version: "1.0.1" name: prun description: Parallel delegation fan-out. The Claude session coordinates (on whatever Claude model is currently selected, e.g. Opus or Fable) while task units run in parallel on workers (never on the coordinator). Codex (codex exec, a separate abundant account) is the prioritized default; Sonnet is reserved for units needing Claude-session-internal capabilities (MCP/email tools, Artifacts, cross-vendor web verification), with the orchestrator deciding per unit. Units may read or write code; workers never commit or push, and the session plus the user are the final integration gate.

prun (parallel run)

Overview

prun fans a task out into independent units that run in parallel on separate-quota or in-session workers, while the Claude session only coordinates. Workers are Codex (codex exec, a separate abundant account, frontier model) and Sonnet subagents (inside the Claude session). Codex is the prioritized default: its quota is separate from the Claude plan and its current model (gpt-5.6 tier) is strong on hard reasoning and code, so most units go to Codex. Sonnet is reserved for units that need something the Claude session uniquely provides (see the Executors rule). The coordinator decomposes the task, dispatches the units, gathers their results, reviews their diffs, and integrates. It never runs a unit itself.

The orchestrator picks the executor per unit; when in doubt, Codex. A Codex unit runs through the separately authenticated Codex/OpenAI account, so the worker run does not draw on the Claude plan at all. A Sonnet unit and the Claude coordinator both consume the current Claude account's quota; the exact split across models and weekly buckets depends on the plan and on active promotions and shifts over time, so check Settings > Usage before relying on any model-specific split. Codex is the default because its worker run is outside the Claude plan; keep Sonnet units targeted because they draw Claude-side quota.

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 has a different quota shape. A Codex unit is dispatched by a shell call to codex exec, so the worker run uses the separate Codex/OpenAI account. A Sonnet unit and the coordinating session both draw the current Claude account's quota, so reserve Sonnet for units that need the Claude session's own tools. The coordinating session also spends a small Anthropic amount while it decomposes, dispatches, reads results, and integrates.

The two relate in two ways, both with the current session as the orchestrator:

• Substitute (quota). When the Anthropic pool is too constrained to run a Workflow, use prun

with Codex-only units, or keep any Sonnet units small and targeted. This shifts the heavy fan-out to Codex while leaving only the coordinator and any chosen Sonnet work in the Anthropic pool.

• Complement (diversity). When a Workflow is affordable and you want cross-vendor perspectives,

run a Claude panel through the Workflow and a Codex panel through prun. Use the same structured contract and the same question on both sides, then cross-check. Agreement across vendors is usually a stronger signal than agreement inside one model family, because shared model lineage and tools can share blind spots. Invoke them together in one natural-language request; no special mode is needed. Reserve this for high-stakes work (a review, an audit, a hard design call), since it spends both pools and the coordinator must merge two result sets.

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

Rule: units never run on the coordinator (the Claude session itself). The orchestrator picks the executor per unit, with a strong default toward Codex:

• Codex is the default for almost every unit (code, research, analysis, web fetch). Its quota is

separate and abundant and its frontier model (gpt-5.6 tier) is capability-competitive with the top Claude models, so there is rarely a reason to prefer another worker. Start here.

• **Sonnet is the reserved exception, chosen only when a unit needs a tool the Claude session has but

the isolated Codex worker does not. Codex is an external process, so route to Sonnet when a unit needs a session-internal MCP / email connector (Gmail, Calendar, Drive, Slack), the Artifact tool, or a cross-vendor web-search verification where you want a Claude-side `WebSearch` result to cross-check the Codex one. A normal Sonnet subagent inherits the session's available tools but starts with fresh, isolated context** (it does not see the conversation history), so put any needed state in its unit prompt; if a task truly needs the full live conversation, keep it in the coordinator (an explicit fork inherits that context but also the coordinator's model, so it is not a Sonnet worker). The orchestrator decides per unit; when in doubt, use Codex. Sonnet draws Claude-side quota, so keep these units targeted.

• The Claude session stays the coordinator, never a unit. A single small session-tool task the coordinator can

do inline; reach for Sonnet when you need to run many such units in parallel.

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 (heavy Codex workers contend past roughly a handful at once, and the excess just queues) and Codex quota headroom. 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.

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 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.

No credential scrubbing or sandbox wall: 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. That is the whole safety model.

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).

1. Assign: default the unit to Codex; pick Sonnet only for the reserved cases (session-internal

MCP / email / Artifacts, or cross-vendor web verification). Also pick read-only (scratch) or code-writing (clone) mode. For a web-heavy unit, "Web access" below covers which executor fits.

1. Dispatch in parallel:

• Codex unit: run scripts/dispatch-task.{sh,ps1} in the background (Bash tool,

run_in_background=true). For a code-writing unit, pass the clone dir via PRUN_SCRATCH_CWD.

• Sonnet unit: spawn a background Agent subagent with model: sonnet. It inherits the session's

available tools, including MCP and connector tools; if you set a tools allowlist, include every connector, Artifact, file, shell, and web tool the unit needs. The subagent starts with fresh context, so put any needed state in its prompt. For code-writing it works in a clone too, under Claude's guard.py, which already gates commit/push.

1. 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 unit's own dispatch-task reaps a worker idle past PRUN_STALL_THRESHOLD at the same threshold, so a persistent stall surfaces as a FALLBACK to re-dispatch rather than a leaked zombie. (gather.{sh,ps1} remains for the plain wait-for-all case.)

1. 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 also salvages the tail into the result file automatically under a FALLBACK header), then re-dispatch or flag the user if it is unusable. 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 (Codex)

scripts/dispatch-task.sh --prompt-file <prompt> --result-file <abs result> --unit-id <id>

• Emits exactly one stdout line STATE-DIR <abs-path>; codex stdout+stderr land in <state-dir>/tail.
• If the worker exits without writing a non-empty result file, dispatch-task salvages its captured

<state-dir>/tail into the result file under a FALLBACK header, so a failed result-write never makes the unit silently vanish at gather. Treat a FALLBACK result as "review or re-dispatch."

• Self-heals a hung worker: if the tail stops growing for PRUN_STALL_THRESHOLD seconds (default

600; the same idle signal monitor reports) or the run exceeds CODEX_DISPATCH_TIMEOUT seconds (default 0 = hard cap off, so the idle signal stays primary and an actively streaming long run is not killed), dispatch-task kills the worker's whole process tree, exits 124, and writes the FALLBACK above naming idle-stall or hard-timeout. A non-empty result the worker already wrote is preserved, never clobbered. On Windows the watch+kill runs in the sibling reap-watch.ps1 (an AMSI-safe split of launch from watch+kill; the .sh does it inline).

• Runs codex from a per-unit working dir: a scratch dir by default (read-only units), or the path in

PRUN_SCRATCH_CWD (point this at a throwaway clone for code-writing units).

• Env: CODEX_DISPATCH_SANDBOX (default danger-full-access), CODEX_DISPATCH_REASONING (default

xhigh), CODEX_DISPATCH_ISOLATE_MCP=off to drop MCP isolation, PRUN_SCRATCH_CWD to set the cwd, PRUN_STALL_THRESHOLD (default 600) for the idle-reap threshold, CODEX_DISPATCH_TIMEOUT (default 0 = disabled) for an optional hard wall-clock cap.

Sonnet usage

Sonnet is the reserved executor (see Executors), for units needing session-internal tools (MCP / email connectors, the Artifact tool) or a cross-vendor web verification. Spawn an Agent-tool subagent with model: sonnet. It inherits the session's available tools but starts with fresh context (it does not see the conversation), so put any needed state in the unit prompt. Give it the same return contract and result-file path. For a code-writing unit, point it at a clone dir; commit and push are also gated by guard.py on the Claude side.

gather usage

scripts/gather.sh <result-file-1> <result-file-2> ...

• Prints GATHER-START count=N timeout=Ss, then DONE <abs-path> per file as it lands; exits 0 when

all land, exits 2 with TIMEOUT remaining=<k>.

• A file is "landed" when it exists, is non-empty, and has been quiet for the stable window

(default 10s); no startup-snapshot race.

• Use a fresh result path per unit per run (delete any stale file before dispatch). Have each unit

write its result in one operation.

monitor usage

scripts/monitor.sh <state-dir-1> <state-dir-2> ...

• Takes the STATE-DIR paths from each dispatch (not result files); reads each unit's tail (growth),

result-file (done/fail), and dispatch-pid (liveness).

• Prints MONITOR-START units=N stall-threshold=Ts timeout=Ss, then on the first actionable event

MONITOR-EVENT <all-done|stall|fail|timeout> and one UNIT <name> <status> line per unit (done / failed(fallback) / failed(dispatch-dead) / stalled(Ns) / growing).

• Exit: 0 all done, 3 attention needed (a stall or fail), 2 hard timeout.
• Env: PRUN_STALL_THRESHOLD (default 600, ten minutes; raise it for long code-writing units),

PRUN_MONITOR_POLL (default 15), PRUN_MONITOR_TIMEOUT (default 3600), PRUN_MONITOR_STABLE_WINDOW (default 10).

• Run it in the background; after handling a stall or fail, re-launch on the still-running units so a

resolved unit is not re-flagged.

report-state usage

scripts/report-state.sh   [--root DIR] [--json] [--summary] [--sort path|tail-bytes-desc]
                          [--min-tail-bytes N] [--include-legacy-pid]
scripts\report-state.ps1  (same flags)

Read-only. It inspects prun-task-* directories left behind by earlier runs and writes nothing at all, which tests/test_prun_report.py checks by hashing the tree before and after a run. Reach for it when a fan-out was interrupted and you need to know which unit output survived. --root repeats, and defaults to the system temp directory.

Every unit carries two independent fields instead of one verdict. A single label such as "salvageable" would read as permission to act, and this command cannot support that reading without the process identity it deliberately does not record.

result is unknown for every result_path_state other than resolved, and resolved may also carry it. Only FileNotFoundError proves a target is gone; a denial or an I/O error yields resolved/unknown plus an entry in that unit's errors, so a failed observation is never reported as an outcome. No other pairing can be emitted, and test_no_illegal_pair_can_be_emitted checks that against the table the module exports.

Remaining JSON fields:

Each errors entry is {"stage": <where>, "error": <value>}. The value is an exception class name, or one of two names for a condition that raises nothing: NotARegularFile when the path exists but is a directory, FIFO, or device, and EntryTooLarge when a result-file or dispatch-pid entry exceeds 64 KiB. That size limit reports rather than truncates. A truncated entry can strip down to a real path and be mistaken for a complete one. Consumers branch on stage:

Discovery failures sit apart from any unit, in a top-level discovery_errors array whose entries carry stage (root or unit-entry), the offending root or unit, and error. They are separate because a root that cannot be listed produces no unit to attach a failure to, and used to read as an empty corpus. Any entry in either place sets exit 1.

--summary adds two byte counters that never overlap. missing_or_empty_result covers units whose result path resolved to a file that is missing or empty. unresolved covers units whose result was never classified while their tail still holds bytes. Each counter names what was observed rather than what may be done about it, because neither a missing target nor an empty one proves that no other copy exists or that a live producer will not fill it. Both appear because the second group is easy to lose: across a live corpus of 220 units the first counter read 24.3 MiB while another 0.4 MiB sat in a unit nothing had classified.

Under --json, those counters arrive in a summary object:

--min-tail-bytes hides small units from the listing and moves no unit between classes; unit_count still counts them. --include-legacy-pid stays off by default. A recorded PID may be stale, or reused by an unrelated process, so it can never show that a worker is alive.

Exit codes: 0 every root was listed and every unit inspected cleanly, 1 at least one entry was recorded in a unit's errors or in discovery_errors while everything readable was still reported, 2 a usage error. An unreadable root is never reported as an empty one.

snapshot-tail usage

scripts/snapshot-tail.sh   --unit DIR [--dest DIR | --output FILE] [--json]
scripts\snapshot-tail.ps1  (same flags)

Copies one unit's tail into a ZIP holding exactly two members, tail.bin and manifest.json, both stored without compression. Only a regular file, or a symlink to one, may be snapshotted; a directory, FIFO, or device exits 4 and publishes nothing. Without that rule a device such as /dev/null reported zero bytes and published an empty archive as a complete capture, and a FIFO with no writer blocked the open indefinitely. The copy is byte-for-byte, so a tail carrying NUL or CR arrives unchanged. Given neither --dest nor --output, the archive lands in a per-user state directory: %LOCALAPPDATA%\anywhere-agents\prun\snapshots on Windows, and $XDG_STATE_HOME/anywhere-agents/prun/snapshots elsewhere, falling back to ~/.local/state when that variable is unset.

On POSIX the command creates the directory mode 0700 and the archive mode 0600. A snapshot extends the lifetime of prompts and tool output, so a directory that already exists and is group- or world-accessible is refused, with the chmod that fixes it named in the message.

Publication goes through os.link. That is the one portable operation which is both atomic and refuses to replace: os.replace overwrites, os.rename differs by platform, and checking first races. An existing destination therefore exits 3 and leaves the file byte-identical. Six concurrent attempts on one name produce exactly one winner. Any other link failure exits 6 rather than falling back to an operation that could overwrite.

The read is bounded by source_size_at_open, and it is best-effort. Equal counts do not establish that the source held still, because bytes can arrive from different generations of a growing file and still total the same number. Read complete_bounded_read as "the reader returned source_size_at_open bytes before EOF", never as "the source was unchanged" or "this is a consistent point-in-time copy". A truncate-and-regrow sequence can also total exactly that many bytes.

JSON output adds published, the final path, and warning, which is null on a clean run. A warning appears when the archive is linked into place but the temporary file could not be removed. The snapshot is valid in that case, so the command still exits 0.

Exit codes: 0 published, 3 the destination already existed, 4 the tail could not be opened or is not a regular file, 5 archive validation failed, 6 publication failed. Every failure other than 3 leaves no file at the final name.

The safety sentence

Snapshotting a tail is the only safe operation offered here. This output does not establish that deleting, overwriting, or promoting any unit is safe.

report-state prints those words on every run, in both text and JSON. snapshot-tail does not repeat them, so apply them yourself after a successful capture: holding a snapshot does not make the unit disposable. Deciding that a unit is finished needs process identity, which this slice records nowhere. See anywhere-agents#29 Part B.

Return contract (every unit writes this)

# <unit-id> result
Conclusion: <one line>
Files: <files created/modified in the clone, or "none (read-only)">
Open items: <blockers or follow-ups, or "none">
Verification: <what was run/checked/searched, or "none">
 
<body: the findings, survey, analysis, or change summary>

Ledger

Keep a simple run ledger (a file in a scratch area) recording each unit: id, executor, mode, prompt file, state-dir / clone-dir, result file, status (dispatched / done / failed), start/end. Use it to report progress and to relaunch only units whose result is missing or fails validation.

Where a unit's own files go: four kinds of file belong under an agent-io directory inside the scratch area. They are the per-unit prompt, the result file, the shared-context file every worker reads, and the run ledger. The directory name tells the writing-style hook to skip them, because none of that text is the coordinator's prose to rewrite. A unit prompt is an instruction to a worker, and a result file holds what the worker sent back. Anything the fan-out produces for a human reader stays outside agent-io.

Web access

Both executors reach the web by different paths, each with its own strengths, so assign per unit.

Codex runs on the user's local machine, so its requests leave from the user's local network rather than the cloud fetcher's egress IP, often a residential IP. That can reach some pages a cloud fetcher gets 403 on, though a hardened site can still block on bot score, fingerprint, or rate. It also surfaces pages a cloud fetch would miss. Web access comes from --sandbox danger-full-access (built-in browser path, confirmed under MCP isolation). Codex quota is abundant, so the extra unit is cheap.

Sonnet units get web from an agentType granting built-in WebSearch and WebFetch. Claude's WebSearch is strong at broad discovery (finding the right page when the URL is unknown), but discovery alone is not a session-internal capability, so treat Sonnet here as a reserved path for an explicitly wanted Claude-side cross-check or for recovery after Codex discovery falls short, not as the default for discovery.

Routing heuristic (apply the Executors rule; when in doubt, Codex):

• Fetch or discover on one path: use a Codex unit first, whether or not the URL is known. Its

local-network path also reaches some pages a cloud fetch gets 403 on.

• Codex discovery fell short, or acceptance needs a Claude-side result: add a targeted Sonnet

unit and its WebSearch.

• A high-stakes fact that might be stale or blocked: run Codex first, then add a Sonnet

cross-check when the value of a second vendor's view justifies the Claude-side quota.

A Codex web-fetch unit can use curl. Report the HTTP status per URL so a cloud-vs-local block shows up in the result. In Windows PowerShell, name the binary curl.exe, since a bare curl can resolve to the Invoke-WebRequest alias instead:

curl -sSL -A "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/124.0 Safari/537.36" -o <body-file> -w "%{http_code} %{url_effective}\n" <URL>

curl.exe -sSL -A "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/124.0 Safari/537.36" -o <body-file> -w "%{http_code} %{url_effective}\n" <URL>