Auto-continue

ht autocontinue controls the auto-continue engine — the piece that decides whether to auto-send Continue to an agent on every turn-end notification. The CLI exposes everything the Settings panel exposes plus per-surface controls (manual fire, pause, resume) that the UI doesn’t have.

autocontinue status

ht autocontinue status
# engine          heuristic
# dryRun          false
# cooldownMs      3000
# maxConsecutive  5
# model           anthropic/claude-haiku-4-5-20251001
# apiKeyEnv       ANTHROPIC_API_KEY
# paused          (none)

ht autocontinue status --json

Snapshot the engine config and the current paused-surface list. Cheap; no I/O.

autocontinue audit

ht autocontinue audit
# 14:02:11  fired      surface:1  next plan step: M3
# 14:02:18  skipped    surface:1  cooldown — 1842ms remaining
# 14:02:33  paused     surface:1  manual pause via ht/UI
# 14:02:55  resumed    surface:1  manual resume via ht/UI

ht autocontinue audit --limit 5
ht autocontinue audit --json

Print recent decisions from the in-memory audit ring (cap 50). Default 20 entries; --limit N (1–50) tightens it. Each row is <time> <outcome> <surface> <reason>. (model) suffix marks decisions where the LLM was consulted.

Outcome	Meaning
fired	Engine sent Continue (or the model’s instruction) into the surface.
dry-run	Engine would have fired but dryRun is on — instruction logged, no text sent.
skipped	Heuristic / cooldown / runaway gate said wait. The reason carries which.
paused	Administrative event from pause (CLI or UI).
resumed	Administrative event from resume.

autocontinue set

ht autocontinue set --engine heuristic
ht autocontinue set --engine hybrid --dry-run false --cooldown 5000
ht autocontinue set --max 10 --model claude-sonnet-4-6
ht autocontinue set --api-key-env MY_CLAUDE_KEY

Persist a partial settings update. At least one flag is required:

Flag	Setting	Validation
--engine <X>	engine	off · heuristic · model · hybrid
--dry-run <bool>	dryRun	true / false / 1 / 0 / yes / no
--cooldown <ms>	cooldownMs	clamped to 0–60000
--max <n>	maxConsecutive	clamped to 1–50
--model <name>	modelName	any non-empty string
--api-key-env <var>	modelApiKeyEnv	any non-empty string

Returns a one-line echo of the new state:

ok — engine=hybrid dryRun=false cooldown=5000ms max=5

The Settings UI re-renders next time it’s opened; the engine reads its config fresh on every dispatch, so changes take effect immediately for the next turn-end.

autocontinue fire

ht autocontinue fire surface:1
# fired  next plan step: M3

ht autocontinue fire surface:2
# skipped  no plan published

Force a dispatch on <surface> using the same lookup pipeline used for turn-end notifications (most-recently-updated plan in the owning workspace, last 12 lines of the surface tail). Useful when:

• Testing a heuristic / model decision without waiting for an agent to fire a real notification.
• Driving the engine from a script that knows the agent finished but hasn’t emitted an ht notify.

The output <kind> <reason> matches the audit ring shape so you can read the decision in one line. Note that fire still respects every gate: cooldown, runaway, paused, dry-run, engine=off all apply.

autocontinue pause

ht autocontinue pause surface:1
# ok — paused: surface:1

Stop auto-continue for a specific surface. Subsequent dispatches return paused until you resume (or until the user types into that terminal — notifyHumanInput from a real keystroke does not clear a manual pause). Useful for:

• Pinning an agent that’s about to do something destructive.
• Pausing one surface while leaving the rest of the workspace responsive.

The pause is per-surface; the engine setting (engine: heuristic) stays on for everything else.

autocontinue resume

ht autocontinue resume surface:1
# ok — no surfaces paused

Clear the pause on <surface>. Side-effect: also resets the runaway counter for that surface, so a surface that was paused after hitting maxConsecutive can fire again immediately.

Recipes

Enable in dry-run, then go live

ht autocontinue set --engine heuristic --dry-run true
# … exercise some agent turns, watch `ht autocontinue audit` …
ht autocontinue set --dry-run false

Pause everything during a long-running build

for s in $(ht list-surfaces --json | jq -r '.[].id'); do
  ht autocontinue pause "$s"
done

# … long build …

for s in $(ht list-surfaces --json | jq -r '.[].id'); do
  ht autocontinue resume "$s"
done

Watch the engine react to a forced fire

ht autocontinue set --engine heuristic --dry-run true
ht autocontinue fire surface:1
ht autocontinue audit --limit 1

Read more

• Auto-continue feature overview — engine modes, dry-run, the LLM provider, settings.
• ht plan — publishing the plan that the engine reads.
• Plan panel feature overview — the sidebar widget that surfaces plans + the audit ring.