[Phototonic]

Summary

Live auto-capture previously saw tool names and inputs alongside assistant claims, without the tool results needed to qualify those claims. It now includes bounded V1/V2 textual results and invocation status in the existing summary request. The prompt distinguishes plans, attempts, failures, pending/interrupted work and observed outcomes, preserving useful failure/recovery lessons and making missing evidence explicit.

• Sanitize before truncation; retain up to 32 real calls, with 2 KiB per result and a 16 KiB tool-evidence block (smaller under small context budgets). Preserve early context and recent attempts, deduplicate repeated part identities and label omissions.
• Limit generated summaries to 8 KiB UTF-8, types to 64 bytes and tags to eight 64-byte values. Oversized responses use existing fallback/retry; raw update rewrites over 8 KiB add the bounded draft without changing the existing memory.
• Exclude attachments, arbitrary metadata, synthetic/ignored parts and compaction summaries. No new raw-log store, extra verification model, configuration setting or historical migration.

These are partial excerpts, and model summaries can still be wrong. Update rewrites are size-checked before shared-client redaction, so replacement text can expand a rewrite beyond that initial 8 KiB check. Existing external-provider history retains requests and raw responses before capture validation; separate attempts remain until existing cleanup runs (seven-day default expiry). This does not bound raw provider responses or total database size, and deleting rows need not shrink database files.

Test plan

• Exercised the change locally (or N/A for docs/meta-only)
• Noted any follow-up or known limitations

Validation was completed in an isolated Linux cloud checkout at 17c4eb7708d460d1afc302f7968e39a08c69dc29. The exact tested patch was transferred unchanged for publication; tests were not rerun on the publishing desktop.

• Focused capture tests: 31 passed, zero failures; 110 assertions.
• bun test --timeout 30000: 1,162 passed, four platform-specific skips, zero failures; 8,464 assertions across 156 files.
• bun run build, bun run check and patch-apply checks passed.
• Synthetic persistence checks cover failure/recovery evidence, privacy across V2 block boundaries, bounded output, oversized rewrites and repeated-output/provider-history growth and cleanup. Controlled responses test wiring; illustrative summaries do not certify a configured provider.
• Default five-second profile-runtime timeouts reproduced on untouched upstream; unrelated timeout settings are unchanged. No Windows/macOS or live-provider validation is claimed.

Checklist

• Branch is based on main (do not open PRs from commits already on main)
• bun test passes (or N/A)
• bun run typecheck passes (or N/A)
• bun run check passes (format + lint + typecheck; or N/A)
• Docs / README updated when behavior or contribution process changes