Start and supervise the Gibson development loop against a wired repo: cost-routed runners (Grok volume, Codex second/review, Claude judgment, Devin merge captain), error budgets, escalation, kill switch, and a plain-English digest for the owner. Final stage of the /gibson pipeline; also standalone ('start the loop', 'run the fleet on <repo>', 'resume gibson development').
Scanned 9/28/2026
Install to Claude Code
npx -y skills add The-AIE/the-gibson --skill gibson-run --agent claude-codeInstalls into .claude/skills of the current project.
Are you the author of Gibson Run?
Add the live security badge to your README — it updates automatically with every re-scan.
[](https://www.skillsdirectory.com/skills/the-aie-gibson-run)More formats (shields.io, HTML) on the badges page.
---
name: gibson-run
description: "Start and supervise the Gibson development loop against a wired repo: cost-routed runners (Grok volume, Codex second/review, Claude judgment, Devin merge captain), error budgets, escalation, kill switch, and a plain-English digest for the owner. Final stage of the /gibson pipeline; also standalone ('start the loop', 'run the fleet on <repo>', 'resume gibson development')."
---
# gibson-run — the loop itself
The loop is `scripts/loop.sh` in the Gibson clone (doctrine: `docs/11`). This
skill's job is to invoke and SUPERVISE it — not to reimplement it.
## Starting
```bash
# default shape (routing table from gibson-resources):
./scripts/loop.sh --runner grok --repo <target-worktree> \
--escalate-after 2 --reviewers codex,claude
# add ONLY after the owner's explicit ACU-spend OK for this run: --supervisor devin
```
**Two preconditions before the first invocation, no exceptions:**
- `<target-worktree>` must be a dedicated worktree/clone — `loop.sh` writes
`gibson/loop-state.md` + journal into the repo it's pointed at, so pointing
it at the canonical checkout violates Law 3.
- `--supervisor devin` makes `loop.sh` run `ensure` (creates a billed session
if none exists) AND every branch handoff spends ACUs even against a live
session. Both are spend: get the owner's explicit OK to enable the supervisor
for this run before passing the flag — a pre-existing live session does NOT
waive that, since the handoffs themselves bill.
- `--runner`: implementer from the routing table (Grok default — flat-rate).
- `--escalate-after 2 --reviewers codex,claude`: two consecutive failures buys
a cross-vendor second opinion BEFORE the error budget kills the lane —
judgment tokens spent exactly when cheap volume has failed twice.
- `--supervisor devin`: finished branches hand off to the cloud supervisor
(docs/22) via `handoff:` in loop-state. **Default handoff = Devin reviews the
handed-off branch and opens/owns the PR but leaves the merge to a human.**
The handoff now pins the exact head SHA (loop-state's `handoff_sha`, or the
remote tip): the driver fetches that commit if this clone lacks it, requires a
distinct-vendor review of that exact SHA against the repo's resolved default
branch, and passes `--sha` so the supervisor rejects the handoff if the tip has
moved. Any failure in that chain blocks the handoff and leaves the branch
queued in loop-state — no review, no handoff. Automated
non-Tier-C merges require the `--merge` handoff mode per docs/22 — and note
honestly: `loop.sh` currently invokes handoff WITHOUT `--merge`, so full
auto-merge needs either that flag wired into the loop or a direct
`devin-supervisor.sh handoff --merge` call. Don't tell the owner merges are
automated until that's true. Tier C parks at the human gate in every mode.
- No Devin wired? Drop the flag; the loop's own review path applies, and the
digest tells the owner what wiring Devin would take.
## Supervising (the orchestrator's own loop)
Check in on a cadence (10–30 min): read `gibson/loop-state.md` + journal tail.
- **Progressing** → log one line, next check-in.
- **Same failure twice across check-ins** → that's a harness gap (docs/09):
capture it as a lesson in `memory/LESSONS.md`, adjust, don't just watch it
burn budget.
- **Loop stopped** (error budget, HALT, or done) → diagnose from the journal;
restart, reroute the failing lane to a different vendor, or report — never
silently leave the fleet dead.
- **Owner-gate items** (Tier C parks, product-direction questions from
builders) → batch into the digest; interrupt only per `docs/14`.
## The digest (the owner's entire view)
Plain English, Ask Contract for anything needing a decision. Shipped / in
review / parked-awaiting-you / blocked+why / cost notes (which pools carried
the load; flag if Claude-cap headroom or Devin ACUs are running hot). Never
show raw diffs or expect the owner to read code — describe behavior changes.
## Hard lines
- Kill switch respected instantly: the `gibson/HALT` file is the permanent stop
`loop.sh` checks every iteration, and `GIBSON_HALT=1` is checked
unconditionally beside it — neither depends on `gh`. The `gibson-halt` label
is only a soft cue, honored when `gh` happens to be available; making a stop
permanent still means the HALT file.
- Law 5 always: reroute so no lane ever reviews its own work, even when a
vendor is down. **The supervisor handoff path is now machine-enforced:
`loop.sh` runs a mandatory distinct-vendor review of the exact SHA it is about
to hand off, and a missing, stale, or failed review blocks the handoff — the
branch stays queued in loop-state rather than reaching Devin. The remaining
gap is the loop's ordinary per-hat path (no supervisor configured): a
same-vendor self-pass there is still only caught by you, so keep checking the
journal each cadence for reviewer-hat entries by the authoring vendor and
reroute.**
- Cost discipline (docs/15): if Grok saturates, overflow to Codex. Claude's
pool buys judgment first and skilled feature work when the capability bar
genuinely requires it — what it never buys is VOLUME work a flat-rate pool
could grind, and never quietly out of convenience.
- One repo, one loop: check `gibson/loop-state.md` for a live claim before
starting a second driver against the same target.
Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.
No comments yet. Be the first to comment!