max 97.8 s)
wall time in successful attempts	887 s
failures where summary tokens equal shadowed tokens	39 of 59
attempts in turn 20 alone	48
user /compact commands	5, all 5 failed

19 of the 20 successful compactions are single-node spans on the previous checkpoint — the last one shadows 2,238 tokens and produces a 2,237-token framed checkpoint, which is exactly the equality the guard then rejects:

summary seq	selected span	span is prior checkpoint	shadowed tokens
708	9–246	no	38,883
873	709–709	yes	2,453
877	874–874	yes	2,442
922	878–878	yes	2,432
…	…	yes	…
1338	1256–1256	yes	2,238

Expected behavior

A compaction candidate that cannot shrink should not be retried in a loop. Concretely:

1. Do not offer a compaction checkpoint (a user/message whose data.source.plugin == "compact") as a summary span, or treat such a span as non-compactable when it is the only candidate.
2. When the shrink guard rejects a summary, record that the selected region is non-shrinking for this surface generation and stop retrying it — do not re-run the same doomed summarization on the next agent/pre-step.
3. Bound the total work: 48 full summarization calls inside one turn is not a bounded retry policy, and each call costs ~38k input tokens on top of the wall time.

Relation to existing reports

• Proposal: suppress repeated pressure summaries after non-converging compaction #5669 ("suppress repeated pressure summaries after non-converging compaction") describes the successful-but-still-above-threshold case. This is a different failure mode: here the compaction throws before commit, so nothing is landed and there is no result to suppress — the loop is driven by the throw.
• Overflow recovery produced no compaction attempt after CONTEXT_WINDOW_EXCEEDED (no compaction/start recorded; session stays dead) #6672 reports overflow recovery with no compaction attempt at all. This session is the mirror image: 59 attempts, all rejected.

Evidence bundle

https://gist.github.com/gorban/4725c87390cb391047a2f14b56b98d28

• EVIDENCE.md — how to recompute every number above from the log
• session-2496881f-sanitized.jsonl — full session log (credential-redacted)
• session-2496881f-checkpoint-chain.tsv — the span table
• session-2496881f-stats.json — counts, wall time, failure messages

Recompute the chain directly:

jq -r 'select(.type=="compaction/summary")
       | "\(.seq)\t\(.data.shadowedRange.start)-\(.data.shadowedRange.end)\t\(.data.shadowedTokenCount)"' \
  session-2496881f-sanitized.jsonl