rlaope/oh-my-hermes / skills/ulw-maestro
ulw-maestro
Verified[omh] Coding owner already chosen, handoff pending: prepares the handoff for the coding agent you already chose, composing its prompt from that agent's own installed skills; never selects the owner and never executes the work itself. Use when the user says: ulw-maestro, coding handoff, prepare the handoff, prepare a coding handoff, hand off the coding work, external executor handoff, handoff prompt, delegation prompt.
Maestro
This is a Hermes-native maestro workflow skill.
Why This Exists
maestro exists so a handoff to an already-chosen external coding CLI carries that CLI's own installed skills, a stated dispatchability boundary, and a captured session id instead of a guessed prompt; absent an explicit coding-owner choice, work runs inside the Hermes harness and no external coding CLI is selected, and this engine only loads once that explicit choice is already made.
Do Not Use When
- No coding owner is chosen yet for this run; the Hermes harness stays the default and this engine never picks one.
- The request is a concept question about maestro, prepared handoffs, or a coding-agent name, or a filename that happens to contain one -- answer directly instead.
- The user wants advice on which coding owner to pick -- ask, don't compose.
- The user is asking whether an owner CAN run right now -- use
omh-executor-runtime-readinessinstead. - The request is lane-splitting or a full delivery cycle rather than one lane's handoff -- use
ulw-work, which enters this engine for lanes with an external owner. - The user wants a coding task handed straight to the Claude Code, Codex, or opencode CLI in the terminal, with no prompt composed from that CLI's own installed skills -- load Hermes' bundled
claude-code,codex, oropencodeskill instead of this one.
Examples
Good example:
- Prompt: $maestro codex already agreed to take this -- compose the handoff prompt for the retry-queue fix.
- Expected behavior: Confirm codex as the accepted owner, discover its installed skills, compose a role-arranged prompt with the required sections, and state the dispatchable handoff mode.
- Why: The coding owner is already explicit and the work needs a skill-aware prompt, not owner selection.
Bad example:
- Prompt: 맡길 사람 아직 안 정했는데 그냥 maestro로 프롬프트 만들어줘.
- Expected behavior: Ask
choose_executorfor the coding owner before composing anything; never pick one on the user's behalf. - Why: No coding owner has been explicitly chosen yet, so composing a handoff would select the owner silently.
Completion Checklist
- The selected coding or runtime owner is named before any implementation claim.
- Prepared handoff, dispatch, execution, verification, review, CI, and merge states are separated.
- The final status cites observed runtime evidence or keeps the work prepared_not_observed.
- When Hermes is the selected coding owner this engine does not apply -- Hermes-native selection uses the Hermes runtime path, never this engine.
- Dispatch never merges: collect each unit's fanout_unit_result/v1 evidence, verify the integrated combination of units (not just each one alone -- disjoint file scopes can still conflict at integration), and report merged/unmerged per unit in the closing brief. Merging the unit branches remains an explicit operator or reviewing-agent action; a dispatch receipt is never merge evidence.
- The closing brief includes the observed
omh_run_summarysummary_text verbatim, or an explicit run-summary not_available line, before any specific next-action question -- never a model-estimated number.
Recovery Notes
- If this request lacks an explicit coding-owner choice, no handoff stands yet; mark prompt composition unresolved and hand back the choose_executor question without selecting a CLI.
- If the scratch-edit permission probe fails, retain the composed prompt and discovered skill names, mark execution not_observed, and hand back the denied permission and required grant change before retrying.
- If dispatch returns no session or thread id, retain the dispatch receipt, mark steering unresolved, and hand back the explicit unsteerable note and any observed unit result without claiming a resumable session.
Workflow Lane
- Current lane: Coding handoff - coding owners, handoffs, review, CI, and merge evidence.
- If intent belongs to another lane, hand back to
omh-routingor name the adjacent workflow. - Shared product, routing, compatibility, and evidence rules:
skill_view(name='omh-routing', file_path='references/skill-common-rail.md').
When to Use
Use once a lane's coding owner is an explicit external CLI and the work needs a prompt composed from that CLI's own installed skills, its readiness and permission checked, and its session captured for steering.
Strong routing signals: `$maestro`, `ulw-maestro`, `coding handoff`, `prepare the handoff`, `prepare a coding handoff`, `hand off the coding work`, `external executor handoff`, `handoff prompt`, `delegation prompt`, `コーディング委任`, `委任プロンプト`, `ハンドオフを準備`, `外部の実行エージェントに渡す`, `코딩 위임`, `위임 프롬프트`, `핸드오프 준비`, `외부 실행기 위임`, `코딩 에이전트에 넘기`, `코딩 에이전트에 넘길`, `передай кодексу`, `отдай задачу агенту`, `подготовь хэндофф`, `подготовь делегирование`, `编码委托`, `移交提示词`, `准备交接`, `交给编码代理`, `siapkan handoff coding`, `buat prompt delegasi untuk agen coding`, `susun instruksi implementasi untuk eksekutor coding`, `siapkan serah terima ke agen coding pilihan saya`, `bantu bikin brief untuk eksekutor coding`, `kemas tugas coding untuk agen yang sudah dipilih`, `chuẩn bị bàn giao coding`, `tạo prompt ủy quyền cho agent coding`, `lập hướng dẫn thực hiện cho trình thực thi coding`, `chuẩn bị bàn giao cho agent coding đã chọn`, `giúp viết brief cho trình thực thi coding`, `đóng gói công việc coding cho agent đã chọn`
Catalog Metadata
Category: execution
Phase: external-handoff
Hermes role: handoff-guide
Quality tier: handoff-gated
Reasoning demand: heavy
Quality bar:
- Planning is not execution permission. Handoff requests authorize preparation only; explicit execution requests authorize only their scope, subject to readiness and permission probes. Clarify missing authority.
- Require an explicit owner choice made in this request: named in it, or confirmed when asked; a default the person set in the setup profile also stands. A choice counts only for the request it was made in: an earlier request's owner, recommendations and plan mentions do not count, so ask again. For a missing, ambiguous or unready owner, ask
choose_executoronce and stop; never choose for the user. - Owner selection alone is not dispatch permission:
prepare a Codex handoff onlyoruse Codex, do not dispatchstays preparation-only.Use Codex to implement this nowsupplies both the owner choice and dispatch permission within scope, with no redundant confirmation. After readiness and permission probes, invoke the fanout-dispatch bridge (omh coding runfor one unit); clarify missing owner or action authority first. - State the handoff mode before composing: the claude-code handoff record is prompt-only (
coding_prompt_handoff/v1-- that record is never dispatchable and never described as a run; a user-commanded Claude Code run goes through the bridge below, and only the fanout-dispatch bridge --omh coding fanout dispatchor itsomh coding runsingle-run entry -- ever spawns a CLI), codex is a dispatchablecoding_executor_handoff/v1, and omx-runtime/pi-family-runtime/omc-runtime arecoding_runtime_handoff/v1. - Compose the prompt from the selected profile's DISCOVERED skills via
omh coding executor-skills --profile <profile>: arrange the returned skills by the unit's role recipe, one named skill per step, using each skill's own invocation string verbatim (/name,/pack:namefrom its manifest,$namefor a codex pack) -- never a guessed prefix. Empty discovery gets one explicit line -- "no installed skills discovered for ; prompt composed generically" -- then compose generically. Loadreferences/executor-prompt-composition.mdfor the full procedure. - A discovered skill is declared, never observed: a
SKILL.mdon disk is evidence the file exists, not that the receiving agent loads, enables, or honours it -- its own registry is the authority. - Hold every composed prompt to the executor prompting contract: the ten required sections in order (Goal, Do, Don't, Known context, Unknowns and decision rule, Expected result, Test, Progress and blockers, Evidence boundary, Task), a greppable
Docs consulted:block (URL plus version, or the explicit none-line), and the six-section session summary shape on report-back. - Keep the composed prompt cache-stable: an invariant head that stays byte-identical across units and re-dispatches, with only the tail varying.
- Before real dispatch, observe execution (a
--versionor no-op call) and read the configured model from the executor's own config or output; a binary on PATH plus an auth file isprepared, neverobserved. Run a bounded permission probe before the real dispatch. - When the user names a model for this delegated run (for example "opus로 돌려줘", "fable로 돌려줘", "use opus"), pass it through
omh coding run's--modelflag (or the unit'smodelfield underomh coding fanout dispatch) using the executor's own accepted identifier -- codex and claude-code both take--model, so an alias likeopusor a full id likeclaude-opus-5-5reaches the CLI unmodified; Claude Code'sopusalias tracks the provider's recommended Opus (Opus 5.5 needs Claude Code v2.1.280 or later; on Microsoft Foundry it resolves to Opus 4.6). - That named model is handed to the executor verbatim, unvalidated; an unknown or unentitled value surfaces as the executor's own observed exit failure, never a silent fallback to the dispatch-model preference or the executor's own default.
- The fanout-dispatch bridge --
omh coding fanout dispatchfor a multi-unit split, oromh coding runfor one unit -- is the only executing surface, explicit per invocation, and it never merges; preparing, composing, or showing a prompt is never dispatch, and a dispatch receipt is never review, CI, or merge evidence. - To carry the files the Hermes session already touched into the handoff, cite
session_file_activity/v1(omh quality-evidence file-activity --hermes-session <id> --json): the workspace files its read_file, write_file, and patch calls named, with the outcome Hermes recorded -- never file-content, diff, test, review, CI, or merge evidence. - Capture the executor's session id at dispatch (
--output-format json->session_idfor Claude Code,--json->thread_idfor Codex) and carry it into every status line; a missing id is reported as unsteerable, never silently attached. - Observe to a terminal state with ONE bounded watcher, never a status poll in your own turns: after dispatch, start
omh coding fanout status --fanout-id <fanout-id> --wait --timeout 600once -- whenterminalis in your tool list or tool_search finds it, asterminal(command=..., background=true, notify=true), otherwise in the foreground. Then do independent work or end the turn, unless the spawn result says this session cannot receive the completion: then wait on that process the way the result names before ending the turn. It re-reads the roster itself and its exit code says why it stopped: 3 when the roster's ownstuck_unitsnames a unit, 0 whenall_units_terminalis true and every unit isverified, 1 when all are terminal and one is not, 65 when they all verified but the journal had unreadable lines (a unit may be missing from the roster), 124 when the timeout passed first. Any other code means the watcher itself did not finish (an error, a crash, or a kill). On any exit, read the roster once withomh coding fanout status --fanout-id <fanout-id> --jsonand act on the units it names; after a code the watcher did not choose, start a fresh watcher only if units are still running. Per unit,terminalis the answer andunit_stateis why -- alifecycle_stateofunit_verification_observedorintegration_ready, or a recordedfailure_diagnostic, is what makes a finished unit terminal -- withlast_event_age_secondsthe time since that unit's last observed output,progress.seconds_since_new_outputthe time since its output last grew, andcapacity.next_actionthe reason a refused unit was refused. A unit that isprogress_stalled,awaiting_input,account_limit,permission_blocked, ordata_missingneeds intervention NOW, not more waiting: a live process with no new evidence is not progress, and re-running under the same account, the same credentials, or the same missing objects repeats the failure exactly. Never end a turn on "waiting for the worker" while a unit sits in one of those states; exit 3 is that moment.all_units_terminalis also false when the roster is empty and when a unit has neither a marker nor a summary row, which is why the watcher has a deadline: on exit 124 a unit whoseunit_stateis stillunknownis a missing record to chase, not a unit to keep waiting on, only a unit stillrunninggets a fresh watcher, and a wait with no bound is the stall it was meant to catch. - A finished dispatch is an event to act on in the same turn, not a status to report: verify that unit's result, record the outcome on the plan (done, or blocked with its reason), then run the recovery or start the next item. Never announce a continuation that has not actually started -- a closing sentence promising the next step, with no dispatch and no plan change in the same turn, is the failure this rule exists for.
- Write every steering delta as more than a restated brief: name the changed constraint, the new evidence, the required action, and whether the verification target moved.
- A mid-run user message is an interjection, not a stop: answer it briefly and, in the same reply, continue the run — re-read the phase todo when one is active and dispatch or advance the next pending step, or name the armed wait it is waiting on -- handle, bound completion signal, deadline -- instead of re-reading status. Only the user's explicit stop or cancel, or the engine's own completion gate, ends the run; when the interjection changes scope, say so and update the declared plan or todo instead of silently abandoning it. A mid-run message is the latest steering for the active task, not automatically a replacement objective: it replaces the objective when the user says so and steers the current one otherwise.
- A follow-up that needs new authority, materially expands the scope, or changes external state not already authorized is described first and started only on the user's approval: the turn ends by naming that next action and asking whether to take it, as one question carrying the choices the user has, never by declaring what will not be done; persistence never broadens the authorized scope. A refused escalation is answered the same way, with a safer alternative inside the boundary or the authorization the boundary asks for — never a workaround or an indirect execution.
- The closing brief scales to the change: one or two sentences for a simple change. Lead with the result and validation; omit abandoned approaches unless they explain a tradeoff, and internal bookkeeping. Disclose requested-but-not-done work and material prepared/unmerged work; omit unrelated non-action footers. Required accounting stays outside this scaling. Order: result -> relevant evidence/limits -> observed accounting (when required) -> specific question (if useful). For a real unresolved user decision or a useful briefing discussion, offer one specific, content-linked question. Completed work needs no question; discussion grants no authority. Never estimate usage.
- Entered from an
ulw-worklane, own that lane's handoff only -- lane framing, disjointness, integration verification, and the closing brief stay withulw-work; report back in that lane's evidence vocabulary. - Close with the localized
omh_run_summarysummary_text verbatim, or an explicit run-summary not_available line, before any specific next-action question -- never an estimated number.
Handoff policy:
Convert an explicitly chosen external coding owner into a prepared handoff: claude-code as a prompt-only coding_prompt_handoff/v1 (never dispatchable, never described as a run), codex as a dispatchable coding_executor_handoff/v1, and omx-runtime/pi-family-runtime/omc-runtime as coding_runtime_handoff/v1. This engine loads only after that choice is made -- absent an explicit coding-owner choice, work runs inside the Hermes harness and no external coding CLI is selected -- and it never substitutes for the Hermes harness path or picks the owner itself.
Executor readiness:
- When accepted work mutates code, check
executor_readiness/v1for the selected Codex, Claude Code, Hermes, or oh-my runtime path before first dispatch. - If readiness is
missingorblocked, ask the user to choose another coding agent, configure PATH, continue in Hermes, or keep a prompt/runtime handoff; retry only after that state changes. - A readiness probe is not dispatch, implementation, verification, review, CI, merge-readiness, or merge evidence.
Delegation transparency:
- When delegating, show the composed delegate prompt in a fenced code block in the status message; truncate a long prompt to a bounded preview ending with
... [truncated, N chars total]. - Name every delegated or parallel lane's model and, when the host exposes it, its reasoning effort inline as
(model effort)in status and briefing lines — including runtime-native subagents; when no effort is exposed, show the model alone as(model)rather than writing a placeholder likeunknownbeside a known model, and never emit empty parentheses. Carry token and elapsed figures the same way in these narration lines: report observed figures and omit unobserved ones — when the user asks for a figure directly, say it was not observed instead of omitting it; a rendered status-board column keeps its ownunknowncell. - Capture a resumable session or thread id at dispatch and report it in the status message: for non-interactive Claude Code pass
--output-format jsonand readsession_idfrom the result (resume withclaude -p --resume <session-id>); for Codex pass--jsonand readthread_id(resume withcodex exec resume <thread-id>, repeating--skip-git-repo-checkoutside a git repo). Never leave a delegate run with no recorded way to resume or steer it. - Before dispatch, grant the executor session every permission the task will need — file write/edit, command/test execution, and the working directory — on the dispatch command itself, not through settings-file guesses: for non-interactive Claude Code pass
--permission-mode acceptEditsor an explicit--allowedToolslist (--dangerously-skip-permissionsonly inside an isolated worktree or sandbox), and the equivalent sandbox/approval flags for other CLIs.acceptEdits: trueis not a settings key and~/.claude/settings.local.jsonis not a file Claude Code reads — user scope is~/.claude/settings.jsonand project scope is the dispatch cwd's.claude/settings.local.jsonwith rules underpermissions.allow. Prove the grant with a bounded scratch-edit probe run before the real dispatch: a permission denial in a non-interactive run recurs identically on retry, so never redispatch until a changed grant is proven, and surface an ungrantable permission as a blocker before dispatch, not after minutes of silence.
Required inputs:
- explicit coding-owner choice made in this request
- task or unit description
- the chosen profile's discovered executor skill set
Expected outputs:
- a composed executor prompt arranged by the unit's role recipe
- the handoff mode and dispatchability state named up front
- a captured session or thread id, or an explicit unsteerable note
Artifact expectations:
- prepared external handoff record when a wrapper can record it
Safety rules:
- Never prepare a handoff without an explicit owner choice made in this request -- a routing recommendation, a plan mention, or an earlier request's owner is not a choice for this one.
- Prepared, composed, or shown is never dispatch, execution, review, CI, or merge evidence.
- Never route a Hermes-owned lane through this engine; the Hermes harness stays the default coding path.
- Never carry a discovered skill's description text into a composed prompt -- only its name and invocation string ever leave discovery; the description stays inside the classifier.
- Never dispatch without an explicit user dispatch command; the fanout-dispatch bridge --
omh coding fanout dispatchor itsomh coding runsingle-run entry -- is the only executing surface.
Runtime Evidence
Preferred harness for this skill: coding-handling.
omh runtime record --skill maestro --harness coding-handling --status started
Record observed delegation results; otherwise return not_available or not_observed.
Prepared OMH routing is not execution, review, CI, merge-readiness, or merge evidence.
- Treat wrapper memory/context summaries as advisory local context, not proof of opaque Hermes memory reads or changes. Preserve workflow intent and stop conditions; verify before claiming completion. Reply in the user's own words and the host's own voice: its SOUL.md persona owns reply language, tone, speech level, and sentence endings, progress updates included (where it sets no language, use the one the user wrote in), and OMH shapes structure and content only. When the requested output is delivered, put any open points inside the result and end with no question or offer; only when the person asked to choose or decide and that choice is still theirs, end with one specific question about it, not a list, a generic offer, or a footer; the shared rail's Reply Language And Host Voice and Turn Ending sections hold the full rule.
Use Hermes-native subagent/delegation features when available: native subagents -> Hermes delegation when available, otherwise sequential lanes.
Shared product, compatibility, topology, memory, harness, and execution rules: skill_view(name='omh-routing', file_path='references/skill-common-rail.md'). Load it when applicable; otherwise name an unavailable capability.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 471k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |