Conductor
Every other role is seated as a subagent and accumulates lessons in its own file. The conductor is the role the main session adopts while parallel agent lanes run — and it was seated nowhere, so its learnings existed and nothing loaded them. This skill closes that gap: it is the loading mechanism, and the place new lessons land.
ticket-triage schedules across tickets; this governs what happens within a run: is this lane doing what it was told, and what do you do when it is not.
The measure problem
A lane will satisfy the done-when you wrote, not the one you meant. Measured, one repo, one epic: fourteen lanes, zero files moved — every brief said "nothing reads this file any more", every lane achieved that honestly, and the owner’s actual measure (the file has moved) was never stated. The discipline: a done-when phrased as a property ("it is no longer used") is satisfiable by argument; one phrased as an observable ("this path no longer exists", "this command exits non-zero") is not.
Four measured deviation shapes
Each grounded in a real incident, described by what it looks like from outside:
| Shape | The tell |
|---|---|
The control that pins the blocker |
A lane asked to remove an obstacle ships a test asserting the obstacle holds. Route: report obstacles as a file and a line, never as a new artifact. |
The report that is evidence, not fact |
A lane claimed a script repointed — one of its five path defaults was. Verify anything load-bearing before merging: run the artifact, demand red-before-green, diff the branch’s file list against the claim. |
The stale premise |
The lane builds on a "fact" in your brief that is false. Say which facts you measured and which you inherited — a lane that measures will correct you, and that is it doing its job. |
The collision visible only after both merge |
Two lanes each removed one entry from a shared manifest; git merged both cleanly and the result was wrong. Partition by what will collide (shared assertions), not what will be edited. |
Routing and verdicts
A routing table (relay / re-scope / pause the lane not the round / let it run) with an explicit what-NOT-to-relay list; findings flow lane → conductor → lane, never lane → lane. Verdict-reading rules for results you did not run: never take a piped command’s exit code, a pass proves nothing without its coverage report, and never touch a tree while a check runs on it. The parallelism ceiling is a measurement (one 24-core box: five concurrent builds sustainable, six put load at 54 and killed three), not a preference.
Three-tier learning
| Tier | Lives in | Holds |
|---|---|---|
generic |
the plugin’s SKILL.md |
deviation shapes, routing, verdict reading — portable, no repo nouns |
per-repo |
|
this machine’s numbers, this repo’s instruments |
graduated |
|
per-repo lessons that held ~3 runs |
The tier test: strip the repo’s proper nouns — a rule that survives is generic; one that evaporates is per-repo. Numbers are almost always per-repo.
Install
/plugin install conductor@alexmskillsComposes with ticket-triage (its round file logs relays), roles (graduation target), and dev-crew; all optional.
Positioning
A 49-entry verified research sweep (2026-09-12) found nothing else occupying this niche: the ecosystem’s orchestrators are plan-time artifacts or fan-out mechanics, and the closest sibling gates quality by refusing parallelism. Independent validation of the problem: the MAST taxonomy (arXiv 2503.13657) measures 32% of multi-agent failures as inter-agent misalignment and 23.5% as verification failures — the two things this skill manages — and Agent-as-a-Judge (arXiv 2410.10934) measures verify-with-tools at 92% human-alignment versus 60–70% for reading the agent’s report, which is this skill’s "run the artifact yourself" rule, quantified.
Claude Code vs Codex
Tier 2 — needs tool translation.
| Claude Code | Codex | |
|---|---|---|
Supervision |
The main session watches Agent lanes. |
The client’s wait/result mechanism, within its wait limit; keep working locally while lanes run. |
Reconcile and relay |
Lane reports and messages. |
After a wait timeout, reconcile with the agent status tools; follow up an existing lane before considering a replacement. |
The difference that matters: The deviation shapes, verdict-reading and ceiling are about lanes, not tools, so they carry over unchanged.
Install, hook trust and verified limits: Codex compatibility.