Morning digest for unattended runs: what ran, what it cost, what changed #323

Activity

1. currentbits

   added

   enhancementNew feature or requestNew feature or request

   plan:todoPlanboard statusPlanboard status

   plan:doingPlanboard statusPlanboard status

   and removed

   plan:todoPlanboard statusPlanboard status

   on Aug 17, 2026
2. currentbits commented on Aug 17, 2026

   

   currentbits

   on Aug 17, 2026

   OwnerAuthor

   More actions

   Shipping in three parts. Contract landed on main (734b98f): DigestRun/DigestResult in src/shared/ipc.ts + the pure ranking in src/digest.ts (merge-ready / needs-you / discard, risk flags, wasted-spend total) with 9 unit tests.

   • Contract — done: main collects evidence, the renderer ranks it, so the receipt re-ranks without re-walking git.
   • Main process (grok worker) — electron/digest.js: window since last "mark reviewed" (12h default), per-thread cost, diff/commit stats vs the base branch, and check evidence scraped from the run's own tool calls. That last bit stands in for Verification loop: agents must prove their work before a thread goes green #296's verification stage until it lands: a command that actually ran is evidence, a claim in prose is not.
   • UI (grok worker) — DigestView: three ranked groups, cost + change stats + check badge + risk chips per row, and the discard pile's cost in the headline.
3. currentbits

   added

   plan:donePlanboard statusPlanboard status

   and removed

   plan:doingPlanboard statusPlanboard status

   on Aug 17, 2026
4. currentbits commented on Aug 17, 2026

   

   currentbits

   on Aug 17, 2026

   OwnerAuthor

   More actions

   Landed on local main (4b72ecf), not yet pushed.

   Main process — electron/digest.js: window since the last "mark reviewed" (12h default, persisted as store.digestSeenAt); per-thread cost/turns from the session ledger; diff + commit stats vs the base branch via the existing gitTryAsync/parseShortstat helpers (never throws — any git failure is zeros); check evidence scraped from the run's own tool calls. Still-running threads are skipped, except one stalled on a permission prompt — that is precisely the overnight case worth waking up to.

   UI — DigestView: three ranked groups, and per row the cost, change stats, a check badge and risk chips. The discard pile's cost rides in the headline (6 runs · $4.12 · $1.80 wasted) so "tokens burned for nothing" is a number, not an inference.

   Verification: 790 renderer tests, 899 electron tests (the one failure is the pre-existing web-integration reconnect baseline on main), tsc --noEmit clean, plus a manual smoke of the real git-stats path that the unit tests stub out.

   Two ceilings, both marked with ponytail: comments in the code:

   • Cost is the thread's cumulative session cost, not the window's slice — the store keeps no per-thread/per-window ledger, and adding one wasn't worth it for a receipt read at breakfast.
   • Check evidence is "a test/build command actually ran, and its tool call did not error". That is a real signal and a much better one than prose, but it is a stand-in until Verification loop: agents must prove their work before a thread goes green #296 gives threads a genuine verification stage with artifacts. When Verification loop: agents must prove their work before a thread goes green #296 lands, DigestRun.checks is the single field to repoint.
5. currentbits

   added a commit that references this issue on Aug 17, 2026

   Merge origin/main into the observability branch (#280)

   ...

   401ac3e
6. currentbits commented on Aug 18, 2026

   

   currentbits

   on Aug 18, 2026

   OwnerAuthor

   More actions

   Round 29 data point: the state-of-agentic-coding July 2026 report (lbotinelly/state-of-agentic-coding, citing Anthropic's 2026 Agentic Coding Trends report) quantifies the workflow this issue serves: morning-batched Codex cloud tasks hit 85-90% success on scoped maintenance work, developers use AI in ~60% of work but fully delegate only 0-20% of tasks, and ~27% of AI-assisted work is net-new work that would not otherwise have happened. Takeaways: the morning batch is a proven pattern with a known success envelope (scoped maintenance succeeds; ambiguous work does not), so the digest should separate "cleanly done" from "needs a decision" rather than presenting one undifferentiated list, and automations should bias toward the scoped-maintenance class of work overnight.
7. currentbits commented on Aug 22, 2026

   

   currentbits

   on Aug 22, 2026

   OwnerAuthor

   More actions

   Verified shipped on main (audit 2026-08-22): implementation present in the working tree, not just on a worker branch. Closing — the planboard never auto-closed these (#632).
8. currentbits

   closed this as completedon Aug 22, 2026