Model limits and session transfer

Each model has a limit: subscription quota, token package, limit on requests per minute. Pulse shows it in chat and, when the limit is close, helps you continue working on another model or with another agent without losing context.

Chip limit in chat

A chip appears next to the token consumption in the session composer and in the graph panel: “limit 77% · 2 hours”, “balance 1.2M”, “limit exhausted”. Color - the same thresholds as for context filling: yellow from 60%, red from 85%. Hover your cursor to see all limit windows at once in a pop-up tooltip.

How does Pulse know about the limit:

Agent Source What is visible
Claude Code subscription quota comes in ACP adapter overhead 5 hour window, weekly on all models, Opus and Sonnet weekly windows, overage credits: percentage and reset time
GigaChat GET /balance after each turn and on the “Check” button the balance of package tokens for each model; 402 from the server - “model balance is exhausted”
Local model and other OpenAI-compatible headers x-ratelimit-* in the answer (OpenAI, LiteLLM, vLLM behind the proxy) tokens and requests per minute, reset time
Codex the adapter does not send limits outward no chip

The limit applies to the account, not to the session: all agent sessions see the same value, and the agent card (Settings → Agents and models) contains a badge. When the window goes to “critical” or “exhausted”, a notification comes and an entry appears in Activity.

If you have a subscription without a quota or the server does not report limits (Ollama, LM Studio), the chip simply will not appear - this is not an error.

Chip menu: what to do when the limit is running out

The menu shows all limit windows, advice and two ways to continue:

  • Another model of the same agent - normal model change. The session context is preserved, there is no need to transfer anything. Helps when the window of a particular model is exhausted (for example the Opus weekly window).
  • Transfer to another agent — with a transmission report, see below. It is necessary when the limit is common for all agent models.

There's a link there too "What eats the limit..." — she opens Settings → Consumption.

Report-transfer

The new agent didn't see the conversation, so Pulse doesn't throw him into the void. The order is:

  1. The former agent is asked to write a report: the purpose of the task, what has been done, the current state, how to check what is left, pitfalls. It does not touch the files.
  2. The progress stops, the session switches to the new agent and the selected model.
  3. The card “Transmission: Claude Code → Codex” is added to the chat - the report inside it opens when clicked.
  4. Your next step goes to the new agent along with this report.

If the previous agent no longer responds (limit exceeded, error, silent for more than two minutes), a report on the transcript is written by the service model from Settings → AI — for this you need consent to send the code to the AI provider. If it is not there, the report contains a mechanical summary: the last request, the last response and the error.

Changing an agent with an agent chip works exactly the same - it's the same transfer.

The session’s working folder is shared, so the edits of the previous agent are already in place: the new one continues with them, rather than starting over.

Settings → Consumption: what eats up the limit

The page answers not “how much was spent”, but “why”. Period - day, week or month.

It is calculated approximately and only based on sessions on this machine: work on other devices and outside of Pulse is not included here. The context at the time of the request is taken as a measure of the consumption of one turn - each turn sends the entire context to the model anew, so long sessions are more expensive than short ones with the same number of turns.

The reasons are independent: one turn falls into several at once, so the shares add up to more than 100%.

Reason What does this mean What to do
Turns with more context (from 150k or more than 60% of the window) the context is sent in its entirety every turn start a new task instead of continuing a long session
Sessions where subagents worked each subagent has its own requests and its own context for simple subtasks use a cheaper model in the agent version
Sessions longer than 4 hours the context accumulates, the limit is spent in the background close a task when the topic has changed
Turns in 3 or more parallel sessions parallel sessions share one limit run one at a time if the result is not needed immediately

Below is “who spends” (agent and model, share, tokens, cost where the agent reports it) and “the most expensive tasks” for the period.

Money limit: what it really is

The task and swarm cost limit (Settings → Management, budget chip in the task) is not the same as the provider limit above. And he promises less than he seems.

It works retroactively. The cost of a turn is known only from the provider’s report, and the report comes after the turn. This means there is nothing to stop the waste before it happens: the limit refuses next progress, and the one that crossed the threshold has already been paid for. This is not a calculation error, and it cannot be corrected - the price comes later.

For some agents it never works. The built-in Pulse agent - both the local model and GigaChat - does not report the cost at all: GigaChat gives away the rest of the token package, and not the price of the turn. The dollar limit set there looks like protection that is not there. For Codex and Kimi, the price did not appear in the observations, and this was not checked separately; it is impossible to say in advance whether the limit will work or not. What condition your agent has is shown by the line “Informs the cost” in his card: see. Agents and models.

What to limit instead. These three limits do not depend on the provider’s report and work for any agent:

Limit Where is it set? What does
Tokens (task and swarm) Settings → Management stops based on token consumption; The built-in agent has an accurate measure, ACP has an estimate based on context snapshots, marked with the sign “≈”
Turns budget chip in task does not allow you to start a turn beyond the number
Agent minutes budget chip in task stops based on time worked

The amount of expenditure where the price is not known for all turns is written with the word “from”: this is the lower limit, not the total.

Tokens on the graph

In Agent Graph, under the session node there is the context that the model is currently holding, and under the turn node - how much the context added this turn (“+4.9k”). So you can see exactly what turn inflated the session.

Adjacent

  • Agents and models — context window, keys, agent options.
  • Management and budgets — task cost limit: it stops sessions, unlike the provider’s limit. What he doesn’t promise is in the “Money Limit” section above.
  • Memory and context - how Pulse compresses the history as it approaches the window.