Operate a caller-selected Gas City 1.4 with upstream registry packs and native run-centered surfaces while keeping GC runtime state out of AgentOps verdicts. Triggers: "using gc", "gas city", "drive the mayor", "dispatch through gc".
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
Operate a caller-selected Gas City 1.4 with upstream registry packs and native run-centered surfaces while keeping GC runtime state out of AgentOps verdicts. Triggers: "using gc", "gas city", "drive the mayor", "dispatch through gc".
Use Gas City only when the caller explicitly selects it. Treat it as a
replaceable execution adapter, not a correctness or completion boundary.
The adapter cannot select AgentOps semantics, issue a binding verdict, or turn factory completion into delivery or validation proof.
Choose the factory first
AgentOps supports both Gas City and the
Agentic Coding Flywheel as external
software-factory runtimes. Use this skill only for Gas City. If the caller
selects the Flywheel, switch to using-flywheel
and its native workflow instead of wrapping it in Gas City.
AgentOps supplies skills and evidence contracts to either factory. It does not
need its own Gas City formula or role pack. Install or link AgentOps skills into
the provider runtime before starting workers; the upstream Mayor, coordinator,
and workers can then discover and select plan, implement, test,
validate, and other AgentOps skills normally.
Gas City 1.4 operating model
Gas City 1.4 is run-centered. The supervisor serves the dashboard and typed,
paginated session/run APIs. Every graph-owning city or rig scope needs its own
core.control-dispatcher; that deterministic worker advances formula control
beads. Agent workers claim routed work. The upstream gc.mayor skill is the
guided coordinator; gc.run-operator launches and supervises formulas.
The normal AgentOps path is:
Install and pin the upstream gascity workflow and rig-role imports.
Add the project as a rig, prepare its stock maintainer runtime, and make
AgentOps skills visible to its provider sessions.
Create a caller-owned source intent bead and hand its id to the Mayor,
which authors the workflow beads and dispatches the upstream build-basic,
continuation, review, or implementation formula that matches the available
artifacts.
Read run, session, bead, artifact, and verdict state. Completion is never
inferred from chat or pane prose.
Prepare and qualify a rig before its first build with the shipped AgentOps
CLI (no repo checkout required):
ao gc prepare --city /path/to/city --rig /path/to/rig
ao gc check --city /path/to/city --rig /path/to/rig
The command verifies the exact official workflow and role pins, snapshots the
upstream validation scripts and schemas unchanged inside the rig's .gc
runtime, installs only small AgentOps-owned wrappers at the formula check
paths, selects an existing Python that can import PyYAML, and links the
AgentOps skills into the city and rig Codex sinks. Skills come from the
enclosing AgentOps checkout when one is present, otherwise from the installed
skills root; pass to pin a different directory. It never
modifies the GC binary, cache, formulas, roles, or upstream pack.
issues only native inspection commands, writes no adapter files, and fails
before model spend when that runtime contract is missing or drifted.
--skills-source
check
prepare also pre-seeds Codex trust for every session directory that exists
when it runs — the city and rig roots, each .gc/agents/** session home, and
each rig worktree root — so a Codex session in one of those directories does
not block on the interactive trust dialog. Both persisted layers are seeded in
$CODEX_HOME/config.toml: workspace trust ([projects."<dir>"] trust_level = "trusted"), without which Codex silently reports that directory as having no
hooks at all, and per-hook trust
([hooks.state."<hooks.json>:<event>:<m>:<h>"] trusted_hash = "sha256:..."),
which is what the pack's per-provider .codex/hooks.json would otherwise
prompt for. Hook digests are read back from Codex's own hooks/list, never
recomputed.
Trust is judged by value, not by the presence of a table. prepare appends
only entries that are missing and refuses, naming the entry, when one exists
but does not confer trust — an explicit trust_level = "untrusted", a hook
Codex reports as changed since it was trusted, a recorded hook Codex still
rejects, or a hook recorded enabled = false (a disabled hook is not a trusted
working hook). It never overwrites an operator decision, and re-running is a
no-op. It also fails rather than continue if Codex returns an empty or
unrecognized hook list. The trust store itself is never edited in place: the
merged content is parsed in memory first, then installed with the CLI's durable
atomic writer, so no failure path can leave a partially written Codex config.
ao gc check verifies the same pre-seed from local state only — it runs no
Codex subprocess and writes nothing, deriving each expected hook key from the
directory's own hooks.json — and names the specific deficient directory or
hook using the same rule prepare seeds to.
Two named limitations.
check cannot detect a stale hash. Because it never asks Codex, a
recorded trusted_hash that no longer matches the hook's current content
reads as satisfied and still raises the trust dialog in a real session. Only
prepare sees that — Codex reports the hook as changed and prepare
refuses. A green check therefore means "trust is recorded", not "trust is
fresh".
Homes created after prepare are not covered. Discovery is by
filesystem marker, so the guarantee covers session directories that exist at
prepare time. A session home Gas City materializes later still carries
untrusted hooks on its first spawn; prepare names the configured agents
that have no home yet.
The operational rule that follows from both: run prepare, start the city,
then run prepare again (it is idempotent) before dispatching.
Preferred pack and registries
The built-in main registry catalogs official packs. The community registry is
optional configuration:
gc pack registry list
gc pack registry refresh
gc pack registry search --all
gc pack registry show main:gascity
gc pack registry add community https://registry.gascity.com/registry.toml
gc pack registry search --registry community --all
search reads the local registry cache; show reports release provenance and
exact import commands. gc import add declares a source/version, and gc import install resolves it into packs.lock. Prefer an exact accepted release for
reproducible cities.
AgentOps prefers the official gascity build pack, the workflow family visible
in the public Maintainer City factory. The current accepted reference is
gascity 0.1.6 at commit
3b3b89f2011e06d84459aa7bea1552382f13930a:
dashboard: https://factory.gascity.com;
workflows: build-basic, build-from-*, implement, review, issue, and PR
flows;
stock rig roles: gc.run-operator, gc.implementation-worker, planners,
reviewers, and publisher;
scope-local formula control: core.control-dispatcher;
guided coordination: the upstream gc.mayor skill.
Install the workflow pack at city scope and its sibling roles pack on every rig
that runs work, following the exact commands returned by
gc pack registry show main:gascity. Keep the stock gc.* namespace; do not
nest or rename the roles behind an AgentOps pack.
Work enters the city through the Mayor. The caller authors ONE source intent
bead with acceptance, then hands the Mayor its id — the Mayor decomposes,
authors the workflow beads, and dispatches. The caller never runs gc sling
itself; an operator-slung run bypasses the coordinator that owns retries,
re-dispatch, and tending for that workflow.
gc bd create "Add a --json flag to the export command"
gc mail send mayor -s "Build ago-XXXX" \
-m "Decompose and launch build-basic for bead ago-XXXX with push=true open_pr=true." --notify
Or, in an interactive Mayor session:
Use skill gc.mayor
A multi-behavior intent arrives as a plan manifest (see plan's manifest mode):
one caller-owned document with stable slugs, dependency edges, and per-child
acceptance and write scopes, tracker IDs left TBD. Mail the Mayor the
document path once it is on the rig's mainline; the Mayor materializes the
epic and children from the manifest and owns their dispatch order. Reference
example: docs/plans/2026-07-30-ponytail-whole-repo-contraction.md.
Direct gc sling remains a debugging tool for a city with no live Mayor; a
run started that way has no coordinator and the operator inherits its tending.
AgentOps skills are tools available to those factory agents, not a replacement
workflow. Explicitly name a skill in the bead or prompt when its behavior is
required. The current upstream decomposition does not automatically propagate a
free-form Required Skills section from the caller-owned source bead into every
generated work item. Inspect the decomposition before implementation; put a
required skill name on the actual work item or worker prompt when its use is an
acceptance condition. Skill presence and skill invocation are different facts.
Upgrade an existing city to 1.4
Before starting its orchestrator, run once per city:
gc doctor --fix
gc import install
gc supervisor stop --wait# macOS when an older direct supervisor remains
gc start
Then confirm:
gc version reports 1.4.0 from the intended path;
gc doctor has no blocking failures;
each graph-owning rig has an unsuspended core.control-dispatcher;
imports and packs.lock resolve;
ao gc check accepts the contained maintainer runtime and
AgentOps skill links;
on macOS, the supervisor LaunchAgent resolves to the same executable as the
selected gc binary;
old standalone-dashboard bookmarks or reverse proxies are removed.
A stale registered city may block every start. Repair that city with gc doctor --fix, or explicitly unregister it if it is intentionally retired.
Retire an old HQ/canary by exact registered name or path, without stopping the
machine-wide supervisor needed by its replacement:
unregister fails rather than silently accepting an unknown target. Preserve
the city directory until its Beads state is backed up or confirmed disposable.
Create the replacement from the upstream Gas City template, install its pinned
imports, and verify it with gc cities --json, gc --city <new-city> status,
and gc --city <new-city> doctor --json.
Orchestrating through the Mayor: the tending loop
After handing intent to the Mayor, the orchestrator runs five verbs. Each verb
has one owner; crossing owners is the recurring failure class this section
exists to stop.
Verb
Owner
Surface
Monitor
orchestrator
$API/runs/census and $API/runs/<run-id> on a fixed cadence, plus gc mail inbox for Mayor replies. failed > 0 in the census, a run in failed/canceled, or unread Mayor mail is the act signal; everything else is a tick.
Observe
orchestrator
On an act signal, walk the visibility layers in order — census, run detail, bead graph, session roster, pane truth — and stop at the first layer that explains. Do not start at pane truth.
Nudge
orchestrator, once
A ready bead: dispatch once to its gc.run_target. A routed bead with a live session: gc session wake <run_target> once. A second nudge on the same subject means the diagnosis is wrong — mail the Mayor instead.
Redirect
Mayor
Priority, scope, cancellation, or model/provider changes travel by mail with bead/run ids. The orchestrator never re-slings, edits workflow beads, or patches a live run.
Rework
GC first, then Mayor
Failed review findings re-enter the run through its native fix loop (review_fix_formula, default fix-loop-base); bounded gated retries are gc converge loops. Only a TERMINAL failed/canceled run — or a completed run whose result misses caller acceptance — goes back: mail the Mayor the run id and the failure evidence for re-decompose and relaunch.
Rework the orchestrator performs by hand (editing a failed run's worktree,
re-slinging its formula, closing its beads) creates a second uncoordinated
author for the same intent; the Mayor's relaunch then races it.
Stall protocol
First classify the bead.
Still ready: dispatch it once to its gc.run_target, then stop and inspect.
Already routed/in progress: re-slinging is a NO-OP. Wake its owning worker
once:
gc session wake <run_target>
Then capture the exact tmux pane named by session state and run gc doctor.
Never repair a city from inside that city.
Never create pack-owned sessions by hand. gc session new for a singleton or
scaled agent (core.control-dispatcher, role workers) makes a mis-scoped
session that squats the canonical name in start-pending and blocks the
reconciler from spawning the real one — extending the exact stall being
repaired. Session lifecycle belongs to the reconciler and demand scaling.
When the city itself needs tending (a stalled reconciler, sessions that never
leave draining, model or provider rewiring), send the request to the Mayor:
gc mail send mayor -s "<subject>" -m "<request with bead ids>" --notify
The upstream pack may leave a future affinity-bound step assigned to a session
that has already drain-acked. Diagnose this only from outside the city:
ao gc recover-affinity --city /path/to/city --rig /path/to/rig
The default is a dry run. If every listed assignment is correct, repeat with
--apply. The bounded repair only clears the assignee on a currently ready
formula bead whose gc.session_affinity=require session is no longer live. It
does not sling, retry, close, restart, or select work.
Visibility: four layers
Supervisor/run state — gc dashboard, run detail, gc status, and
gc session list --json. Run detail unifies the stage ladder, structured
transcripts, token rate, and estimated burn rate. A roster may still report
active while a provider is wedged.
Programmatic run status comes from the supervisor's typed run API — the
same data the dashboard renders. gc status prints the API base; neither
gc status --json nor any other CLI subcommand carries run objects.
Poll run status and census for progress; a nonzero failed count is the
first machine-readable failure signal. The dashboard's run page
(/city/<city-name>/runs/<run-id>) is the human view of the same objects.
Bead graph — gc bd --rig <rig> ready --json and show <id> --json.
This is workflow-state truth, but a claimed bead cannot reveal a wedged pane.
Pane truth — tmux -L <socket> capture-pane -pt <session>. This exposes
trust prompts, update nags, API/DNS failures, and interactive wedges. A pane
parked on Codex's Do you trust the contents of this directory? (or the
later Press t to trust all hooks dialog) means that session directory was
not pre-seeded — the workflow queues dispatches as pending with no active
worker and reports no error. Almost always the home was created after the
last ao gc prepare; re-run prepare, then restart that session.
ao gc check names the untrusted directory or hook before you spend a
dispatch on it. Gas City also appears to auto-answer this dialog by sending
keys into the pane, so a wedge may clear on its own — treat that as a race
you do not want to depend on, not as a reason to skip the pre-seed.
Health machinery — gc doctor, gc order history, storage health, and
events. This proves metabolism, not semantic acceptance.
When layers disagree, trust the more direct observation: pane over roster for a
session wedge, bead/run state over prose for workflow completion.
gc status may return a partial no_agents_running snapshot while
gc session list --json shows a live Mayor or worker. Treat that as an
observability disagreement, not permission to restart. Use session and pane
truth for liveness, bead/run state for workflow progress, and Doctor for
metabolism. A supervisor with abnormal CPU, a timed-out native stop, or a
recurring hook rewrite remains an upstream operational defect; this helper
reports it but never kills or patches GC processes.
The caller-owned input bead and the generated workflow root have separate
lifecycles. A successful build-basic run may close its workflow root while
leaving the input bead open. Likewise, push=false and open_pr=false produce
a successful no-op publish while the approved commit remains in its source
anchor worktree. Neither state is semantic completion by itself.
Boundaries
GC quests, runs, attempts, stalls, cancellations, and internal close state
stay in GC. They never become AgentOps Plan, Candidate, RPI, or verdict state.
A GC close or completed run is not AgentOps completion. Only a fresh Validate
context issues the semantic result or, when requested, persists verdict.v2.
This skill performs no automatic selection, retry, semantic validation, Git,
integration, closure, release, or delivery.
The operator lane into a city is a closed set: author source intent beads,
gc mail (work dispatch and city tending both go to the Mayor),
gc doctor [--fix], supervisor start/stop from outside,
ao gc prepare|check|recover-affinity, and reading state. The Mayor authors
workflow beads and dispatches; creating, scaling, or repairing pack-owned
sessions by hand is outside the lane, and the reconciler owns session
lifecycle.
