agy-worker
Skill2 starsChanged 45 days agoMITView on GitHub
Use when Codex or Claude Code should delegate repository exploration or implementation to Google Antigravity CLI (agy), then review, verify, repair, and deliver the result.
2 other projects use a version of this.
Adopted by
Most recently taken up by
- cagdasyurekli/agy-workerword for word, first seen Oct 17, 2026
- swjturay/codex-agy-delegatortheir own version, first seen Jul 25, 2026
Other versions2 projects use a version of this.
Word for word · 1
- cagdasyurekli/agy-worker★ 2100% same
Edited · 0
None found.
Wrote their own version · 1
- swjturay/codex-agy-delegator★ 4same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repositoryOne other file it gives its agents.
- AGENTS.mdAGENTS.md
The file
Show its rules and text
We split this skill into the rules it gives an agent
14 rules. About 0 are common, 4 are distinctive, and 10 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈4
Distinctive · estimate
This author's own choices. Worth a look.
10
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 24
Use "$PIPELINE/doctor.sh" --repo /absolute/path/to/target for offline readiness.
- Very specific: line 37
Prefer --provider-scope FILE --approve-transmission-sha SHA256 for bounded jobs.
- Distinctive: line 71
Keep raw logs and controller state outside the worktree and prompts.
- Distinctive: line 76
Read Security and compatibility before a first live dispatch or changing execution exposure.
- Very specific: line 100
Use workflow.sh run --preview, approved run, read-only status, and verify-finalize.
- Very specific: line 122
Use paths that exist in the reviewed worktree and keep the scope file outside it with mode 0600.
- Very specific: line 126
Run driver-selected build/test commands and Python imports in an isolated verification copy with PYTHONDONTWRITEBYTECODE=1.
- Very specific: line 140
Use initial --allow-scoped-repair for approved multi-turn scoped work.
- Very specific: line 164
Stop for missing exact approval.
- Distinctive: line 167
Never use dangerous permission/approval-bypass flags or disable the host sandbox for AGY.
- Distinctive: line 168
Do not modify user configuration as a code change.
- Very specific: line 170
Before changing AGY-facing flags or claims, run "$PIPELINE/ground-truth.sh" and inspect current help.
- Very specific: line 172
Read result.structured_output, not echoed schema or empty display text.
- Very specific: line 178
When status offers finalization for a bound candidate, verify-finalize records the driver's result without a Git change and is the facade's …
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Delegate repository work and verify the result
Delegate substantive exploration or implementation to agy, then inspect its diff
and run driver-owned checks. Unknown files, architecture or a first test command do
not prevent useful exploration; worker envelopes are not evidence.
SKILL_ROOT contains this file. In Claude Code it is ${CLAUDE_SKILL_DIR},
substituted in instructions rather than exported by Bash. Resolve the package:
PIPELINE="$(bash "$SKILL_ROOT/scripts/resolve-pipeline.sh")" || exit $?Use "$PIPELINE/doctor.sh" --repo /absolute/path/to/target for offline readiness;
ready proves neither authentication nor task success. See the Package README.
Authorize provider work
Obtain human approval for the task, exact provider-readable content, transmission and isolation modes, caller-selected model and budget. One upfront approval may cover predictable same-scope repairs and mechanical digest refresh. New scope, content exposure, destination, isolation, permissions or budget needs fresh authority. SHA values bind the documented controller inputs, not the entire human decision; refreshing a still-applicable binding is not another approval request. Goal and hand-authored JSON are not ordinary-use prerequisites.
Prefer --provider-scope FILE --approve-transmission-sha SHA256 for bounded jobs. It binds reviewed read entries, their content digest, and a write subset in a fresh owner-private mode-0700 Gitless stage.
Whole-worktree dispatch requires --approve-whole-worktree LAUNCH_APPROVAL_SHA256. Every disposable-worktree entry is provider-readable and may reach Google/Gemini; --add-dir, prompt denylists and gate path policies do not narrow that read boundary.
Neither workflow.sh run nor the advanced agy-worker.sh initial dispatch has an implicit transmission mode or an implicit --provider-isolation.
The whole-worktree digest binds content, kinds, full file mode bits, symlink target hashes, the
readable manifest, provider isolation and native grant profile. The scoped digest
binds canonical read/write policy, readable path/kind manifest, selected bytes and
executable bits, isolation and grant profile; scoped mode rejects symlinks and does
not bind full POSIX permissions. The controller rechecks the approved boundary
before provider start. The human approves one launch_approval_sha256 over canonical
launch_authority: those content inputs, destination and Git base, exact normalized
task and constructed prompt, fixed transport templates, workflow/edit mode, model/effort,
cycles and time budgets, scoped repair, self-verification manifest digest, provider-env
names, slash policy, additional directories and provider schema digest. Preview the
exact task and all launch options; reuse them unchanged for the approved run. Only
trailing LF bytes are removed from UTF-8 task input; spaces and CRLF are preserved.
Environment values and private verification commands are never serialized into the
approval payload. Use the private prelaunch review checklist
in Project lifecycle and verification, including any optional self-verification
commands, IDs and limits; keep that manifest out of the provider preview and prompt.
Retired dispatch and workflow job formats are rejected; finish or discard them with their creating release, without migration.
--provider-isolation session|native is caller-owned with no implicit value: the driver names the mode and its consequence in the approval request and never selects session on its own. session uses the existing AGY session with normal user filesystem/network authority plus ambient HOME, enabled user plugins, hooks and MCP servers; staging and reconciliation are not host isolation.
The controller compares a bounded observation of the verified main source checkout
against one immutable job baseline before later launches and before candidate
admission or retention. No-candidate failures and cancellations without a report
defer the post-exit comparison until the next launch; a cancelled repair defers it
until its inherited candidate is next used. A detected change blocks
the candidate; ignored untracked files and writes elsewhere on the host are outside
this check. The main-workdir helper skips this comparison, while public launch
preview still requires a linked disposable worktree. See
Security and compatibility.
Explicit --provider-isolation native requires supported macOS scoped containment; it never falls back to session mode. Preserve the recorded isolation mode and grant profile across repairs.
Provider-scope approval grants neither provider execution, Git action, driver acceptance, nor publication.
Exclude secrets, denied paths and unrelated private content from every approved entry; telling the worker not to read an approved entry is not a control.
Keep raw logs and controller state outside the worktree and prompts. Each launch requires capability preflight and immediate executable-binding recheck. Model and effort stay caller-owned. Child environment opt-ins require approval per variable name. Installation grants no provider or Git authority.
Read Security and compatibility before a first
live dispatch or changing execution exposure. It owns native network/Keychain limits (native needs a one-time dedicated agy keychain),
verifier environments and no-follow boundaries. AGY 1.4.2 was observed starting workspace MCP commands from .agents/, _agents/, .agent/ and _agent/. Preview rejects known
executable configuration under all four roots. Each launch checks actual cwd ancestry in
both modes. Prefer scoped mode excluding it, then a disposable worktree without it;
ancestors must be clear. Session mode can start enabled user-plugin MCP servers and run
global and enabled-plugin hooks in headless mode with slash commands disabled; hooks can
observe or rewrite tool calls, including shell commands. The guide details reserved paths,
native plugin/hook exclusion evidence and limits. Read the required
launch notices
before initial, resume, continue or restart attempts; notices do not repeat approvals.
Choose and run the workflow
| Intent | Workflow | Cycle budget | Driver action |
| Explore, understand, review or plan | explore | default 2, allowed 1..2 | Spot-check findings; state coverage limits. |
| Bounded feature, refactor or tests | task | default 2, allowed 1..2 | Review the diff; run relevant checks. |
| Project build or broad audit-and-fix | project | default 5, allowed 1..5 | Review changes; run build, tests and lint. |
Select only needed files and direct tests, name exact files/functions and the expected change, size the budget to that scope, and split large changes into bounded jobs. See the maximum-runtime observation and cleanup test.
Use workflow.sh run --preview, approved run, read-only status, and verify-finalize; the facade does not choose a model, assurance, repair or external action.
Project lifecycle and verification owns copyable commands, Verification v2 candidate bindings and recovery.
Quick path: save the reviewed scope file outside the target repository with mode 0600, then run the facade with a unique job ID:
TARGET=/absolute/path/to/approved-repository
JOB_ID=job-12345
SCOPE=/absolute/private/provider-scope.json
TASK='the exact approved bounded task'
"$PIPELINE/workflow.sh" run --preview --repo "$TARGET" --job-id "$JOB_ID" --provider-scope "$SCOPE" --provider-isolation session --task "$TASK"This creates the disposable worktree and private state. Review launch_approval_sha256,
then repeat the same command without --preview, adding
--approve-transmission-sha "$LAUNCH_SHA". A minimal owner-private
scope file, with entries sorted by path, is:
{"schema_version":1,"kind":"agy-worker-provider-scope","read":[{"path":"src/parser.py","kind":"file"},{"path":"tests","kind":"tree"}],"write":[{"path":"tests","kind":"tree"}]}Use paths that exist in the reviewed worktree and keep the scope file outside it
with mode 0600. The facade derives an owner-private state path for later
status and verify-finalize; see the linked guide for the complete sequence.
Run driver-selected build/test commands and Python imports in an isolated
verification copy with PYTHONDONTWRITEBYTECODE=1. Inspect the bound candidate's
actual diff and gate binding directly. Never manually edit, delete, or chmod the
bound candidate; send needed
repairs to the same worker conversation. Never execute an envelope's
commands_run or tests_run; bind only sanitized driver findings to the candidate.
Gate and verify-finalize --verify-argv verifiers run in the bound candidate: choose
only commands known to be read-only there. Snapshot rejection detects changes to bound candidate state
after they happen; it does not isolate the candidate. Verifiers run untrusted
candidate code with the user's authority; the gate does not detect changes outside
the candidate, including Git hooks and configuration. Feed separate copy test results
into Verification v2.
Repair observable failures in the same conversation within budget; never silently
fall back to direct-driver work after provider failure or exhausted budget.
Use initial --allow-scoped-repair for approved multi-turn scoped work. Without that
grant, a changed scoped candidate is result/finalize-only. Preserve useful work;
restart requires an explicit user decision.
Self-verification is optional advisory feedback, not acceptance. An unknown first
check is not a hard stop. Reuse checks only for unchanged candidate bytes and relevant
environment; the driver independently decides the final assurance.
For trust-boundary changes, use the short design note and independent review in
Material planning governance.
Explicit delegation-first requires running the delegation-policy.sh evaluator before substantive repository work.
Controller records are local: the runtime cannot infer prior work or approval and
must never silently authorize direct-driver fallback after missing approval, a hard
stop, preflight/provider failure or exhausted budget.
Claude Code: in non-interactive claude -p or SDK runs, dispatch in the foreground
(up to the host's ten-minute ceiling) or keep the turn alive until it finishes;
never end the turn while a job runs. In interactive sessions, a long dispatch may
use Bash run_in_background: true while the main session remains active. Do not
duplicate a job or treat a Bash timeout as provider failure.
Bash permission approval is separate from transmission approval. Read
Claude Code host operation.
Hard stops and delivery
Stop for missing exact approval; secrets or denied/unrelated private content in
approved inputs; writes escaping the worktree, entering .git or traversing symlink
boundaries; or unapproved Git, publication, installation, account or external actions.
Never use dangerous permission/approval-bypass flags or disable the host sandbox for AGY.
Do not modify user configuration as a code change. Never weaken the evidence gate.
Before changing AGY-facing flags or claims, run "$PIPELINE/ground-truth.sh" and
inspect current help. Its default version/help phase is local; --account is separate.
Read result.structured_output, not echoed schema or empty display text.
Deliver verified, partially_verified, rejected or blocked with the checks
actually run and remaining gaps. Offline checks do not prove live provider behavior,
completeness, release state or general correctness. Use
Troubleshooting for the failing boundary.
When status offers finalization for a bound candidate, verify-finalize records the
driver's result without a Git change and is the facade's required closure step.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

