Read Claude Code sessions as first-class peers in the Agents roster #84

Depends on fleet-instance-presence. This adds a second source to the presence model
that change defines. It does not alter that model's transport, and it adds no new one.

Why

The human runs Claude Code and opencode-skein side by side, all day, across worktrees of
the same repos. fleet-instance-presence makes skein instances visible to each other and
solved half the problem. The other half is that roughly half the running agents are not
skein instances, and to the roster they do not exist. An /agents view that omits them
answers "what is everything doing" wrongly, which is worse than not answering.

The work needed here is small, because Claude Code already publishes exactly this data
locally. Verified on 2026-09-11 against Claude Code v2.1.268:

• claude agents --json prints every live session — interactive and background — as
  {pid, cwd, kind, startedAt, sessionId, name, status}, with --all and --cwd <path>.
  It is a supported, documented CLI surface and does not require a TTY.
• The same data is on disk at ~/.claude/sessions/<pid>.json, which additionally carries
  version, peerProtocol, peerFeatures, messagingSocketPath, entrypoint,
  statusUpdatedAt and the session's display name.

So this change is an adapter, not an integration. Read the CLI, normalize, merge.

What Changes

1. Owner widens from a literal to a union

packages/opencode/src/agent/presence.ts defines
Owner = Schema.Literal("opencode-skein"). It becomes
Schema.Literal("opencode-skein", "claude-code").

This is the load-bearing edit. Everything downstream — /agents, the TUI view, skein's Go
aggregator — already keys off owner, so widening it is what makes a Claude session a
peer rather than a special case. Consumers that cannot handle a foreign owner must be
found and fixed as part of this change, not discovered later.

2. A Claude Code presence source

A new source module reads claude agents --json (preferring the CLI over the on-disk
registry; the registry is the fallback when the binary is absent) and maps each session
into the existing Info record:

Fields Claude does not publish stay absent, never guessed. A roster that invents a
model name is worse than one that admits it does not know.

The session's name is carried through as the display name. This matters more than it
looks: the human's /fanout workflow names sessions <slug>-<role>, so the name is
frequently the only thing that says what a peer is actually doing.

3. Freshness, not liveness

Claude sessions are discovered by polling, so a record is a snapshot. Each record carries
its observation time, and a record whose source process is gone is reported unreachable
— the same distinction fleet-instance-presence already draws between a crashed instance
and a cleanly exited one. Claude's registry is keyed by pid, so liveness is a cheap
kill(pid, 0)-equivalent check rather than a timeout.

4. Work identity via specsync topology

A directory is not a unit of work. Where specsync is available, its
topology -json (specsync change agent-topology-command) is read and each peer is
annotated with the change slug, tracker issue and branch matched by worktree path — for
skein sessions and Claude sessions alike, since the join is on directory, not on owner.

This is the payoff: /agents stops being a list of directories and becomes a list of
who is working on which change. Absent specsync, peers still list with directories.

5. Scope

Local machine only. Claude's registry and sockets are per-user and per-host; there is
nothing to discover over mDNS, so this source is not advertised to the network and does
not extend fleet-instance-presence's network surface.

Capabilities

MODIFIED Capabilities

• agent-presence: gains a second owner and a local, poll-based source for it.

Non-Goals

• No messaging, prompting, or control. Read-only. Every control capability is
  published false. Talking to a Claude peer is claude-peer-messaging, deliberately
  separate because it depends on a private protocol and this does not.
• No writes into ~/.claude/. Nothing in this change creates, edits or deletes a file
  under Claude Code's state directory.
• No reading Claude transcripts, prompts or tool output. Metadata only — the same
  content-suppression rule fleet-instance-presence already enforces, and it is tested
  the same way.
• No claiming or mutual exclusion. Seeing a peer on a slug is not reserving it.
• Not a Claude Code plugin, hook, or MCP server. skein reads a public CLI; Claude Code is
  not modified and need not know skein exists.

Risks

• claude agents --json is a CLI contract that can change. Mitigation: parse
  defensively, treat unknown fields as ignorable and missing fields as absent, and fail
  the source (not the roster) on a shape that no longer matches. A broken Claude source
  must degrade to "no Claude peers", never to a broken /agents.
• Polling cost. Mitigation: poll on roster read with a short cache, not on a timer.
• Path disclosure. Claude cwd values are absolute paths, already the same class of
  data fleet-instance-presence publishes. Because this source is local-only and never
  advertised, it does not widen that exposure.

Impact

• Changed: packages/opencode/src/agent/presence.ts (Owner), and every consumer that
  assumed a single owner.
• New: a Claude source module under packages/opencode/src/agent/, a topology reader, and
  their tests.
• /agents gains an owner column and a slug column.

Disposition (2026-09-18)

Shipped. Shipped. 3.1/3.2 (specsync topology annotation) descoped: specsync topology does not exist; directory + branch are already on every peer.