Run a remote Trinity agent task in a sequential, bounded loop — a fixed number of iterations or until a stop signal, with optional response chaining. Fires server-side via run_agent_loop (caller can disconnect), then polls and renders progress. The remote, durable counterpart to Claude Code's local /loop. Add `local` to run the same bounded loop natively in this session instead — inline back-to-back iterations, or a hand-off to the built-in /loop when a cadence is given.
Run a remote Trinity agent task in a sequential, bounded loop — a fixed number of iterations or until a stop signal, with optional response chaining. Fires server-side via run_agent_loop (caller can disconnect), then polls and renders progress. The remote, durable counterpart to Claude Code's local /loop. Add `local` to run the same bounded loop natively in this session instead — inline back-to-back iterations, or a hand-off to the built-in /loop when a cadence is given.
{"version":"1.7","created":"2026-06-09T00:00:00.000Z","author":"Ability.ai","changelog":["1.7: Teach the server-side loop guardrails — on_failure abort|continue + max_consecutive_failures (#1167), max_duration_seconds / max_cost_usd / no_progress_threshold hard stops — in Phase 2 and Guardrails; queued_timeout receipt framing (~25s, not 60s); point one-shot deferred checks at set_reminder and completion-notification at agent.task.* events instead of loops","1.6: Local mode — a `local` token runs the same bounded loop natively in this Claude Code session: inline back-to-back iterations by default, hand-off to the built-in dynamic /loop when a cadence is given; offered as the fallback when Trinity is unreachable and the task doesn't need the remote","1.5.2: Disambiguate 'fire-and-forget' in Phase 5 — here it means disconnecting from a *server-side remote loop* the backend keeps running (safe), NOT spawning a >~10-min job **inside a single headless run** and ending the turn (which reaps every background task/monitor — that work must be decoupled to an OS-level job; see /trinity:onboard → Long-running jobs inside a run)","1.5.1: Reworded the connection prerequisite to the canonical block shared with /trinity:sync and /trinity:onboard — delegates to /trinity:connect (the single connection owner), reconnect via /mcp, never the CLI/curl","1.5: Resolve the remote from the unified `.trinity-remote.yaml` registry — read the `default` remote's `agent` under `remotes:` (with legacy single-remote fallback). Provenance now accurate: the file is genuinely written by both /trinity:onboard and /trinity:sync","1.4: Default local watch after firing — dynamic /loop (ScheduleWakeup) polls get_loop_status, reports changes/stalls/terminal state; skipped on fire-and-forget or when local /loop is unavailable","1.3: Until sentinel must be tied to verifiable evidence (not self-assessment); stall detection in status/observe; timeout_per_run recommended for Until mode; durable artifacts in agent files, not {{previous_response}}","1.2: No @agent now defaults to this agent's remote copy (.trinity-remote.yaml / name match); fire without confirmation when the task is clear","1.1: Surface modifiers in argument-hint and Usage (count, every-cadence, until); document 1h delay ceiling with redirect to schedules","1.0.1: Quote argument-hint — unquoted brackets broke YAML frontmatter, making the skill invisible","1.0: Initial version — modeled on Claude Code's /loop, backed by the run_agent_loop server primitive"]}
/trinity:loop — run a remote agent in a sequential loop (or run it locally with local)
ℹ️ First, set expectations: before anything else, print one short line with this skill's version and its most recent change — the top entry of metadata.changelog above — e.g. loop vX.Y — recent: <summary>. Then proceed.
The remote counterpart to Claude Code's built-in /loop. Where /loop re-invokes your local session on a cadence, /trinity:loop hands one bounded, sequential loop to a remote Trinity agent: you fire it once with run_agent_loop, get a loop_id, and can walk away — the backend runs every iteration in order, optionally chaining each response into the next, and exits on a hard cap or a stop signal.
Use it for iterative refinement, agentic retry, and bounded polling that must outlive your session (and the sync MCP window — a plain chat_with_agent hands back a queued_timeout receipt at ~25s; a server-side loop outlives all of that). When the task lives on this machine and doesn't need to outlive the session, add local to run the same bounded loop natively here — see Local mode.
Usage
/trinity:loop [@agent] <message> start a remote loop (default verb)
/trinity:loop local <message> run the loop natively in this session
/trinity:loop status <loop_id> show per-run progress
/trinity:loop stop <loop_id> request a graceful stop
No @agent means this agent's remote copy — the usual case is looping your own remote counterpart on Trinity.
Modifiers, anywhere in the message:
<N> times | x<N> | max <N> iteration cap (max_runs, 1–100; default 5)
every <N>m | every <N>s pause between iterations (delay_seconds, up to 1h)
until <condition> | stop when … Until mode — exits early on the [[DONE]] sentinel
Examples:
/trinity:loop @researcher draft section {{run}} of the report, 5 times
/trinity:loop refine the report until every TODO marker is resolved — stop when none remain
/trinity:loop @ci-agent run the test suite until it passes, max 10
/trinity:loop @monitor poll the deploy every 2m until it's healthy
/trinity:loop local run the test suite until it passes, max 10
/trinity:loop status loop_a1b2c3
/trinity:loop stop loop_a1b2c3
Trigger
User wants to:
Run a remote agent task N times, in order, optionally feeding each answer into the next
Loop an agent until it signals done / a condition is met
Poll something on the remote on a bounded cadence without holding a connection open
"loop this agent", "run it until it passes", "iterate 5 times on the remote", "keep retrying until done"
Run the same bounded loop locally — "loop this here", "retry locally until it passes", no remote needed → Local mode
Do not use this for a single remote turn (use chat_with_agent) or a parallel batch (use fan_out). This is the sequential primitive.
Prerequisite — Trinity MCP connection
The mcp__trinity__* tools below need a live connection — Local mode needs none of them and skips this section entirely. If they're unavailable (or a call errors with no connection), first check whether the task actually needs the remote; if it doesn't, offer to run it in Local mode instead. Otherwise run /trinity:connect — it authenticates and writes .mcp.json (or refreshes it from your stored profile) — then reconnect with /mcp (full restart only as a fallback) and resume. Never fall back to the Trinity CLI or curl.
Parsing (in priority order)
Parse the input into verb, @agent, and the loop spec. Mirror /loop's parsing discipline — resolve the cheap structural tokens first, then read intent.
Leading verb: if the first token is status or stop, that's the verb and the next token is a loop_id. Jump to that section. Otherwise the verb is start.
local token: a local / locally / here token (or phrasing like "run it locally") routes to Local mode — strip it and keep parsing; steps 3–5 apply unchanged, then jump to the Local mode section instead of the start phases. If an @agent token is also present, remote wins — an agent-targeted loop can't run in this session; say so and drop local.
@agent token: an @name token anywhere selects the target agent. Strip it from the message.
Iteration count: a count phrase (5 times, x5, 10 iterations, max 10) sets max_runs. If none is given, default to a sensible cap (max_runs: 5) and say so.
Mode signal: an until-condition (until it passes, until tests are green, until done, stop when …) → Until mode. Otherwise → Fixed mode.
Cadence: a every <N>m/every <N>s clause sets delay_seconds. Strip it from the message. The API caps it at 3600 — if the user asks for a period over 1 hour, don't silently clamp: tell them a loop isn't the right tool for that cadence and point them to a Trinity schedule (create_agent_schedule with a cron expression) instead.
If the remaining message is empty, show the usage block and stop.
The two modes
/trinity:loop has exactly the two modes /loop has — the loop body is server-side instead of cron/ScheduleWakeup, but the shape is identical.
/loop (local)
/trinity:loop (remote)
Engine
Fixed-interval (/loop 5m) — runs every tick, you stop it
Fixed mode — runs exactly max_runs times, then max_runs_reached
max_runs, optional delay_seconds
Dynamic (/loop check the deploy) — model self-paces, ends on a condition
Until mode — runs until a response contains stop_signal, capped by max_runs
stop_signal (sentinel [[DONE]]) + max_runs safety cap
— (no chaining)
Chaining (either mode) — feed each answer into the next
{{run}}, {{previous_response}}
Fixed mode
The default. Set max_runs: N. The loop runs N times and stops with stop_reason: max_runs_reached. Use for "do this N times" and bounded polling (add delay_seconds).
Until mode
Triggered by an until-condition. Set a stop_signal (use [[DONE]]) and a max_runs safety cap (the loop will not run forever — the cap always wins). Then rewrite the message so the agent emits the sentinel when the condition is met — this is the crucial step. For example:
run the test suite. If every test passes, paste the passing output and end your reply with [[DONE]]. Otherwise, report what failed.
Tie the sentinel to verifiable evidence (test output, an HTTP 200, zero remaining TODOs), not self-assessment — models skew positive grading their own work, so a gameable condition ("until it's good", "until it's tight") will exit on the first pass. If the user's condition is subjective, rewrite it into something checkable or keep the cap low.
The loop exits early with stop_reason: stop_signal_matched on the first iteration whose response contains [[DONE]].
Chaining (orthogonal to mode)
If the task refines or builds on prior output ("refine", "improve each pass", "continue from the last draft"), template the message with the substitution helpers:
{{run}} → the 1-indexed iteration number
{{previous_response}} → the trailing 2000 chars of the previous iteration's response (empty on run 1)
Example: Draft section {{run}} of the report. Stay consistent with what came before: {{previous_response}}
{{previous_response}} is lossy (trailing 2000 chars) — fine as a hint, not as the artifact. When iterations build a real artifact (a draft, a migration, a report), instruct the agent to keep it in a file in its workspace and re-read it each run: Refine the draft in report.md — read it first, improve it, write it back. The remote agent's filesystem persists across runs; the loop context doesn't.
Local mode
A local token runs the same bounded, sequential loop natively in this Claude Code session — no Trinity connection, no run_agent_loop, no loop_id. Use it when the task lives on this machine (local tests, local files), when Trinity is unreachable, or when a quick bounded iteration doesn't justify a remote round-trip. The parse is identical — Fixed vs Until, the cap, an optional cadence — only the engine changes, and there are exactly two:
Inline — no cadence (the default)
Run the iterations back-to-back, right now, in this session. This is a plain sequential loop that you execute yourself:
Announce the plan in one line — local loop — Fixed, 5 runs or local loop — Until (<condition>), cap 8 — then start immediately; the same no-confirmation rule as Phase 3 applies.
For run 1..max_runs: do the task, then print one status line — run N/M — <one-line result>.
Until mode: after each run, check the condition against verifiable evidence (same rule as remote — test output, an HTTP 200, zero remaining TODOs, never self-assessment) and stop early the moment it's met.
Finish with a terminal report: runs completed, stop reason (cap reached / condition met), and the final result.
Chaining is native here — the previous iteration's full output is already in your context, so skip the {{run}}/{{previous_response}} templating entirely. The durable-artifact rule still holds: when iterations build a real artifact, keep it in a file and re-read it each run.
Keep inline loops modest: every iteration is a full task executed in this session, serially. If the task is heavy (a long test suite × 20 runs) or must outlive the session, that's exactly what the remote loop is for — say so and offer it.
Paced — a cadence was given
every <N>m|s means waiting between iterations, and this session shouldn't sit blocked. Hand off to Claude Code's built-in /loop skill (dynamic form — no interval, it self-paces via ScheduleWakeup) with the loop contract baked into the prompt:
<task>. This is a bounded loop: run at most <max_runs> iterations, roughly every <cadence> — keep count across iterations. [Until mode: when <verifiable condition> is met, report the evidence and end the loop.] After the final iteration, report runs completed and stop reason, then end the loop.
Local wakeups have a ~60s floor and a 1h ceiling. A sub-minute cadence (every 30s) can't pace locally — run it inline without pauses, or use the remote loop (delay_seconds handles seconds fine). Over 1h → a schedule, same redirect as the remote parse.
The built-in /loop owns stopping: tell the user "stop the loop" (or Esc) ends it. There is no loop_id — /trinity:loop status|stop don't apply to local loops.
If the built-in /loop skill isn't available in this harness, fall back to inline without pauses and say so.
Choosing local vs remote when the user didn't say: the remote loop survives disconnects and runs on the agent's own filesystem and budget; the local loop is immediate, chains losslessly, and needs no connection. If the loop must keep running after this session ends, it's remote — no exceptions.
Verb: start
PHASE 1 — Resolve the target agent
If an @agent was given, use it.
Otherwise, default to the remote copy of this agent — the primary use case is looping your own remote counterpart. Resolve it:
Read .trinity-remote.yaml at the repo root if present (the shared remote registry written by /trinity:onboard and /trinity:sync) and use the agent of its default remote under remotes:. For a legacy single-remote file (top-level agent:, no remotes:), use that agent directly.
Otherwise call mcp__trinity__list_agents and match this agent's own name (from template.yaml or the working directory name).
Only if no remote counterpart can be found, fall back to mcp__trinity__list_agents: exactly one agent → use it; several → AskUserQuestion with the agents as options (label = name, description = its purpose/status). Prefer a running agent — a loop needs the agent up.
PHASE 2 — Design the loop
From the parse, assemble the run_agent_loop arguments:
agent_name — from Phase 1
message — the task, with {{run}}/{{previous_response}} woven in if chaining; for Until mode, with the explicit "end your reply with [[DONE]] when …" instruction appended
max_runs — the count, or the default cap (1–100)
stop_signal — [[DONE]] for Until mode; omit for Fixed mode
delay_seconds — from any cadence clause (0–3600); omit if none
timeout_per_run — set it for Until-mode loops (a hung iteration silently stalls the whole sequence); size it to the task, e.g. 600 for a test suite. For Fixed mode, set only if the task is long-running; else omit to inherit the agent default (10–7200)
on_failure — abort (default: the first failed iteration ends the loop) or continue (tolerate failed iterations, bounded by max_consecutive_failures, 1–100, default 3). Set continue for agentic-retry loops where a failed run is part of the plan — otherwise the loop dies on the very failure it was built to retry
max_duration_seconds / max_cost_usd — loop-level wall-clock and cumulative-cost hard stops enforced server-side; set them whenever max_runs × model is expensive
no_progress_threshold — server-side doom-loop detection: the backend stops the loop after N consecutive near-identical responses
model — only if the user named one (e.g. claude-opus-4-8); else omit
If a key parameter is genuinely ambiguous (e.g. mode unclear, or no agent could be resolved), ask one focused AskUserQuestion. Don't interrogate — infer sensible defaults and state them.
PHASE 3 — Fire
If the task is clear — agent resolved, mode and cap unambiguous — call mcp__trinity__run_agent_loopimmediately, without asking for confirmation, and show the resolved plan alongside the handle it returns:
Agent: researcher
Mode: Until ([[DONE]]), cap 10
Delay: none
Message: run the test suite. If every test passes, end your reply with [[DONE]] …
Confirm first only when something genuinely warrants it: a key parameter was ambiguous and you had to guess, or the cost is outsized (large max_runs × expensive model — see Guardrails). One question max, then fire.
PHASE 4 — Report the handle
The server starts iterating right away — like /loop, the first run happens now, not on some later tick. Report:
the loop_id (the user needs it to check or stop — always surface it, the way /loop surfaces its job ID)
the mode and cap
how to come back: /trinity:loop status <loop_id> and /trinity:loop stop <loop_id>
that they can disconnect; the loop also appears on the agent's Loops tab in the Trinity web UI
PHASE 5 — Local watch (default)
After reporting the handle, start a lightweight local watch by default — skip it only if the user said fire-and-forget (or equivalent), or the session is headless. ("Fire-and-forget" here is safe: you're disconnecting from a server-side loop the backend keeps running — not the same as spawning a heavy job inside a single headless run and ending the turn, which reaps every background task/monitor spawned in that turn. A >~10-min job must be decoupled to an OS-level cron/systemd/sidecar — see /trinity:onboard → Long-running jobs inside a run.) Invoke Claude Code's built-in dynamic /loop skill (no interval — it self-paces via ScheduleWakeup) with a watch prompt like:
check trinity loop <loop_id> via mcp__trinity__get_loop_status; post one line only when something changed (run N/M · status · cost) or the loop looks stalled (near-identical consecutive responses — suggest /trinity:loop stop <loop_id>); when status is terminal, report stop_reason and the final response, then end the loop
Pace the watch to the loop: hint the expected iteration time (delay_seconds + typical run duration) in the prompt so wakeups aren't wasted. Tell the user the watch is on and that they can stop it anytime (say "stop watching" or Esc) — the remote loop keeps running either way. If the local /loop skill isn't available in this harness, skip the watch and just point to /trinity:loop status <loop_id>.
PHASE 6 — Observe on demand
If the user asks to look right now, poll mcp__trinity__get_loop_status and render the per-run summary as a table — run_number · status · cost · duration · response preview. Re-poll on request rather than busy-looping. Stop polling once status is terminal (completed / stopped / failed / interrupted) and report the stop_reason and final response.
Watch for stalls: if consecutive responses are near-identical (same failure, same output, no state change), the loop is burning budget without progress — flag it and suggest /trinity:loop stop <loop_id>. The cap alone won't catch a stalled loop running under the limit; loops designed in Phase 2 with no_progress_threshold get stopped server-side automatically, so this manual watch is the fallback for loops fired without it.
Verb: status
Call mcp__trinity__get_loop_status with the loop_id and render:
header: status, stop_reason (if terminal), runs_completed / max_runs, total cost
a per-run table: # · status · cost · duration · response preview
the last full response below the table
a stall warning if recent responses are near-identical with no progress — suggest stop rather than letting it run to the cap
If no loop_id was given and none is in this session's context, ask for it — there is no MCP "list loops" call, but loops are visible on the agent's Loops tab and in the execution timeline (tagged with loop_id) in the Trinity web UI.
Verb: stop
Call mcp__trinity__stop_loop with the loop_id. It returns:
stopping — the in-flight iteration finishes, then the loop exits with stop_reason: user_stopped. Report that the current run will complete.
already_done — the loop had already reached a terminal state; report its final status via get_loop_status.
Recipes
Iterative refinement (Fixed + chaining): Improve this draft. Previous version: {{previous_response}} · max_runs: 4. Four progressively-refined passes.
Agentic retry (Until): Attempt the migration. If it succeeds, end with [[DONE]]; otherwise fix the error and we'll retry. · stop_signal: [[DONE]] · max_runs: 8. Stops as soon as it works, bounded at 8 tries.
Bounded polling (Until + delay): Check the deploy health endpoint. If healthy, reply [[DONE]]; else report status. · stop_signal: [[DONE]] · delay_seconds: 120 · max_runs: 30. Polls every 2 min for up to an hour, exits the moment it's healthy.
Guardrails
The cap is the safety net. Always set max_runs even in Until mode — stop_signal is best-effort (the agent has to emit it); the cap is guaranteed. Bounds: max_runs 1–100, delay_seconds 0–3600, timeout_per_run 10–7200. The server-side hard stops (on_failure/max_consecutive_failures, max_duration_seconds, max_cost_usd, no_progress_threshold) are guaranteed too — prefer designing them in over watching.
One loop, one handle. Always echo the loop_id. A loop the user can't find is a loop they can't stop.
Sequential, not parallel. If the user wants the same task across many agents/inputs at once, that's fan_out, not this.
One deferred check ≠ a loop. A single "check back in 2 hours" is set_reminder (one-shot self-trigger, 60s–30d out), not a 1-run loop; and "tell me when the worker finishes" is a subscription to its agent.task.completed event (trinity#1578), not a polling loop.
Cost compounds. N iterations = up to N task executions against the agent's budget. For large max_runs × expensive model, set max_cost_usd and say so in the Phase 3 plan before firing.
