[Lexus2016]

What this fixes

Two defects in the single-agent run loops, reported against v7.9.0.

1. A turn that strands a background task reported success

An agent that started something with run_in_background and then ended its turn — "I'll wait for it and continue" — reports a clean end_turn, i.e. subtype:'success'. That is the one rung the auto-continue ladder does not cover: the ladder fires only on NON-success. Nothing else resumes a headless claude -p — the process exits and takes the background shell with it — so the chat printed ✅ Done over work that had not happened, and the continuation the agent promised could never arrive.

run-continuation.js adds two pure decisions, and the loops spend at most one extra run collecting the task.

Detection is structural, never textual. A Bash call carrying run_in_background is the same JSON on every install; user-facing prose is written in the UI language (buildSystemPrompt pins that explicitly), so a regex over English phrases like "I'll check back" is dead on a French or Ukrainian one.

The debt is incremental and turn-level. Four simpler shapes were tried first and every one of them shipped green while still being broken — they are listed in the module header and each has a test named after it:

shape	what it did wrong
one-way latch	charged an extra run to every agent that obeyed the new instruction
per-run counters	the rescue run started from zero, so a text-only second walk-away claimed success — the same bug, one run later
raw call counts	two polls of one job cancelled a second, genuinely abandoned launch
batch total	banked a surplus harvest, so reading a leftover tasks/<id>.output paid for a launch that came afterwards

applyBackgroundTool(state, name, input) folds each tool call into {debt, seen}: a harvest decrements only an existing debt, once per shell id, and a surplus is dropped.

The harvest side cannot key on BashOutput. Measured on CLI 2.1.231, a background launch answers Output is being written to: <hash>/<uuid>/tasks/<id>.output and the agent collects by reading that file — BashOutput is never called.

A bounded nudge needs an honest ending. When the rescue run also walks away, the turn says what is still running, as a real --- status line (statusLineKind() needs the fence, or the SPA paints its own Done badge over the warning).

2. error_max_turns far below the requested budget no longer says "raise Max turns"

Measured against CLI 2.1.231, a run capped at N reports num_turns === N + 1, so a genuine exhaustion lands at the cap. A stop at 3 turns against a 50-turn dial is a cap imposed somewhere else — a different CLI version, a settings.json, a hook on the machine the agent runs on — and sending that user to the dial is actively wrong advice. Applied to the first retry notice as well as the exhausted one: hit the 50-turn limit (used 3) contradicts itself.

The SSH auto-continue path also now logs (it logged nothing at all, on either branch) and names the budget and what was spent, which the local CLI loop has always done. Relates to #67.

Delivery channels

--system-prompt is dropped whenever there is a session to resume (claude-cli.js:252), and is absent entirely for a task with no bot. So each runner takes the channel that actually reaches it:

• chat, Telegram — buildSystemPrompt
• Kanban/scheduled (every engine) — the task prompt, next to TASK_VERIFICATION_SUFFIX
• multi-agent workers — agentPrompt (they resume the orchestrator's session)
• bots — botSp on a fresh session, the standing user block once resuming

The harvest rescue stays in runCliSingle/runSshSingle; duplicating the debt accounting into a third and fourth loop is how the local and remote paths drift apart. The unattended runners get the instruction only — which is the half that matters most there, since nobody is reading that chat.

Known limits, both documented in the module header

• A harvest is not paired to a launch: the shell id exists only in the tool result, while onTool sees tool_use blocks. The mirror case (a leftover read before the launch) is closed by the no-banking rule.
• Shell-syntax backgrounding (cmd &, nohup) inside a foreground Bash call is not detected — separating it from a && b and 2>&1 needs a shell parser. The system-prompt rule covers it in words.

Verification

• npm test → 70 files, exit 0. New test/run-continuation.test.js (69 checks); every wrong shape above has a test named after it.
• Six source-text pins on which expression carries BACKGROUND_TASK_INSTRUCTION, plus a negative pin that taskBotSp must not. Three commits in this branch were green no-ops before those pins existed.
• Verified live against real CLI 2.1.231, all three shapes in one script: obedient agent → debt=0, no nudge; walk away then text-only rescue → debt=1, stranded notice instead of "Done"; launch two collect one → debt=1, nudge.
• Reviewed across 8 rounds by three non-Claude models (codex, agy, grok). Four critical and six warning-level defects were found and fixed during review; the residual findings are the two documented limits above.