YosefHayim Route Request (mid-orchestrator)
You are a dispatcher, not a second implementation of every skill, and not a second conversation.
Turn messy freeform (typing or dictation) into a ready agent prompt that reuses existing skills.
Default: reuse existing skills. Only suggest creating a new skill when the job is repeated, stable, and not covered (then capture-workflow / skill-from-feedback).
How this should feel (product contract)
The user should not feel interrupted mid-session by a long “routing meeting.”
| Mode | When | What happens |
|---|---|---|
| A — Input-side (preferred) | STT release or “refine before send” | Draft is refined before a real agent turn. Refined text lands in the focused input (same caret STT already uses). User glances, hits Enter, same session continues with the real work. |
| B — Same-turn silent (typed already sent) | Messy freeform already submitted | In this same turn, classify → pick primary skill → execute that skill immediately. No multi-message “here is my routing card, wait for you.” At most one short line: Using finish-and-ship… then do the work. |
| C — Explicit advice only | User asked “which skill?” only | Routing card + refined prompt; do not execute until they say “do it.” |
Wrong: open a separate routing chat, force Cmd+C / Ctrl+V as the main UX, or burn a full turn on a long plan when they already wanted the job done.
Right: refine → input (mode A) or refine → same turn execute (mode B). Clipboard is only an implementation detail inside STT paste (macOS already pastes via clipboard + ⌘V into the caret—user does not manage that).
Relation to existing voice stack
- STT already inserts text with
type_text(clipboard + paste into focused field). - Optional Control double-tap can refine clipboard via
prompt_refinement(“review” mode)—related, but not the full story. - Goal for seamless STT: after final transcript, run route-aware refine, then
type_text(refined)so the input box shows the ready prompt. User proceeds with Enter. Session never “switches characters.”
Product wiring lives in the voice worker / bag config; this skill defines what the refined string must contain. Do not invent a parallel product.
Safety
- Do not invent skills that are not installed or cataloged.
- Do not skip safety of the target skill (no silent main commits, no remote delete, no deploy unless authorized).
- Prefer the smallest skill that fits.
- One primary skill; others are supporting only.
- Never strip quoted literals, paths, URLs, or code from the draft (same spirit as
prompt_refinementvalidation).
Workflow
1. Capture the raw request
Keep user wording. Note workspace if known. Strip secrets from logs, not from the refined prompt the user needs.
2. Classify (one primary)
| Class | Freeform signals | Primary skill |
|---|---|---|
| Ship / branch / PR / report when done | new branch, open PR, commit push | finish-and-ship |
| Whole messy repo / every feature / backup main | multi-feature cleanup | messy-repo-orchestrator |
| Parallel agents / worktrees | many lanes, fan-out | sdlc-tasks-executions setup-lanes |
| Land concurrent lanes | salvage, integrate worktrees | sdlc-tasks-executions land-lanes |
| Local UI prove | launch local, playwright, don’t deploy | preview-and-prove |
| Prod live prove | redeploy, is live, curl | deploy-and-prove |
| Lean / ceremony | deslop, AI slop, ban payload | deslop-v2 |
| Style system (existing code) | CODE-STYLE, grill with docs | grill-me-code-style-with-docs |
| Kill ports | free ports, metro 8081 | kill-ports-local-dev |
| Bootstrap Code folder | clone all GH repos | workspace-bootstrap |
| Cloudflare ops | wrangler, D1 (not prove live) | cloudflare-ops |
| Session skill mining | repeated prompts | agent-session-auditor |
| Fix a skill | skill misfired | skill-from-feedback |
| Bench A vs B | tokens, turns, same tasks | agent-benchmark |
| Unsure / voice dump | which skill | this skill → then primary |
See REFERENCE.md.
3. Build the refined prompt (the deliverable)
The refined prompt is a single paste-ready agent message, not a markdown report:
- Lead with skill trigger when helpful:
$finish-and-shipor clear “use finish-and-ship” - Explicit gates: branch name, open PR vs merge, deploy yes/no, report yes/no
- Exact strings/paths in quotes
- No filler, no “I will now…”, no multi-section essay
Mode A (input-side): only that string matters → inject into caret.
Mode B (same-turn): use that string as the effective request and run the primary skill now.
Mode C (advice): show a short routing card plus the same paste-ready string.
4. Mode A — input-side inject (preferred seamless path)
When the host/product can write the input field (STT worker after final transcript, or a refine-before-send binding):
- Produce refined prompt only (no long card in the agent transcript unless debugging).
- Insert into the focused agent input (existing
type_text/ caret path). - Stop. User edits if needed and submits. No second session.
5. Mode B — same-turn execute (typed/sent messy freeform)
When the user already sent a messy ask in this session:
- Classify in one line max (optional).
- Load primary skill SKILL.md.
- Execute immediately under that skill.
- Do not wait for “ok run it” unless the ask was ambiguous on a destructive gate (merge, deploy, delete remote).
6. Mode C — advice only
If they only asked which skill / refine for later: short card + refined string. No execution.
7. New skill?
Only if repeated, stable, and uncovered → capture-workflow or skill-from-feedback. Else keep route-request + primary.
Verification
Routed (mode A): refined string is paste-ready; injected or ready for input; user can proceed without a routing conversation.
Routed (mode B): primary skill’s verification is the source of truth for “done”; routing was not a separate incomplete session.
Advice (mode C): primary id + refined string + “do not use” list present.
Do not claim success because you wrote a long skill essay. Success is correct primary skill + ready prompt + same-session feel.