Summary

Restores the delegate-first orchestration workflow that was deleted in #3224 as an unadvertised built-in agent skill. Users who miss the original Orchestrator can type /orchestrate <request> to load the playbook into the conversation; the skill stays out of the default system-prompt skill index so it doesn't clutter the UX for users who don't miss it.

Also pulls the ChatInput placeholder tip carousel from side-questions-2hr3 ahead of that branch landing, with a tip surfacing /orchestrate so users have a passive discovery path for the new hidden skill.

Background

#3224 ripped out the Orchestrator agent wholesale because the scaffolding around it (LLM auto-router, settings UI, agent picker entry, post-propose_plan button, two helper methods on taskService) carried real maintenance cost. The original PR explicitly chose not to migrate users.

But the primitives Orchestrator orchestrated all remained:

• task / task_await / task_list / task_terminate (toolDefinitions.ts:2181, 2289)
• task_apply_git_patch (toolDefinitions.ts:1449)
• sub-agent patch artifact flow + git am integration
• Max Task Nesting Depth setting

So the workflow itself never broke — only the prompt and the chrome that pointed agents toward it. A markdown skill is enough to put the workflow back without resurrecting the scaffolding.

Implementation

Orchestrate skill

• src/node/builtinSkills/orchestrate.md — adapted from the deleted orchestrator.md (recovered from e623bb8af^). Frontmatter sets advertise: false. Body keeps the original sections: delegate-first hard rules, delegation guide (explore / exec / plan / desktop), task brief template, dependency analysis, parallel + sequential patch integration loops, conflict reconciliation playbook (including the git am --abort-only-when-in-progress nuance), and the "never read session storage" guardrail. Adds a prerequisites note about Max Task Nesting Depth.
• builtInSkillContent.generated.ts — regenerated via bun scripts/gen_builtin_skills.ts.
• agentSkillsService.test.ts — two existing assertions enumerate the built-in skill list; both now include "orchestrate".
• builtInOrchestrateSkill.test.ts — new focused test confirms the skill is registered, has advertise: false, and its body still carries the load-bearing directives (regex spot-checks on delegate-first, task_apply_git_patch, dry-run, Max Task Nesting Depth) — wording can drift, but a wholesale gutting of the playbook would fail.

Tip carousel (pulled forward from side-questions-2hr3)

• src/browser/features/ChatInput/placeholderTips.ts — rotating placeholder tip list keyed off a wall-clock 20-minute bucket, so switching workspaces mid-bucket never reshuffles the visible tip. Includes the /orchestrate tip alongside /btw, /haiku, /+high, /compact, /fork, /plan, /goal, /clear --soft, /new, /vim, /truncate.
• src/browser/features/ChatInput/placeholderTips.test.ts — covers single-bucket stability, bucket-boundary rotation, modulo wrap, degenerate-input fallback, and /orchestrate presence.
• src/browser/features/ChatInput/index.tsx — workspace variant uses getPlaceholderTip() instead of "Type a message..." on non-mobile; mobile keeps the plain placeholder because the on-screen keyboard already squeezes the input.

What this PR intentionally does not restore

• Tool restriction. The old agent did tools.remove: [file_edit_.*] so it physically could not edit files. A skill can't strip tools from its caller. The skill prose enforces the rule strongly ("Do not implement features/bugfixes directly in this workspace") but it's a soft constraint.
• "Start Orchestrator" button on propose_plan cards and the planExecutorRouter LLM auto-router. Recreating those is a separate, scaffolding-heavy feature; users who want the post-plan handoff can type /orchestrate Implement the plan. themselves.
• Distinct agent picker entry / Workflow icon. Users stay in their current agent (Exec); the chrome was incidental.

This is the same hidden-skill pattern already documented for /deep-review in docs/agents/agent-skills.mdx.

Validation

• bun test src/node/services/agentSkills/ — 36/36 pass.
• bun test src/node/services/tools/agent_skill_{read,read_file,list}.test.ts src/node/services/tools/skills_catalog_read.test.ts — 89/89 pass alongside the skill suite.
• bun test src/browser/features/ChatInput/placeholderTips.test.ts — 5/5 pass.
• make static-check — green.

Risks

Low. No schema changes, no IPC changes, no migrations. Behavioral surface:

1. One additional unadvertised built-in skill is discoverable. agentSkills.list already returns unadvertised skills, so slash autocomplete surfaces /orchestrate once the user starts typing it, matching the documented behavior of unadvertised skills.
2. The workspace ChatInput placeholder is no longer the static string "Type a message..." on desktop. The text rotates every 20 minutes against a wall-clock bucket; the value is deterministic per bucket so tests / screenshots inside a single bucket are stable.

The legacy config keys (planSubagentExecutorRouting, planSubagentDefaultsToOrchestrator) deliberately removed in #3224 are not reintroduced.

Generated with mux • Model: anthropic:claude-opus-4-7 • Thinking: high • Cost: $11.13