Ambient awareness

Stable mode injects a feed of recent sessions into every new session at start. The agent's first context window already contains what the team worked on recently: session IDs, titles, work states, message counts, and the opening lines of each conversation. Paired with messaging, this gives a fleet of sessions ambient awareness of each other — every session knows its neighbors exist, what state they are in, and how to reach them.

Stable mode is enabled through the snippet system's wizard, or directly:

cast install stable      # enable (solo)
cast stable team         # switch mode
cast install stable --disable

Solo, team, off

Stable is a three-way choice, not a toggle:

A global flag widens the feed from the current project to everything you can see. In team mode the feed includes teammates' sessions with their names attached, so a fresh session starts knowing that a colleague's agent is mid-flight on the same subsystem — before it duplicates the work.

How injection works

There are two injection paths, one per agent family, sharing a single builder so the feed parameters and format can never drift:

• Claude Code: cast install stable registers cast stable-context as a SessionStart hook. When a session boots, the hook fetches the feed and prints a <stable-context> block that Claude Code folds into the opening context.
• Codex: the daemon builds the same block and passes it as developer instructions when it spawns the thread.

The injected block looks like this:

<stable-context mode="team">
This gives you bigger-picture visibility on what has been and is
being worked on by the team.

── Session binding Convex error ────────────────────
   jx7az96 | ● needs input | 41 min ago | 408 msgs | ~/src/codecast
     1: [user] …first line of the conversation…
     2: [assistant] …first reply…
…
Use: cast read jx7az96 <range>    # read messages by line range
</stable-context>

Each entry carries the session's short ID, so the agent can immediately cast read any of them for detail — or cast send one a message. The feed is oriented toward action, not decoration: "needs input" means the ball is in someone's court, "working" means an agent is mid-flight, and the agent reading the feed is expected to use that.

Recorded, visible, failure-safe

What was injected is recorded against the conversation, and the web renders it as cards at the top of the transcript — so a human reading the session later sees exactly which sessions the agent knew about at boot.

Injection is an enhancement, never a boot blocker. The fetch has a hard timeout, and any failure — network down, backend unhealthy, not authenticated — means the session simply starts without the block. Specific sessions can be excluded from the feed (the daemon uses this to keep a spawned worker from seeing its own siblings' noise when that would mislead it), and the fetch over-fetches by the exclusion count so the feed never shrinks below its normal size.

Why this changes fleet behavior

Without stable mode, a session knows only what its prompt says. Coordination has to be pushed: someone (human or orchestrator) must tell each worker about the others. With stable mode, coordination can be pulled: any session can notice a neighbor in the feed, read its transcript, and message it. The messaging guide shows the patterns this enables — delegation, peer collaboration, and fleets where workers find each other by label. Awareness comes from the feed; action comes from cast send.