[dmmdea]

What

mr-hook no longer injects the skill hint on background-task notification turns. Human prompts are unchanged, and so is the quota banner.

Why

Claude Code delivers background-task notifications (a finished background command or subagent, a Monitor event) through UserPromptSubmit, the same event as a typed prompt. The hook ranked them as if they were requests. Measured on one long session: the hint fired 25 times on notifications and 7 times on the operator's own prompts. Every hint stays in context and is re-read on every later call.

The shape, measured on real transcripts

• Every notification prompt starts with <task-notification>. That holds for all 1,710 queued ones (queued_command, commandMode: "task-notification") as well as direct user turns. Three tag layouts occur: background command, finished agent (with <note>/<result>), and Monitor event (<event>, with no tool-use id).
• The hook receives queued notifications verbatim: the usage log's prompt_hash equals the sha256 of the transcript's queued prompt, and the lengths match.
• Other machine-originated shapes exist and are not covered here: <cross-session-message …> and <agent-message …>. Both are queued with commandMode: "prompt", there were 10 in total, and they carry actual requests.

Change

• cmd/mr-hook/machineturn.go: isMachineTurn returns true when the prompt starts, after leading whitespace, with <task-notification>.
• cmd/mr-hook/main.go: on a machine turn, retrieval is skipped, so no embedder call is made and the row logs mode:"machine-turn". The index load, the identity check and the quota banner (including the delegate proposal) run exactly as before. This keeps the PR feat: render the quota banner only when it carries news (semantic ChangedAt) #69 semantics: the banner shows on a changed_at within ±15m, a reset within 15m, the first prompt of a session, or the GLM latch.
• The offload nudge is untouched: 0 of 3,274 unique real notifications can trigger it.
• There are no new writes of any kind. The hook still writes only its usage row.
• Consumers of the new mode: mr-outcomes skips rows with no surfaced ids, and a reader of cands gets none, the same as on a gated-empty row.
• Docs: the README hook table plus one key-property bullet, one SECURITY.md sentence, and the mode list in usagelog.Record.

Tests (red first, then green)

Before implementation, with an inert stub, the notification tests got the gstack-qa hint and failed. The human control passed, and the GLM test passed its banner assertion and failed only on the leaked hint.

test	asserts
TestIsMachineTurnShapes	three real layouts, leading whitespace, a human prompt quoting a notification, a truncated tag, empty
TestHintSilentOnTaskNotificationE2E	built binary, against an embedder that scores every prompt 1.0: no output, 0 embedder calls, mode:"machine-turn"
TestHintSilentOnQueuedNotificationE2E	the queued agent-finished layout, with and without leading whitespace
TestHintStillShownOnHumanPromptE2E	control arm, same setup: the hint surfaces and mode:"embed"
TestGLMLatchBannerOnNotificationTurnE2E	established session, so the first-prompt rule cannot be the cause: banner shown, hint absent, orchestrator state dir byte-identical, no Claude-side state created

Mutation proof

The harness requires each mutant to compile (go test -run '^$') before a failing run counts, and restores the file with a hash check after each mutant.

mutant	result
guard never true at the call site	caught
machine branch dead	caught
retrieval still runs after the machine branch	caught
row mislabelled	caught
isMachineTurn always false	caught
no leading-whitespace trim	caught
prefix widened to contains	caught
tag widened to <task-	caught
quota banner suppressed on machine turns	caught (GLM test)
a machine turn writes orchestrator state	caught (GLM test)
non-compiling control	rejected, not counted

Gates

• go test -count=1 ./...: all 53 packages ok.
• go vet ./...: clean.
• gofmt: clean on the LF blobs of the touched Go files, and every staged blob is LF.
• semgrep scan --config auto --error <changed files>: exit 0, 0 findings. The new test's launcher of its own freshly built binary carries a targeted nosemgrep with its reason. The existing E2E launchers in untouched test files still trip the same rule 5 times; they are out of scope here.
• B12 raised consciously: +28 non-test LOC (22 in machineturn.go, 6 in main.go), 30924 → 30952, with the prior note kept as PRIOR.

Deploy (after merge, not done here)

From a clean checkout of merged main: bash scripts/deploy-fleet.sh. It prints built mr-hook @ <rev> for each binary and ends with fleet is uniform at revision <rev>.

Risks

• A human prompt that itself starts with <task-notification> gets no hint. That is unlikely, and the cost is one missing hint.
• If a future Claude Code release changes the notification envelope, the hint quietly comes back on those turns. That fails in the direction of today's behaviour, not towards suppressing human hints.
• Cross-session and agent-message turns still get the hint (see above).

🤖 Generated with Claude Code