agent-core: compaction summary candidates (head-tail, bridge, state-sections, short, carry-all, validate), behind --compaction-candidates

[jaredLunde]

Six compaction-summary candidates from the summary research, each its own commit and its own --compaction-candidates name. All off by default; with none named, every request, summary and persisted session is byte for byte what main produces.

Name	What it does	Evidence it answers
head-tail	The re-render keeps a long tool result's first 500 and last 1,500 characters, plus up to 30 test-tally or error lines from the elided middle, instead of only its first 2,000.	66% of long test results in the eval had their pass/fail tally only after character 2,000, so the summarizer never saw it.
bridge	The spliced summary message opens with "Checkpoint of earlier work. Messages after it are newer and win where they differ. Work listed as Done is on disk: don't redo or re-verify it unless something newer contradicts it." in place of the "Compacted from N tokens" line. The count moves to CompactionProvenance::tokens_before, which the session store prefers; previous_summary and the export strip the bridge, so it never stacks.	Codex, opencode and the other major harnesses add such a line; nothing told our model how the checkpoint relates to the verbatim messages after it.
state-sections	Adds ## Tests & Errors (latest) and ## Failed approaches / don't retry before Key Decisions, and turns Next Steps into one ## Next Step that must follow from the state ("Verify and finish" allowed). On the shipped template (cached instruction under either wire's framing, and the re-render) via shape_default_format, and on the state template.	Only 55% of eval summaries mention a test tally; 20% of compactions re-fail the same command within 30 calls; summarized agents often fail to stop.
short	A ~700-word target with per-section caps.	Median summary length grew 771 → 2,167 words by round five; the evidence on length disagrees, so it is measured.
carry-all	Every genuine user message rides <user-requests> (each ≤ 8,000 chars, ≤ 32,000 in all; over budget the first plus the most recent), instead of the first and latest three.	Summaries keep ~17% of user constraints against 93–99.7% when the user's words are re-appended verbatim; Codex carries them all.
validate	After the summary is final and before the carry is appended, deterministically appends what it lacks: the latest test tally and command (when the kept tail has none), the user's constraint lines, and a next step.	A cheap, deterministic version of judge-based summary repair (+2.6 to +8.8 points in the literature).

Includes an end-to-end compact test for bridge + validate, plus unit tests per candidate.

Eval results (probe-based replay + continuation) are reported separately; nothing here changes defaults.

🤖 Generated with Claude Code

https://claude.ai/code/session_01JimHGjsfk2Ktm5GxyZJKKk

[jaredLunde]

Closing: no candidate met its ship rule in the probe eval (bridge, short and validate dropped; sections, head-tail and carry-all inconclusive or untested). Not landing dead flags.