Parallel Run Fan-Out

prun (parallel run)

Overview

prun fans a task out into independent units that run in parallel while the current session only coordinates. Every worker is an Agy process running Gemini through the Antigravity CLI, on the Google AI plan authenticated in agy. The coordinator decomposes the task, dispatches the units, gathers their results, reviews their diffs, and integrates. It never runs a unit itself.

No Claude-side workers. A Sonnet subagent, a Workflow agent, or any other Agent-tool worker draws on the same Claude account as the coordinating session, so a fan-out of them spends that account's quota at the rate of the fan-out. That is the pool the coordinating session itself needs, and it drained fast once prun routed units to Sonnet. Codex is not a prun executor either; its higher-cost quota is reserved for the default /vet gatekeeper role. Exact plan buckets can change, so inspect current Agy quota before a large batch.

Relationship to the native Workflow tool

The native Workflow tool fans a task out across Claude subagents under a deterministic script, with structured output, judge panels, and resume. A Workflow run counts against the Anthropic plan's usage and rate limits, and its agents use the session model unless the script routes a stage to a different Claude model.

prun is the fan-out that stays off that account. Its units run on Agy and use the Google AI plan; the coordinating session spends only the small Anthropic amount it needs to decompose, dispatch, read results, and integrate. prun therefore never starts a Workflow or a Claude subagent, not as a unit, a fallback, or a second panel. When the user explicitly asks for a Claude panel, that is a Workflow run the user asked for, and it happens outside prun. A cross-vendor read on staged work is what /vet is for.

When the Agy Gemini group cannot accommodate the next batch, queue or defer units instead of moving them onto the Claude account. Read the meter with agent-quota, including snapshot age and reset times, and do not silently shrink a genuinely parallel task to an arbitrary two or three workers. The dispatcher's own quota route, described under dispatch-task usage, already stops a unit from launching into an empty group.

When to use

Use prun when the task splits into independent units that can run at once (different modules, separate research questions, parallel analyses). Units may be heterogeneous, and there can be many of them: a dozen or twenty in parallel is normal when the task warrants it.

Do not use prun when the task is one sequential unit, or units depend on each other's output, or a unit's result cannot be checked without redoing it.

Executors

Rules:

• Every unit runs on Agy. Research, verification, extraction, cross-checks, and code-writing units in a throwaway clone all go through dispatch-task-agy. The dispatcher gives a unit the same unattended capability as the /vet Agy reviewer, so it can verify numbers, run experiments, and fetch the web. Agy defaults to the newest gemini-*-flash-high the account lists (gemini-3.8-flash-high as of 2026-09) at the CLI's maximum high effort.
• Never a Claude-side worker. Do not spawn an Agent-tool subagent (Sonnet or any other model) or a Workflow agent for a unit, including as a fallback when the Agy pool is short. Those workers spend the coordinating session's own Claude account. When Agy cannot take a batch, queue it or tell the user.
• Codex is excluded from prun. Its quota is intentionally reserved for the /vet reviewer role. Do not route a prun unit to codex exec, even if a legacy dispatcher remains on disk for compatibility with old state directories.
• Session-internal tools stay outside the fan-out. An Agy process cannot use the coordinator's MCP, email connectors, or Artifact tool. Gather what a unit needs from those tools in the coordinating session before dispatch, and put it in the unit prompt; leave a small action that needs them to the coordinator as one inline step. A task whose substantive work needs those tools throughout is not a prun task.
• Keep the Agy pool busy with follow-up turns. Units return at different times. When one returns while others are still running, dispatch a follow-up unit rather than idling, provided the follow-up discharges real work: an acceptance criterion the result left open, a claim it made without evidence, a source it cited but did not fetch, a check it proposed but did not run, or the next independent unit in the queue. A slower sibling is not by itself a reason to invent work. --continue-from <state-dir> resumes the same conversation, so the follow-up keeps the earlier context; a fresh prompt with a fresh result path is the alternative. Record each follow-up in the ledger like any other unit.
• The Claude session stays the coordinator, never a unit. Independent substantive work belongs in Agy workers.

Why Agy alone. Its pool is separate from the Claude plan, it is fast, and it adds an independent model family without spending the higher-cost Codex pool used by /vet. The earlier split put Sonnet beside Agy on the grounds that the two draw on separate pools. They do, but Sonnet's pool is the coordinator's own Claude account, so every Sonnet worker spent the quota the coordinating session runs on, and a wide fan-out consumed it quickly. The coordinator still reviews every result and every diff. Check current quota before a large batch, but do not convert changing meter readings into an arbitrary low worker cap.

Concurrency

The orchestrator decides the unit count autonomously. Partition the task by dependency structure (split only along genuinely independent boundaries) and balanced workload (roughly equal-sized units, each worth a full worker run). High autonomy is the intent: do not target a fixed number, and do not cap artificially. A dozen-plus in parallel is fine when the task genuinely decomposes that way.

Two soft bounds, not hard rules: local CPU/RAM (enough concurrent workers eventually contend and the excess queues) and the headroom of the Agy pool. agent-quota reads the current snapshot of both Agy groups. The usual real ceiling is integration bandwidth, since the orchestrator must read and reconcile every result, so prefer fewer well-scoped units over many tiny ones. Over-splitting into trivial units wastes worker startup and tends to produce thin results. Dispatch in batches that fit the runtime's concurrent-worker limit and the available quota, and leave the rest queued; a runtime's in-flight limit is separate from how many units a run may have in total.

What a unit may do, and the one rule

A unit may read or write code, run commands, and fetch the web, with full access. The single hard rule: a worker never commits, pushes, or runs destructive git (commit, push, branch/tag mutation, reset --hard, clean). Everything else is allowed. The final gate is the Claude session integrating the results and the user deciding; workers never touch the real repo history.

This is enforced structurally, not by trust:

• Read-only / research units run from a per-unit scratch cwd, so accidental writes stay out of the repo. dispatch-task-agy.py does this by default.
• Code-writing units run inside a throwaway local clone of the repo with its remote removed:

  git clone --local -c core.longpaths=true <repo> <clone-dir>   # longpaths: Windows MAX_PATH safety
  git -C <clone-dir> remote remove origin

  The worker edits freely in the clone. An accidental git push has no remote to reach (GitHub / Overleaf stay untouched); an accidental git commit only lands in the throwaway clone. The coordinator reads git -C <clone-dir> diff, integrates the wanted changes into the real tree, and the user approves the actual commit. That is the only gate.

The worker's environment is scrubbed. dispatch-task-agy.py builds the child environment with worker_env(), which drops every name containing KEY, TOKEN, SECRET, PASSWORD, PASSWD, CREDENTIAL, APIKEY, or AUTH, and every name starting AWS_. ANTIGRAVITY_* is kept, because the CLI's own session plumbing lives there. The names withheld from a run are listed in <state-dir>/env-withheld; values are never written anywhere. A unit that genuinely needs one variable through gets it with PRUN_KEEP_ENV=NAME1,NAME2.

This exists because the dispatcher previously passed os.environ.copy() straight through. A worker only has to list its environment once, in an env command or a traceback, for a live key to enter a third-party model's context. On 2026-09-20 a scan found two live credentials in 90 files across 18 Agy conversations over three days. A worker never calls a model, so it never needed them.

Beyond that: the user writes the prompts, the clone has no path to the real remotes, and the Claude session plus the user are the integration gate. Scrubbing removes the credential class of accident; it is not a sandbox, and a worker can still read any file the user can.

Flow

1. Gate: confirm the task splits into independent, checkable units. Else use a single worker.
2. Decompose: write one prompt per unit. State the task; for a code-writing unit, that the working dir is a throwaway clone to edit freely but not commit or push; that the unit writes a result summary to its result file (a fresh path, in one write).
3. Assign: every unit goes to Agy. Gather anything a unit needs from session-internal tools first and write it into that unit's prompt. Pick read-only (scratch) or code-writing (clone) mode, and record the mode in the ledger.
4. Dispatch in parallel: run <python> scripts/dispatch-task-agy.py in the background for each unit. With no --mode it runs accept-edits with --dangerously-skip-permissions in a scratch directory it creates. A caller-supplied workspace, meaning PRUN_SCRATCH_CWD (a throwaway clone for a code-writing unit) or --add-dir (a clone or snapshot the unit should see), requires an explicit --mode accept-edits or --mode plan, so the write-capable mode is a named choice for any directory the dispatcher did not create. --continue-from <state-dir> resumes an earlier unit's conversation for a follow-up turn.
5. Monitor (do not go idle): launch scripts/monitor.{sh,ps1} <state-dir> ... in the background (run_in_background=true) and wait on its completion. It wakes you on the first actionable event: all done, any unit stalled (tail no-growth for PRUN_STALL_THRESHOLD, default 10 min), or any unit failed (FALLBACK result or dead dispatch), printing a per-unit digest. On a stall, surface it to the user with a likely cause (capacity or concurrency pressure; suggest lowering the worker count or re-dispatching) rather than waiting silently; act, then re-launch the monitor on the still-running units until all are done. monitor only observes. The Agy dispatcher relies on the CLI's bounded --print-timeout; it does not scan for or terminate unrelated agent processes. (gather.{sh,ps1} remains for the plain wait-for-all case.)
6. Reconcile, then integrate: before integrating, reconcile the ledger: every dispatched unit must have a non-empty result. If any is missing or empty, do not integrate the partial set; recover the worker's output from its <state-dir>/tail (dispatch-task-agy also salvages the tail into the result file automatically under a FALLBACK header). If no usable result can be recovered, re-dispatch that unit or flag the user. Then the coordinator reads each result plus each clone's git diff, merges the wanted changes into the real tree, runs verification, and asks the user before any commit.

Resolve scripts via this order, first hit wins: skills/prun/scripts/, then .claude/skills/prun/scripts/, then .agent-config/repo/skills/prun/scripts/.

dispatch-task usage (Agy)

<python> scripts/dispatch-task-agy.py --prompt-file <prompt> --result-file <fresh abs result> --unit-id <id>

• Emits exactly one stdout line STATE-DIR <abs-path>; Agy stream events and stderr land in the state directory, the conversation id from Agy's init event is recorded to <state-dir>/conversation-id, and the final response is published atomically to the result path.
• Defaults to the newest gemini-*-flash-high at high effort. The constants DEFAULT_MODEL (gemini-3.8-flash-high) and SECOND_MODEL (claude-sonnet-5-5-high) are family templates: Gemini Flash and Claude Sonnet. The preflight's agy models listing supplies the newest version in each family, so a new Flash or Sonnet release runs without an edit. The effort tier may change too: Agy 1.2.16 retired claude-sonnet-4-6 for slugs that carry the tier, and the old template would still have reached claude-sonnet-5-5-high. The tier is chosen first (the template's own, then high), and the newest version within it second. An older high therefore beats a newer medium. An unset ANTIGRAVITY_DISPATCH_EFFORT follows the chosen slug's tier, because Agy rejects a conflicting --effort. The switch shows as a MODEL-RESOLVE from=... to=... line on stderr and in <state-dir>/quota-note, and <state-dir>/model names the model that ran. Quota routing picks the group first, and floating never moves a unit to the other group. One exception keeps a batch alive after a larger rename. When Agy lists nothing in the Sonnet family, a unit routed there runs on the Gemini default, recorded as MODEL-FALLBACK ... reason=template-unlisted. The Gemini group's quota is checked again first, so a fallback into an empty group exits 75 without launching. A model named in ANTIGRAVITY_DISPATCH_MODEL runs verbatim, and ANTIGRAVITY_PREFLIGHT=off skips the listing, so the templates run as written. ANTIGRAVITY_DISPATCH_EFFORT overrides the effort. The dispatcher passes --effort for the Gemini models only. Agy rejected the flag for the older Claude models. Since 1.2.16 it rejects any effort that conflicts with the tier in a Claude slug, so the flag is omitted for the second group.
• Agy Ultra exposes a second quota group for Claude and GPT-OSS models (Claude Opus 5.5 and Sonnet 5.5 at low, medium, and high, and gpt-oss-120b-medium, as of 2026-10), metered apart from the Gemini group. A unit that names no model goes to whichever group has the freer meter, with the newest Claude Sonnet as the second group's model. A unit is shallow work that either group handles, so the meter decides rather than the model family. The worker is the Agy CLI either way, so a Claude model here spends Agy quota and never the Claude account the coordinator runs on. This is a routing policy the user set on 2026-09-15, after a 198-unit batch spent 77 points of the Gemini five-hour meter in an hour while the second group sat untouched. An agent still does not reach for that group on its own outside this rule: one that did spent 646 generations of it in a day. ANTIGRAVITY_DISPATCH_MODEL disables headroom balancing for the run. The exhaustion rules below still apply to the model it names, including the fallback from an exhausted Claude and GPT group to Gemini.
• The dispatcher checks group quota before launching. The two groups are metered separately, and one dispatch names one model, so a batch aimed at an empty group fails once per unit: on 2026-09-11 four units of a seven-unit fan-out died in a row, each carrying Individual quota reached ... Resets in 34m. Before launching, the dispatcher reads the snapshot agent-quota maintains and decides:
  • No model was named: each unit starts from the Gemini default. When both groups are reported, it moves to the Sonnet model if the second group's lowest remaining fraction is at least 15 points higher, or if Gemini is empty and the second group has quota left. A move on headroom also needs both metered windows of the destination present in the snapshot, since a group entry is its emptiest bucket and an unreported window may be the empty one. An empty own group moves the unit without that evidence, because the alternative is not running at all. Units decide independently, so successive readings can switch the group a batch is using. The line MODEL-BALANCE from=... to=... reason=freer-meter own=... other=... goes to stderr and to <state-dir>/quota-note.
  • A named model in the Claude and GPT group, with that group empty and Gemini not: dispatch the Gemini default instead, and record the swap. The line MODEL-FALLBACK from=... to=... reason=claude-and-gpt-quota-exhausted resets=... goes to stderr and to <state-dir>/quota-note. <state-dir>/model always names the model that actually ran, so the ledger's executor column is not the model the caller asked for when the two differ.
  • A named Gemini model whose group is empty: exit 75 without launching, and say so. A model someone chose is not escalated into the metered group on its own; the message names ANTIGRAVITY_DISPATCH_MODEL for the operator who wants that.
  • Both groups are empty: exit 75 with both reset times.
  • A group the snapshot does not report is unknown rather than empty, and an unreadable snapshot skips the check entirely. The gate stops a dispatch only into a group it read as empty. PRUN_AGY_QUOTA_GATE=off disables it. A run that fails at the backend forces a snapshot refresh before exiting, past the readout's own five-minute TTL, because the meter it just hit is newer evidence than the snapshot. Later units then route on what it recorded. This is not a guarantee: a refresh that cannot run, a meter that is unavailable, and units already in flight can still produce repeated quota errors.
• --mode defaults to accept-edits with --dangerously-skip-permissions, the same unattended capability the implement-review Gemini reviewer already runs with, so a unit can verify numbers, run experiments, and fetch the web without a permission prompt. The default applies only to the scratch directory the dispatcher creates. When the caller supplies a workspace through PRUN_SCRATCH_CWD or --add-dir, the dispatcher refuses to launch until --mode is given, because the write-capable mode could otherwise reach a directory it did not create. Safety stays structural either way: point those at a throwaway clone with no remote or a read-only snapshot, never the real tree. --mode plan is the strictly read-only opt-in; it keeps request-review permissions and never gets the skip flag. In headless use a tool request that needs an approval nobody can give (run_command, read_url, browser tools) is denied, and the process can still exit 0 with a normal-looking result that reports it could not verify. A normal result file therefore does not prove those checks ran; read its Verification and Open items fields. If the worker has not written a non-empty result file, a missing, empty, or shorter-than-20-byte final response produces FALLBACK, and so does a final result event whose status is not SUCCESS. A standing permissions.allow rule in Agy's own settings.json (~/.gemini/antigravity-cli/settings.json, entries such as read_url(*) or command(*)) is the alternative for a plan-mode unit.
• --add-dir PATH (repeatable) adds a directory outside the unit's working directory to its workspace without copying a repository into the scratch area, in either mode; it requires an explicit --mode. Point it at a clone or a read-only snapshot, never the real tree, since accept-edits can write there. The dispatcher resolves each path to absolute and refuses to launch if it is empty or not an existing directory.
• --continue-from STATE_DIR resumes the conversation recorded at <STATE_DIR>/conversation-id for a follow-up dispatch that should keep the earlier turn's context instead of re-embedding the prior result in a new prompt. It still needs its own fresh --result-file; an empty STATE_DIR argument, or one whose conversation id file is missing or empty, is a pre-launch error.
• Requires a fresh result path and refuses to overwrite an existing result. The final response is published to that path, unless the worker already wrote a non-empty result file there itself: then the worker's file is kept and the final response lands beside it as <result>.response.<ext>, so a one-line closing reply never replaces a full result. If no non-empty worker result exists, a failed preflight, launch, worker run, or timeout, or an unusable final response, produces an atomic FALLBACK result with captured tails.
• Both signals decide the outcome. A non-zero process exit fails the unit, and after an exit of 0 the final result event's status is consulted, because Agy exits 0 when it stops on a quota limit and that ERROR event still carries the opening narration in response. Publishing that response would hand the coordinator work that never happened. Any status other than SUCCESS fails the unit and carries the event's error text into the FALLBACK result. The one exception is a status that is missing or blank, which counts as success so that an older Agy keeps working.
• A failed run whose worker had already written its own result keeps that file, because the worker may have finished before the backend stopped. The partial response lands beside it as <result>.response.<ext>, and the dispatcher exits non-zero with the backend error on stderr. Both monitors classify a stable worker-written result as done without reading the backend status. Before integrating such a unit, wait for the dispatcher to finish and check its exit code. When that code is unavailable, read the result event's status and error in <state-dir>/tail and the captured dispatch diagnostics. The sibling response is supporting context: a successful run writes one too, and it records no status.
• ANTIGRAVITY_DISPATCH_TIMEOUT_SECONDS defaults to 2700 and is passed to Agy's bounded --print-timeout. The dispatcher never enumerates or terminates another agent process.
• The dispatcher omits Agy's --sandbox flag by default. On Windows that sandbox starts an elevated admin broker and raises a UAC prompt for every unit that runs a command; a declined prompt fails the command. PRUN_AGY_SANDBOX controls whether the flag is added; it does not disable a sandbox enabled in Agy's own settings (enableTerminalSandbox). Accepted values are 1/true/yes/on to add the flag and 0/false/no/off, empty, or unset to omit it. Values ignore case and surrounding whitespace; anything else exits 2 before state creation or launch. Scratch directories and throwaway clones reduce accidental changes to the working repository. They do not enforce filesystem or network isolation; the worker must follow the prompt's ban on commit, push, and destructive git.
• The Codex worker scripts that prun used before 2026-09-13 are archived in legacy/prun-codex-worker/ in the source repositories and are no longer shipped. They remain available if pricing makes a Codex worker the cheaper pool again. The report-state and snapshot-tail launchers recover old unit state through prun_state.py without them.