R7 When and how long the agent works

How long the research agent works alone, when the next session starts, and why a run ends. Describes the research 1.25.2 and the Strom app 3.21.2. Changes from a later version carry its number.

R7.1 Working alone runs in sessions

When working alone, the agent takes tasks from the queue (R2.2) without a person. One session is one task. A run starts one or more sessions in a row; between them it may ask the condition (R7.4).

• Limits:
  • without an option 1 session; --max <n> at most n; --loop session after session while there is work;
  • --task T…: the chosen tasks in the given order, each one session; a task done or taken meanwhile is skipped;
  • --until HH:MM: after it no session starts, one already running finishes. A time that has already passed today means tomorrow;
  • every run ends when no open task is left in the queue.
• Controls: terminal: the menu Let the agent work on its own, the choices Work through the queue in this order – as many as chosen (1–100, like --max), Only the tasks picked from the list (like --task) and Let it work on while there is work and the condition “…” allows it (like --loop, only with a condition set); the command strom run. The Strom app 3.21.2 does not set up working alone; it shows the state of the run (R7.4), and a session can be finished from it (R7.3).
• Since: research 1.0.0 (--max, --until, --task), --loop from 1.2.0

R7.1.1 Why a run ends

• At the end the run prints N sessions · stopped: …: done, no open tasks, time limit (--until), by hand (R7.3), subscription limit (R7.6), the agent is not logged in (R7.7), the condition's answer (R7.4), or an error (credentials, the agent failed, the model).
• When the remaining tasks need an archive that does not answer, a run with --loop and --until waits (⏸ waiting until …); any other run ends.
• Since: research 1.0.0; waiting for an archive that does not answer from 1.22.0

R7.2 Each session has a time limit

One session lasts at most the set number of minutes, 60 by default. The agent knows the end time in advance, and strom reminds it. At the limit strom stops it and returns the unfinished task to the queue.

• Limits:
  • in the last 10 minutes (a short session: the last quarter) every strom command says how much time is left; in the last 3 minutes (a short session: the last eighth) it asks the agent to write what it found and close the session;
  • at the limit, an agent that can (Claude Code, Codex, OpenCode, Grok) gets 5 more minutes to write and close;
  • a task on which 2 sessions in a row wrote nothing is parked (… parked); sessions ended by a person do not count;
  • the order: --minutes > the variable STROM_RUN_MINUTES > the family tree's setting > the computer's setting > 60.
• Controls: terminal: Settings: the agent, working alone, messages, archives, maintenance → The agent working alone → Time for one task: N min (5–600; 0 goes back without a change); strom config set run.minutes <n> (--for-tree: only this family tree). App: nothing.
• Since: research 1.0.0 (60 minutes, parking a task); reminders and 5 minutes to write from 1.2.0

R7.3 A session can be asked to finish

A person can ask a running session to end. The agent starts nothing new, writes what it found, and closes the session. Then the run ends and no further session starts. Nothing is stopped by force.

• Limits:
  • the agent sees the request at its next strom command; a new reader does not start, and strom does not stop a subagent's work halfway;
  • Ctrl-C in the run's terminal: the first time the same request, the second time an immediate stop (the task returns to the queue). On Windows and with --interactive, the first Ctrl-C stops at once.
• Controls: terminal: strom session finish [N…] (when more sessions run, name which), Ctrl-C. App: Research overview, at the working agent Finish and stop; it is confirmed in the research's terminal. When the run waits for the condition, the button is not there.
• Since: research 1.8.0, app 3.5.0

R7.4 A condition decides whether the next session starts

A condition is a program strom asks before a session: go on, wait, or stop. Only a person sets it, for the whole computer. Without a condition the run goes on while there is work.

• Limits:
  • go on: the session starts; wait: the run waits and asks again; stop: the run ends. Anything else (the program crashes, no answer within 120 s, claude-usage 660 s) is an error and the run ends;
  • when to ask again is said by the condition; otherwise in 15 minutes, never sooner than in 1 minute. Ctrl-C ends the waiting; with --until the given time ends it (time limit);
  • --loop and --until: a question before each session; tasks started by hand: one question (R7.4.1);
  • strom ships the condition claude-usage (R7.5); an own condition is a folder plugins/gates/<name>/;
  • a run without the condition set (--no-gate) or with another one (--gate) needs a person's consent.
• Controls: terminal: Settings → The agent working alone → Working on and on under a condition: … (a condition, or none); strom config set run.gate "<name> [what to give it]", strom config unset run.gate; a person confirms both. App: Research overview shows Paused and resumes at … at a waiting run.
• Since: research 1.2.0, app 3.5.0 (the state Paused)

R7.4.1 Tasks started by hand: “Start anyway?”

• One session, --max, --task and the menu are a person's choice. The condition is asked once, before the first session. When it would not start (wait, stop, error), the question The condition “…” would not start now — …. Start anyway? comes up. Enter means no: the run ends (not started — the condition “…”: …). Yes: the sessions start and the condition is not asked again, except for the cap (R7.4.2).
• When the agent starts the run, a person answers in a system window. A script that no one can be asked keeps to the condition.
• Since: research 1.2.0

R7.4.2 The cap: a hard limit

• A condition can mark its answer as a person's hard limit (hard). The run then ends (the cap of the condition “…”), and no “anyway” goes past it.
• A condition can watch the agent's plan window (watch). It is then asked before each session of every run, and a session at work gets a request to finish (R7.3) when the agent reports the window's use at the given limit.
• Since: research 1.14.0

R7.4.3 The record of the condition's answers

• Each answer has a line in the run's output: · condition …: … (exit …; .strom/gate.log): the condition's name, its answer with the reason and the code.
• At the same time a JSON line in the family tree's .strom/gate.log: when, the condition and what was given to it, the exit code, the verdict, the reason, the waiting, the plan's limits and what the run did (go, wait, stop, anyway). It is kept for 90 days, at most 2 MB.
• Controls: terminal: The agent working alone → What would the condition say now?, strom gate test.
• Since: research 1.14.0 (the record); the question from the menu from 1.3.0

R7.5 The claude-usage condition keeps the week's pace

claude-usage keeps working alone within the weekly limit of the Claude subscription. It compares how many percent of the week have passed since it reset (a day ≈ 14.3 %) with the week's usage. A session starts only when the lead (passed − used) is at least the given number n.

• Limits: the steps go in this order; the first that applies decides:
  1. n is not a whole number 0–99, or the cap is not a whole number 1–100: stop (every run);
  2. the agent is not Claude Code, or Claude Code runs on an API key: go on;
  3. the usage from the agent's report in this run (at most 15 minutes old), otherwise from claude -p /usage. Cannot be read: 3 attempts 30 s apart, then wait 15 minutes. Claude Code is not logged in: stop (R7.7);
  4. a week without a reset time: used 0 % → go on, otherwise wait 15 minutes;
  5. used ≥ the cap: stop, hard (R7.5.4);
  6. the five-hour window used up (100 %): wait until it resets, plus 1 minute;
  7. the lead ≥ what is needed (R7.5.1), after waiting with the extra (R7.5.2): go on;
  8. the last 2 hours before the week resets, and at least 10 points left to 100 % (or to the cap): go on;
  9. otherwise wait (R7.5.3).
• Controls: terminal: Working on and on under a condition → Claude usage → enter n (offers 10; only with Claude Code on the computer); strom config set run.gate "claude-usage 10". Without a number, 0 applies.
• Since: research 1.2.0; unreadable usage and not being logged in from 1.25.2
• From research 1.26.0: the menu and strom config set accept only a whole n 0–99 (in 1.25.2 also a decimal, which then stops the run).

R7.5.1 The start of the week

• What is needed is the smaller of n and half the part of the week passed; the whole n holds from 2·n % of the week. Right after the reset a session starts. An example: n = 10, 12 % passed → 6 needed; a session starts at a usage of at most 6 %.
• Since: research 1.15.0

R7.5.2 The extra after waiting

• When the condition waited because the lead was small, it lets a session through again only with an extra of 2 points (at 15: a lead of 17), proportionally smaller at the start of the week (2 · needed / n). Then it lets sessions through one after another until the lead drops below what is needed.
• It remembers the week of waiting in hold.json in its folder; a new week starts without the extra.
• Since: research 1.22.0

R7.5.3 How long it waits

• Until the first of these moments: the week catches up with the lead plus the extra; the last 2 hours of the week begin (only when at least 10 points are left to 100 % or to the cap); the week resets (plus 1 minute). Never past the reset, always at least 5 minutes.
• The answer says the usage, the part passed, the lead, and what lead it waits for.
• Since: research 1.2.0; the last 2 hours from 1.15.0

R7.5.4 The week's cap

• claude-usage 10 --cap 95: at a week's usage of 95 % or more no session starts, in no run, not even after “Start anyway?”. The run ends and says the usage and when the week resets.
• It is asked before each session of every run (R7.4.2). A running session gets a request to finish only at 99 % of the week, even when the cap is lower. --cap without n: the cap holds, and the pace with n = 0.
• Controls: terminal: strom config set run.gate "claude-usage 10 --cap 95". The menu does not accept a cap in 1.25.2; it does from 1.26.0.
• Since: research 1.14.0
• From research 1.26.0: a running session gets the request to finish already at the cap (--cap 95: at 95 % of the week; --cap 100: at 99 %).

R7.6 The subscription limit stops the run

When the agent runs into the limit of its subscription during a session, the session closes and the run ends: subscription limit (resets …). The run does not wait, not even with --loop. This does not depend on the condition.

• Limits: the task returns to the queue; the reset time shows only when the agent said it. Only a condition can wait.
• Since: research 1.0.0

R7.7 An agent not logged in

An agent that is not logged in to its subscription does not do a session. The run ends: the agent is not logged in — run … once and log in.

• Limits:
  • before the first session strom checks this only for Claude Code with a separate folder (agent.claudeConfigDir or CLAUDE_CONFIG_DIR, other than ~/.claude). The run then ends without a session and without asking the condition;
  • otherwise the claude-usage condition finds it (it stops; for a run by hand Start anyway? comes up, R7.4.1), or a session that ends with a login error.
• Since: research 1.25.2; stopping after a session with a login error from 1.0.0

R7.8 Strom archive: a research the agent does not work on

A research can be only a Strom archive: the data are entered in the Strom app, and the research keeps them with their sources, originals and history. The agent does not work on it and researches nothing. Only a person switches the archive on and off, either way; nothing is deleted in doing so.

• Enforced by the code:
  • the agent's work is refused in an archive: a chat (strom chat), working alone, a new line, a person review, reading scans (strom read, strom clips, strom transcripts), reading a document, identifying places. The refusal says in one sentence where research is switched on;
  • switching to an archive parks the open, started and waiting tasks with the reason “kept for later: the research is a Strom archive”. A new task in an archive waits parked. When research is switched on, the tasks return to where they were: to the queue, or back to waiting for a person;
  • the research cannot be switched to an archive while someone is working on it. The session finishes first;
  • switching research on without an agent on the computer offers to install one. Without an agent the archive stays as it was;
  • a person confirms the switch; the agent does not make it.
• What the app sends to an archive is entered as it comes: S2.4. How the app shows an archive: S1.3.
• Why: an archive is for a family tree without an AI agent: the data from the app have their history, sources and a backup (strom pack), and nothing is paid for reading (the code, CHANGELOG 1.20.0).
• Controls: terminal: the research's menu, the first item; strom mode (shows it), strom mode archive, strom mode research. For new researches strom setup asks (What should a new research on this computer be?); it is changed by Settings → New researches: research, or only a Strom archive.
• Since: research 1.12.0; the choice for new researches also with an agent on the computer 1.15.0; reading scans refused 1.20.0

R7.8.1 A research waiting for an agent

A research whose agent is not on this computer (not installed, or the research was moved from another computer) waits for it. Until then it behaves as a Strom archive.

• Enforced by the code:
  • what the app sends is entered as in an archive; nothing is made for the agent, and the agent's work is refused with a sentence saying that the research waits for an agent and where to install it or choose another;
  • the saved mode does not change and the tasks are not parked. Once the agent is installed, the research goes on by itself with the tasks it had;
  • an agent that is working in the research right now (even one other than the one set) makes it a research.
• Why: an archive is the person's choice and strom does not record it for them. A research without an agent only behaves as an archive, so once the agent is installed it goes on by itself with the tasks it had; none is parked.
• Controls: terminal: the menu, the first item (Keep it only as a Strom archive: keeping the data from the Strom app, or installing the agent); strom mode shows the state. The app shows such a research as Archive (S1.3).
• Since: research 1.15.0