--- title: Hide background check-ins --- # Hide background check-ins ## What it is When a Claude Code session kicks off a background job — a `Bash` command with `run_in_background: true`, a `--cloud` test or build, a Monitor watch, a scheduled wake-up — and that job finishes, the Claude Code harness wakes the agent back up and it posts a short trailing message. That message is almost always a status/progress report of the background work, not an answer to you: _"Tests green (exit 0). Build still running on the VM."_, _"Typecheck passed — 2 of 3. Lint still running."_ Omniscio already handled most of this: it hides the wake-up _prompt_, it gives the reply its own visual turn so it can't bury your previous answer, and it suppresses the **Needs You** inbox flip (that's the separate [waiting detector](waiting-detector.md)). The one remaining bit of noise was that the trailing status line still _showed_ in the transcript. **Hide background check-ins** collapses it: a confirmed background acknowledgment turn renders as a single, quiet one-line marker — just the label and a timestamp, with no preview of the status text — ``` › 🤖 Background check-in 10:27 PM ``` — that you can click to expand into the full turn. The marker hugs its text (it doesn't stretch across the chat), matching the **Auto-waited** notice right beside it. It's also deliberately label-only: the status reply often carries colored emoji (e.g. ✅) that paint their own bright color regardless of the muted row, so a preview was visually noisy — it's dropped, and you expand to read it. **Nothing is deleted.** Turning on **Show system messages** for a session reveals these turns whole, for auditing. **On by default.** Turn it off at **Settings → Sessions → "Hide background check-ins"**. **One special case — the "nothing to do" wake-up.** When a background event wakes a finished session that genuinely has nothing left to do, it posts a short no-op sign-off ("Nothing to do here…", "Nothing further to do.", "…nothing further needed."). For Claude sessions Omniscio already removes the exact instructed sentence automatically (a separate, always-on tidy). This feature additionally collapses the natural rewordings — and is the **one path that reaches Codex / non-Claude sessions**, which never emit Omniscio's exact wording. **And when that no-op comes from a mechanical resume** — a rate-limit-recovery or lb-rebalance "Please continue" that fired on its own, not something you asked for — it now **collapses by default, no setting needed**: a re-verification like _"Everything was already complete… nothing further needed."_ shouldn't take over your screen or push your real answer out of view. Progress reports ("tests green") on such a resume still honor the opt-in setting. **And the "verify, then sign off" wake-up (2026-08-05).** A very common shape: a background job finishes, the session wakes, runs a quick check or two to be sure, writes a sentence or two about what it found, and _then_ ends with Omniscio's exact "Nothing to do here. Everything was already finished before this background wake-up." Because a short preamble comes before the sign-off, the always-on tidy (which only removes a message that is _nothing but_ the sentence) correctly leaves it visible. This feature now **collapses that whole turn to the one-line marker by default** — one click still shows the check it ran — so a double-check that changed nothing doesn't sit between you and the real answer above it. It stays strictly scoped to a short, plain sign-off: a turn that carries a real answer (a heading, a bullet list, code, or a question) is **never** folded, and the permanent (un-expandable) tidy is untouched — this is only the reversible, one-click collapse. **Reworded no-ops — "nothing left to run" (2026-08-05).** The "nothing to do" family now also recognizes the _run_ phrasing — "nothing further/more/left to **run**" — and it takes priority over a progress phrase in the same sentence. So a wake-up note that says both "the deploy finished" and "there's nothing left to run" is treated as the terminal no-op it is, and the real answer above it stays in view instead of being folded away. (Previously that combination was read as a plain progress report, so the one-line answer in a "Push to Remote" session could end up tucked behind the activity pill while the wake-up note became the visible final message — the reported bug.) **Two more paraphrases — "nothing left to check" and "no wait needed" (2026-09-27).** A session answering Omniscio's own _auto-nudge_ check-in closed with _"…the fix is finished and tagged, and the land is the auto-lander's job … No wait needed, nothing left to check."_ Because the nudge row is dropped from render, that reply stood alone in the transcript as a green bubble with no context — machine traffic you had no reason to read. The turn _was_ ack-eligible (the dropped nudge opens it, so the detector did run); it simply matched no rule, because the agent had paraphrased Omniscio's instructed sentence into its own words. Two additions close that: the _check_ phrasing joins the same `nothing to ` shape ("nothing further/more/left/… to **check**"), and a new **`no-wait-needed`** rule catches the terminal "you can stop waiting" sign-off ("no wait needed" / "no … waiting needed") — a cue the family had never covered, since it is anchored on "nothing", "no further **action**" and "no longer". Both sit in the `nothing-to-do` category, so they **default-fold** (`isAutoResumeNoOpAck`) rather than waiting on the opt-in setting. ## Where to find it ### How to use it - **Turn it off:** Settings → Sessions → **Hide background check-ins** (on by default). The change is live — already-open conversations collapse/reveal immediately, no restart. - **Expand one:** click the marker. The full turn appears and the marker becomes a collapse control; click again to re-collapse. - **Consecutive check-ins batch automatically.** When two or more collapsible check-ins appear in a row (common with parallel background jobs — builds, tests, lints finishing one after another), they collapse into a single grouped marker — _"5 background check-ins"_ — instead of five separate one-liners. Click the batch marker to expand all of them, then click any individual marker to see that turn's full content. Two-level collapse: batch → individual markers → full turns. - **See them all (audit):** the per-session **Show system messages** toggle (session ⋯ menu) reveals every background check-in turn in full, regardless of this setting. ## How it behaves ### What it does NOT do - **It never hides a turn that asks you a question.** Any `?` (or a multiple-choice-widget shape) in the reply means the turn is addressing you — it always renders in full. The same "question always wins" rule the waiting detector uses. - **It never hides real work.** A reply with a heading, a bullet list, a long answer, or a code block is a deliverable, not an acknowledgment — it's never collapsed. (This reuses the same backtested "is this a real message?" heuristic that decides what to promote out of the activity pill.) - **It only acts on genuine self-resumes.** A turn is eligible _only_ when it was opened by a resume the agent did not author — EITHER the harness re-invoking the agent after its own background work (a durable `turnBoundary` marker Omniscio stamps) OR a dropped, system-injected auto-response resume (a rate-limit-recovery / lb-rebalance "Please continue" that carries no stamp — added 2026-06-30) — AND the reply matches an acknowledgment pattern. A normal reply that merely happens to say "tests green (exit 0)" in answer to you is never collapsed — both halves must agree. - **It never collapses a turn that is still running.** A live, in-flight status update stays fully visible; only a settled turn can collapse. - **It always keeps the most recent check-in visible.** Only _older_ check-ins of a run fold; the newest one of each run stays fully expanded. So when a waiting session times out and surfaces to your inbox, opening it always shows the last thing it said — the reason it pinged you — never a content-less marker. This mirrors how the **Auto-waited** notice keeps its newest entry of each run visible. - **It never stacks two collapsed rows for one message.** A background self-resume whose reply _also_ declares a wait (_"…Waiting on the build to finish."_) trips both this feature and the [Auto-waited collapse](waiting-detector.md). With both settings on it would have folded into two stacked markers ("Background check-in", then "Auto-waited") that took two clicks to get past; instead the turn collapses to a **single ⏱ Auto-waited marker** — the wait label wins — so it's one click to the message, which still carries its inline ⏱ badge. - **It never hides a turn that ends the session.** When a session archives itself (its agent decides its work is done) or is auto-archived, that turn carries a _"Self-archived — the agent ended its own session."_ / _"Automatically archived — …"_ note — a terminal event you asked to see. Even when the sign-off reads like a routine "nothing to do" check-in, the turn is treated as a first-class landmark and always renders in full, never folded into a background check-in batch. This is why a self-archive that lands at the end of an auto-continue run stays visible instead of being buried in the collapsed group (the 2026-08-18 report). - **It doesn't delete or alter any data.** The full turn is still in the conversation and the database; this is purely a display collapse, reversible by a click or by "Show system messages". - **It does not change which sessions land in your inbox.** That's the waiting detector's job; this feature only changes how an already-settled turn renders. ## For agents ### How it works (technical) Two halves, AND-gated — either alone is not enough: - **Structural half** — the reply turn must have been opened by a self-resume the agent did not author: EITHER a `metadata.turnBoundary` stamp (the durable mark Omniscio writes when the harness re-invokes the agent after background work / a wake-up) OR a turn split off at a dropped, system-injected auto-response resume (a rate-limit-recovery / lb-rebalance "Please continue", which carries **no** stamp — the builder tracks this as `openedByResume` off `currentTurnOpenedByAutoResume`, added 2026-06-30 so a no-op reply on that stampless path is eligible too). This is what distinguishes a genuine background acknowledgment from a normal reply that happens to use status vocabulary. - **Content half** — a small regex library (`BACKGROUND_ACK_PATTERNS` in `src/shared/background-ack/patterns.ts`, categories _exit-code_ / _check-result_ / _progress-tally_ / _still-running_ / _nothing-to-do_) matched against the turn's reply prose by `detectBackgroundAck` (`src/shared/background-ack/detect.ts`). Before the pattern match, three guards reject non-acks in order: a 500-character length cap, a question veto, and a deliverable veto (heading / list / long prose / code fence). Deliberately distinct vocabulary from the waiting detector — that one matches the _hold_ ("standing by"), this one matches the _completion/progress_ report ("exit 0", "tests green", "0 errors", "N of M", "build still running", "merge/install still running") that precedes it, plus the woken _no-op_ sign-off ("nothing to do here", "nothing further to do", "no further action needed", "no longer needed / already worked around", "nothing left to check", "no wait needed"). The turn builder (`src/shared/real-conversation-turns.ts`) AND-gates the two halves and stamps `turn.isBackgroundAck`. The actual hide is a **live render decision** in the shared `computeTurnFoldState` predicate (`src/renderer/src/lib/turn-fold-state.ts`, read by both `TurnGroup` and the scroll index) — the on/off setting is read fresh at render, never baked into the cached turn — so toggling the setting flips the view instantly. The render gate also requires `nextTurnIsContinuation` (more agent activity follows this turn before your next message), which keeps the newest check-in of each run visible — the same newest-of-run carve-out the auto-wait marker uses. One exception (2026-06-30): a woken **no-op** ack (`turn.isWokenNoOpAck` — the reply is Omniscio's exact instructed "Nothing to do here…" sentinel) folds even when it is the newest of its run AND folds regardless of the opt-in setting, because the exact sentinel is unambiguous and a no-op never re-pings — mirroring how `isCanonicalSentinelWait` default-folds. This is what keeps a rate-limit-recovery no-op from standing between you and the real answer above it (the buried-final-answer fix). A **second** default-fold (2026-07-03, extended 2026-08-12): `turn.isAutoResumeNoOpAck` — a nothing-to-do ack (the exact sentinel OR a natural reword like _"…nothing further needed"_, _"…no further action needed"_, _"…no longer needed / already worked around"_, _"…nothing left to check"_, or _"…no wait needed"_) whose turn was opened by a **mechanical auto-resume** (rate-limit-recovery / lb-rebalance, or a spontaneous background wake) OR a Claude `turnBoundary` self-resume split that follows a real answer — also folds by default and even as the newest of its run, and (per the same scroll-skip rule) the picker skips its ref-less marker so the view rests on the real answer above. It is scoped to the `nothing-to-do` category, so a progress ack on a mechanical resume still honors the opt-in setting. Since 2026-08-15 the backend also *suppresses the redundant re-ping* for such a reworded no-op on an already-waiting session (previously only the exact sentinel was), so a leftover-task wake like _"…no longer needed"_ neither re-chimes nor sits between you and the real answer above it — the live `314c48ac` leak. A **third** default-fold (2026-08-05): `turn.isTrailingWokenNoOpAck` — a background-wake self-resume whose reply _ends with_ Omniscio's exact "Nothing to do here…" sentence after a short verify/preamble (the "woke, ran a quick check, then signed off" shape that the whole-message hard-hide and `isWokenNoOpAck` both miss, since a preamble precedes the sentence). Detected by `detectTrailingWokenNoOpSentinel` (a strict complement of `detectWokenNoOpSentinel` — the no-preamble whole-message case returns false, so the irreversible hard-hide stays whole-message-exact) and AND-gated on `isBackgroundAck`, so the same substantive veto (≤500 chars / no heading / list / code fence / question) has already spared any real deliverable. It folds by default and even as the newest of its run, to the same reversible one-line marker — never the irreversible hide. When a turn is flagged BOTH a background-ack and an auto-wait, that same predicate suppresses the background-ack collapse so the turn folds as one Auto-waited marker (the auto-wait fold wins) — see [waiting-detector.md](waiting-detector.md). Audit mode (Show system messages) makes the builder leave the flag false, so those turns always render whole. **Batch grouping** (display-only, 2026-07-29). When two or more consecutive turns are all `ackCollapsible`, the renderer groups them into a single batch marker ("N background check-ins") instead of N separate one-line markers. The grouping is computed by `computeAckBatchGroups` (`src/renderer/src/features/sessions/ack-batch-groups.ts`) — a pure function that scans the turns array for consecutive runs and returns a lead-count map (first turn of each run → count + last turn) and a follower-to-lead map (every other turn → its lead's key). `RealConversationList` runs this in a `useMemo` pre-pass, owns the expand/collapse state in a `useState>`, and passes batch info as props to `TurnGroup`. When collapsed, follower TurnGroups return `null` and the lead renders a `BackgroundAckBatchMarker`; when expanded, each individual turn renders its own `BackgroundAckMarker`. Two-level collapse: batch → individual markers → full turns. **Archive-note landmark veto** (2026-08-18). A turn carrying a self-archive / auto-archive note ("Self-archived — …" / "Automatically archived — …", recognized by `isArchiveNote` in `src/shared/archive-note.ts` on the turn's `system` rows) is a first-class _terminal landmark_: `computeTurnFoldState` computes `hasArchiveNote` once and refuses every whole-turn fold for it — the background-ack collapse, the woken-no-op hard-hide, and the auto-wait fold — exactly as it already protects a compaction divider. Without it, a self-archive reply — which naturally reads as a "nothing to do" sign-off ("…nothing left to receive or decide, so I'll close this session out.") and rides an auto-continue "Please continue" resume — set both `isBackgroundAck` and `isAutoResumeNoOpAck`, so it default-folded into the "N background check-ins" batch and hid the session's own ending two clicks deep (the reported bug). Detection is content-based and system-rows-only, so a normal agent reply can never trip it, and it reads only existing render-time fields, so no turns-cache version bump is needed. See `background-ack-detector-contract.md` (`archive-note-landmark-veto`). Like the waiting detector, the pattern library is proven against your real history by a **read-only live-DB backtest** (`scripts/background-ack-backtest.mjs`) that must show **zero false positives** in the would-hide sample before any pattern change ships. On the build's first run it flagged 870 would-hide turns across the corpus with 0 false positives; the 2026-06-11 broadening (background-operation jobs like "merge/install running", plus clean-count results like "0 errors") took the corpus figure to 1,557, still with 0 false positives; the 2026-07-03 addition (the `nothing-further-needed` reword family + the mechanical-resume default-fold) matched exactly **1** corpus row — the target rate-limit-recovery reply — with 0 false positives; the 2026-08-12 addition (the `no-further-action-needed` rule + the turnBoundary reword-fold) held 0 false positives across 73 would-fold rows; the 2026-08-15 addition (the `no-longer-needed` leftover-task rule + backend re-ping suppression for reworded no-ops) matched exactly **1** corpus row — the target `314c48ac` leak — with 0 false positives; and the 2026-09-27 addition (`nothing to check` + the `no-wait-needed` rule) matched exactly **1** corpus row — the target auto-nudge leak — with 0 false positives across all 60,269 agent rows scanned, and 12/12 adversarial inline cases clear. ## Related - [waiting-detector.md](waiting-detector.md) — the sibling system that suppresses the **Needs You** inbox flip for the same self-resume turns (the _hold_; this page is the _acknowledgment_). Same design: regex rules + a mandatory live-DB backtest.