Skip to content

Make context recovery compact the outgoing request without losing tool evidence - #329

Merged
acoliver merged 3 commits into
mainfrom
issue322-effective-compaction-20261004
Oct 4, 2026
Merged

acoliver merged 3 commits into
mainfrom
issue322-effective-compaction-20261004

Conversation

@acoliver

@acoliver acoliver commented Oct 4, 2026

Copy link
Copy Markdown
Contributor

Product change

Repairs #322 at the outgoing request owner. Recovery previously changed prior-turn history, then copied accumulated current-turn tool results back unchanged. It now reclaims previously observed results through their already admitted durable projections while keeping the newest tool round live. Tool call/result pairing, completed semantic summaries, prompt/system content, usage and output accounting remain intact. An unchanged provider context rejection fails without replay.

The integration head 37dd3509 retains both the original fix a1e597cc and landed #324 main f4adda82. Cache output ownership, transcript emission, summary ownership and compaction are composed at their owning callers. No context/guard enlargement, fallback summarizer, compatibility reader, gate exception or assertion weakening is permitted.

Evidence and limits

The original fix has a failing baseline reproduction with request size 63,308 → 63,308 bytes. Its rebuilt release completed a controlled native edit/test workload and restored continuation after the actual wire request shrank 644,994 → 325,843 bytes under the unchanged 900,000 guard. A separate authenticated Sol300000/localoauth coding task passed, but did not reach compaction. Historical private payload equivalence is not claimed.

Exact integrated-source gates and release workloads are running. An independent full review of this new scope and composition is active on an immutable detached copy. Neither result is approval yet. Fresh current-head CI, strict release/source archive and extracted acceptance remain required before merge. Individual original-source gates and verified archive passed; the original full release aggregate was not run end-to-end. Raw failures and evidence remain under evalwork/results/pr-landing-20261003/.

Closes #322 when reviewed, verified and merged.

@acoliver

acoliver commented Oct 4, 2026

Copy link
Copy Markdown
Contributor Author

Fresh exact-head CI at 37dd3509 passed eight checks, but independent review found two blocking original-scope defects. F1: rebuilding timestamped internal message parts can create a byte-length change without changing provider-relevant content, permitting an unchanged context-rejection replay. F2: a forced-summary provider context rejection bypasses recovery even when earlier admitted results remain compactable. Both have independent compiled witnesses; an intermittent full stable test failure is retained rather than dismissed by a passing rerun.

Native Sol is repairing both on the composed #324 source. Original assertions and guards remain unchanged. One findings-only follow-up will verify the fixes and original scope; no extra broad review cycle. This PR will not merge on green CI alone. Disk capacity also blocked local build/extraction attempts; only safely identified idle owned current compiler outputs may be reclaimed after preserving binaries and evidence, and all failed/unrun commands remain recorded.

@acoliver
acoliver merged commit 621a51a into main Oct 4, 2026
9 checks passed
@acoliver

acoliver commented Oct 4, 2026

Copy link
Copy Markdown
Contributor Author

Merged at 621a51a5c70656015d5b74c0d5380d005a88da4a after the single findings-only follow-up resolved F1 and F2 on cab5f89f. Original defect inputs remain unchanged; separate contract probes establish no timestamp-only replay and a genuinely reduced forced-summary sixth request. Both full root suites passed 1,170 unfiltered tests, with helper repeats counted separately. Final release controlled compaction/edit/test/restoration and a fresh authenticated coding task passed without enlarging guards.

Fresh exact-head run 37176337042 passed eight checks. Raw checkout/release logs bind generated merge c0976ac7 to actual prior main f4adda82 and reviewed candidate. Its complete tree exactly equals shipped main. All 16 strict release phases, extracted offline tests/build and GNU-tar reproducibility passed; artifact ID 11294031961 retains its actual generating commit. Current-head local aggregate/extracted remains unrun, not relabelled. Initial review failures and all original refs/evidence remain.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Native profile remediation hits 900000-byte context guard after one ineffective compaction

1 participant