What your agent is told at session start

When a session starts, kcap gives the agent a short briefing before your first prompt. Nothing in it is required; if the server is slow or down, the pieces are simply absent and the session starts normally. Each piece has its own switch

What is in the briefing

Guidance from past sessions. Lessons that evaluations drew from earlier sessions in this repository and that someone approved for injection: a handful of short rules, chosen by weight, and only those marked for automatic delivery. Longer guidance is delivered as skills instead, which the agent loads on demand

The memory index. Titles and one-line descriptions of memories relevant here: this repository’s, its project’s, the organisation’s, filtered to the audiences you belong to. The agent reads a memory’s body only when a task calls for it

Who else is working here. A notice when a teammate’s live session overlaps the same files or the same work item, so the agent can tell you before you collide

Reminders. Short nudges the agent acts on without you asking: to file the session on its ticket, to declare its plan, to present next work when a task wraps up. On a plan that excludes work items, the work-item nudge is suppressed

Once only. The first session after setup gets a note that the hooks are loaded and an offer of the tour. A newly detected coding agent that is not set up gets one offer to set it up, at most every six hours. An out-of-date kcap gets a one-line note

How much it costs

Injection is bounded. A document approved for injection may be up to 4,000 characters; at most three documents and eight individual facts are included; the whole briefing is capped at 16,000 characters. When the budget is exceeded, lower-weight documents are dropped first, then lower-ranked facts, always whole. These are the server defaults; an operator can change them, and the approval dialog in Knowledge shows the current limits

Turning pieces off

Each piece has a config key. Set it from a terminal with the kcap CLI, for example:

kcap config set disable_memory_index true

The setting is per profile, so it applies to every agent and session using that profile on this machine. kcap config show lists what is set and kcap config unset <key> turns the piece back on. The keys:

Piece	Key
Guidance from past sessions	disable_session_guidelines true
The memory index	disable_memory_index true
Who else is working here	disable_coordination_notices true
The work-item reminder	disable_workitems_nudge true
The plan reminder	disable_plans_nudge true
The first-run notice	disable_first_run_notice true
The new-agent offer	disable_harness_nudge true
The out-of-date note	update_check false

Teams control what gets injected from the other side too: a repository’s Knowledge → Settings decides whether only curated facts are delivered or the strongest uncurated ones as well. See Knowledge

Where the agents differ

Claude Code receives the briefing through its session-start hook. Codex, Cursor, Copilot CLI, Gemini CLI, Kiro, OpenCode, Pi, Antigravity, Muse Code, Grok Build and Mistral Vibe receive it through their own mechanisms, and five of them also carry a short steering block in their instructions file that tells the agent when to prefer kcap’s tools: a PR summary before gh pr diff, a session search before git log archaeology, a memory search before assuming there is no prior art. Your coding agents lists which agent has which