Claude host PR 3: Claude projection v2 (member subagents, skills, hooks, watch preset) (#2438 Phases 2, 3, 4, 6) #2442

Part of #2438 (PR 3 of the epic's PR plan, Waves 2–3). Starts after PR 1 merges. Can be developed in parallel with PR 2; they don't touch the same files.

Goal

"Claude projection v2": Claude Code gets per-member subagents, skills, a slimmer coordinator, opt-in governance hooks and a squad watch preset. Bundled so the Claude projection version is bumped once (harness-adapters §5), with one rollback suite and one upgrade diff for users.

If this gets too big to review, split Phase 3 (hooks) into its own PR for the security review. Hooks are opt-in and advisory, so a split needs no extra version bump.

Settle before implementing

Record the answer in #2438 or the PR description:

• Open question 1: the charter extension namespace for host settings (charter §8): one per host (e.g. X-claude-*) or a single Squad-registered namespace. squad is reserved for the spec project.

Phase 2: charters compile to Claude member subagents

Charter v0.1 §6 and Execution Preferences v0.1 §2 map onto Claude subagent frontmatter:

• Add a Claude target to packages/squad-sdk/src/agents/charter-compiler.ts (today it emits only SquadCustomAgentConfig). (Landed in client/adapters/claude-members.ts, which uses parseCharterMarkdown from the compiler.)
• Carry host-only settings in a charter extension: a tools allowlist (e.g. read-only Scribe and Fact Checker) and isolation: worktree for implementers. These are runtime config, not authorization (charter §13.1).
• squad cast and squad upgrade --client claude write one managed .claude/agents/<id>.md per member, plus the built-ins (Scribe, Ralph, Rai, Fact Checker). Only when claude is in clients[]. If an unmanaged, user-authored file already exists, skip it and report; never overwrite.
• The Claude dispatch section routes to named members, falling back to today's generic subagent with the pasted charter when a member file is missing.
• Extend harness-adapters §7.1 / §7.3 supportFiles with the member files.

Phase 4: skills projection and coordinator slimming

• Write templates/skills/* into .claude/skills/.
• Move on-demand sections (Ralph mode, PRD mode, ceremonies, Issues mode, casting) out of the Claude coordinator into skills. The shared core digest and .github/agents/squad.agent.md do not change.
• Keep skill precedence (.squad/skills/ first) and the untrusted-content rule.

Phase 3: governance through Claude Code hooks

Off by default, behind a .squad/config.json flag. The first release runs lockout in advisory mode (log and warn, never deny).

• PreToolUse: reviewer lockout per governance-review §6. An Edit/Write by an author in the review subject's lockout set is flagged (advisory) with the reason from squad_state. (Advisory, as written: the reason comes from the durable lockout record governance/lockouts.json, managed by squad lockout.)
• SessionStart: run the squad_state_health version/dir handshake and inject roster and routing context. (Injects the roster, routing and the version and team dir that squad_state_health must report; the coordinator runs the handshake.)
• SubagentStop / Stop: emit the coordination-handoff result envelope, trigger Scribe's inbox merge when the inbox is non-empty, and apply the Ralph circuit breaker. (The envelope comes from PostToolUse on Agent: Claude Code delivers SubagentStop context to the subagent, not the parent. A background launch is an acknowledgement, not a result. Stop handles the Scribe inbox and the Ralph circuit breaker.)
• Project the hooks into .claude/settings.json as a Squad-owned, additively merged block (same rules as .mcp.json; never touch user hooks). This makes the advertised native-hooks token in packages/squad-sdk/src/client/adapters/capabilities.ts true.
• Tests: lockout flag/denial path, an override recorded by an authority, the deadlocked state, and hook removal on rollback.

Phase 6: watch preset

• Add a built-in Claude preset for squad watch --execute (packages/squad-cli/src/cli/commands/watch/), e.g. claude -p {prompt} --agent squad --mcp-config .mcp.json --output-format stream-json. Flip ralph-cli-triage-watch.claude from degraded in the matrix. The default runner is unchanged. (--agent-preset claude. The MCP config holds only squad_state, resolved by the CLI, never the repository's .mcp.json.)
• Separately, as research rather than code: find out whether gh-aw supports a Claude engine and record it in Epic: Claude Code as a first-class Squad 2.0 host, and the non-Copilot host that promotes the v0.2 profiles #2438. If it does, re-evaluate the gh-aw.claude cell. (Recorded: Epic: Claude Code as a first-class Squad 2.0 host, and the non-Copilot host that promotes the v0.2 profiles #2438 (comment). The gh-aw.claude cell was re-evaluated and stays n/a, with corrected notes.)

Projection version

• Bump the Claude projection version once for the member files, skills and settings.json block. (CLAUDE_PROJECTION_VERSION = 2.)
• Byte-preserving rollback tests for every new file (harness-adapters §15.7).

Delivered in #2450.

Done when

• All checkboxes above are ticked (or Phase 3 is split out to its own PR).
• The new tests pass, and the regression guards from PR 1 stay green: Copilot output is byte-identical, and repos without claude in clients[] see no diff.
• The matrix has no Claude cell unsupported or degraded where Claude Code has a native primitive (model selection, Ralph watch, hooks).