user context management

Context management

Steiner keeps the conversation within the model's context window so long tasks do not fail on a hard limit.

How context stays small

1. Delegation. Work done by a sub-agent never enters the main conversation; only its compact result does. See Sub-agent delegation.
2. Bounded sources. Skills and project context each have a size cap, so no single source dominates.
3. Bounded tool output. Tools return a bounded preview. Cut-off bash output is saved to a file the agent can page with read, and grep returns one page at a time. See Tools.
4. Compaction. When the conversation gets large, older turns are summarised and replaced with the summary.

Compaction

Automatic compaction starts when the prompt reaches about 70% of the model's context window. Set a different trigger per model with models.definitions.<alias>.advanced.limits.compaction_threshold_tokens; it must be positive and below the context window. See Configuration.

Compaction is automatic: the agent pauses briefly, summarises, and carries on. After compaction the model still has:

• the system prompt and tool definitions
• summaries from earlier compactions
• the most recent conversation turns, verbatim
• your latest message, verbatim, when it fits
• a note of any sub-agents still running
• enabled skills, re-applied after the summary

Everything else is reduced to the summary, so details can be lost. Each compaction loses a little more. If Steiner warns that a session is likely lossy, start a fresh session and carry over what matters. Keep important state in files or git, not only in the conversation.

Manual compaction

/compact compacts now. Add focus text to steer the summary:

/compact focus on the auth refactor

The focus text is added to the summarisation prompt alongside any models.<name>.prompts.compaction override. Automatic compaction is never steered. /compact is refused while the agent is generating.

Compaction history archive

Interactive sessions and oneshot phases save the messages each compaction removes to history.md in ~/.config/steiner/sessions/<session-id>/. It holds user and assistant text and short tool-call summaries, not tool output or model reasoning. The summary tells the model where the archive is so it can grep and read it for a missing detail.

Use /recall to view the current session's archive. --exec runs and sub-agents do not keep an archive. Deleting a session deletes its archive; a forked session starts its own archive.

Compaction fallback

When the main model's prompt cache has probably expired, a compaction can run its summary on a second model instead. This needs both models.profiles.<profile>.compaction_fallback and a positive compaction_fallback_after on the main model's provider (defaults: 5 minutes for anthropic, 60 for claude_subscription, 30 for openai and codex, off for other types). See Configuration.

If the fallback fails, the compaction reruns once on the main model. The compaction banner notes via <provider>/<model> (<reason>) when the fallback ran, or fallback skipped: <reason> when it did not.

claude_subscription limit

If the first response after a claude_subscription compaction still reports prompt usage at or above 90% of the context window, Steiner stops instead of compacting again and asks you to start a new session.

Skills and context

Enabling or disabling a skill mid-session takes effect with your next message and does not rewrite earlier messages. A disabled skill's text stays in the conversation until the next compaction removes it. See Skills.

Session date

Steiner records the local date when a session, --exec run, or oneshot run starts, for example Current date: 2026-09-13 (BST, UTC+01:00), recorded when this session started., and includes it in every request. It does not change during the session, even past midnight. Resuming or forking a session records a new date.