[persona-kit 4/8] Migrate workforce CLI to use @agentworkforce/persona-kit

[willwashburn]

Part of the persona-kit consolidation. Depends on #66. This is issue 4 of 8.

Goal

Rewrite packages/cli/ to consume @agentworkforce/persona-kit's top-level orchestration API (buildPersonaSpawnPlan + executePersonaSpawnPlan) instead of the imperative skills/mount/sidecar/input wiring it has today. CLI becomes a thin shell over persona-kit.

After this PR, the CLI's spawn flow reads top-to-bottom:

const loaded = loadPersonas({ cwd, searchDirs });
const spec = loaded.byId.get(personaId);
const persona = resolvePersonaTier(spec, tier);
const plan = buildPersonaSpawnPlan(persona, { cwd, installRoot, envOverrides });
const handle = await executePersonaSpawnPlan(plan, { cwd });
try {
  const child = childProcess.spawn(plan.cli, plan.args, { cwd, env: plan.env, stdio: 'inherit' });
  await waitForExit(child);
} finally {
  await handle.dispose();
}

Everything fancy that's currently in cli.ts (sidecar resolution, mount apply, skills install, input rendering, configfile materialization) is now persona-kit's job.

Files to touch

packages/cli/src/local-personas.ts

• Delete inline persona-loading logic; import loadPersonas and resolvePersonaTier from @agentworkforce/persona-kit instead.
• Keep CLI-specific concerns: persona-picker UX, override-merging UX (printing what got overridden, etc.), warnings collection.
• The cascade-search-dirs configuration stays here (CLI knows about $AGENT_WORKFORCE_HOME etc.); pass the resolved dirs into persona-kit's loadPersonas.

packages/cli/src/cli.ts

• Replace the inline skills install loop with a call to executePersonaSpawnPlan (or runSkillInstalls if the CLI wants its own orchestration for progress UX — pick one and stick with it; default to the top-level executor).
• Replace inline sidecar writes with persona-kit's writePersonaSidecars.
• Replace inline mount apply with persona-kit's applyPersonaMount.
• Replace inline buildInteractiveSpec calls with buildPersonaSpawnPlan (which composes them).
• Replace inline input env-binding with plan.env.
• The CLI keeps: argv parsing (commander/yargs/whatever), agentworkforce subcommands, persona-picker prompts, session recording, routing-profile lookup (still imported from @agentworkforce/workload-router).
• Drop direct imports of @agentworkforce/harness-kit — everything is via persona-kit now.

Other CLI files

• grep for any other @agentworkforce/harness-kit or persona-symbol imports from @agentworkforce/workload-router and rewire to persona-kit.

packages/cli/package.json

• Add @agentworkforce/persona-kit as a workspace dep.
• Drop @agentworkforce/harness-kit (workload-router still needed for routing profiles + persona catalog).

Tasks

• Rewrite local-personas.ts to delegate to loadPersonas + resolvePersonaTier.
• Rewrite the spawn flow in cli.ts to use buildPersonaSpawnPlan + executePersonaSpawnPlan.
• Verify the CLI's UX is unchanged: persona-picker prompts, error messages, and progress output should look the same. (Persona-kit's executor may need a progress callback — if so, define a small onProgress?: (event: ProgressEvent) => void option on ExecuteOptions and add it in this PR. Document the event shape.)
• grep for residual @agentworkforce/harness-kit imports inside packages/cli/; remove all.
• Update CLI's package.json deps.
• Update any CLI-internal types that referenced harness-kit / workload-router persona types — point them at persona-kit instead.

Tests / Smoke checks

• pnpm --filter @agentworkforce/cli test passes.
• agentworkforce <persona-id> for a persona that exercises every field (skills, mount, claudeMdContent, inputs) — actually run it interactively against:
  • claude harness
  • codex harness
  • opencode harness
• Verify the shell-level outputs match pre-migration: same npx skills add commands run, same files written, same harness argv.
• Failure-path: a persona declaring a hallucinated skill source aborts with a friendly error message and no orphan child process before the harness would've spawned.
• executePersonaSpawnPlan(...).dispose() (called via the CLI's exit handler) leaves the cwd in the same state it was in before the run — no .claude/skills/<name>/ directories, no half-written sidecars.

Constraints

• No user-visible UX regressions. If something looked, sounded, or smelled different to a CLI user before/after this PR, that's a bug.
• harness-kit is still in the monorepo (gets deleted in 6/8). The CLI should not import from it after this PR; the package itself stays.
• workload-router still exists and the CLI still imports personaCatalog and routing types from it. Don't touch those imports.

Verification

• pnpm -r build and pnpm -r test pass.
• CI workflows (if any) that run agentworkforce against fixture personas pass.
• grep @agentworkforce/harness-kit inside packages/cli/ returns zero hits.

Reference

Plan section 1.4.