Fractal

Fractal

A fractal is a tree of autonomous loops, each running in an isolated Git worktree. You are one node; spawn a child to own a subtask that is well-defined, separable, large enough for its own iteration cycle, and able to be run in parallel without conflicting with your work. Pay careful attention to your NODE.md instructions to guide you on how and when to spawn sub-nodes.

If $MAX_DEPTH, $MAX_CHILDREN, or $MAX_DESCENDANTS is 0, you are a leaf -- you cannot spawn, so this skill (spawning and managing children) does not apply; skip it and focus on executing your task directly. Otherwise, default to spawning. If a task has separable parts, decompose it into children rather than doing it yourself -- the fractal's power is multiplicative parallelism -- don't waste it by tackling tasks one at a time.

Don't spawn what you can easily and reliably finish yourself. However, when a child's task is itself complex, be a manager: give it resources and detailed instructions on how to decompose further. Deep trees of focused nodes outperform shallow trees of overloaded ones.

Run fractal node --help and fractal node <command> --help for the CLI.

Limits (check before spawning)

• Width (--max-children). Caps direct children only (not grandchildren). Enforced locally on the spawning node. A child may set a larger --max-children than its parent.
• Depth (--max-depth) and descendants (--max-descendants). Enforced across the entire ancestor chain -- fractal node init checks every ancestor's config and rejects the spawn if any limit would be exceeded. You do not need to split or decrement budgets when spawning; set whatever limits make sense for the child's subtask and let init enforce the ancestors' caps. A spawn that would breach any ancestor's cap fails fast with an error naming the offending node, so glance at $MAX_DEPTH and $MAX_DESCENDANTS before spawning to avoid wasted failed inits.
• Re-entry re-checks. Width and descendant caps count children in play -- active, or idle awaiting start -- so a completed, stopped, exited, killed, or retired child frees its slot. node start --continue and a node unretire that lands idle put a node back in play and re-check both caps exactly like a spawn (refused over cap, no override); depth is structural, so only spawn checks it.

Cost

A --max-cost (per-run USD ceiling -- runs are isolated, and a budget-ended run refuses node start --continue without an explicit new --max-cost) is strongly recommended for every child -- without one the child launches uncapped, with a loud warning at start and bounded only by --max-iters/--timeout; a non-positive --max-cost is rejected. Set it at init (--max-cost, <= your remaining when $COST_BUDGET is finite) and optionally --max-iter-cost. Allocate conservatively; you may over-allocate across children optimistically, but monitor (fractal node cost spent, fractal node cost breakdown) and kill over-spenders. Caps are soft -- a child is not hard-stopped when it nears --max-cost; once it drains into the reserve it gets cleanup guidance to wind down the remainder of the current iteration cheaply, then the loop ends its run at that iteration's boundary (the child runs finish itself only when its requirements already verify -- the reserve guidance's goal-met exception) -- but a single iteration or its subtree can still overshoot, so reining in an over-spender is on you, the parent. Price a cap as solve + wind-down + reserve: a clean finish costs a full iteration of close overhead on top of the last unit of work (typically a few dollars per leaf child -- environment- and model-specific), so a cap sized to the solve alone strands a done child exited at the reserve boundary instead of completed unless it finishes deliberately in the wind-down. Size --max-iter-cost to one wind-down pass as well: the iteration cap, not the run-level cap, is what bounds each close attempt. Iteration caps bind at step boundaries, never mid-step -- plan each iteration to end under the cap, and read a trip as disclosure that an iteration ran hot, not as a hard stop. If a done child strands anyway, prefer raise-cap + continue + stop over a full re-finish -- the re-finish re-pays the whole close severalfold and can still near-miss -- and treat exited with recorded DB finish signals as a defensible terminal once the work is merged and verified. Child-side corollary: fractal node cost remaining reports against --max-cost, NOT against (--max-cost - reserve); a node inside the reserve band reads positive remaining and plans an iteration the loop will never grant. Plan your last productive iteration against (--max-cost - reserve), and if you intend to signal finish, fire it BEFORE draining into the reserve -- wind-down work scheduled for "next iteration" does not happen. In-step overshoot is bounded for claude children (each step runs under a hard per-invocation budget and stops cleanly at it), but the run-level cap stays soft. A child's spend (including its sync) counts against your own budget. Some agents report cost directly; others report token usage, priced from published rates -- so a cost cap on a token-reporting child requires a (priced) --model (the run fails on the first step otherwise). claude, grok, opencode, and omp report cost directly; codex reports tokens. A claude child routed through openrouter is priced from tokens too, so under a cap it also needs an explicit priced --model.

Every dollar figure you read has a scope; know it before comparing two. Budget-guard figures, recorded run-end reasons, and fractal node cost spent / cost remaining are all full-depth per-run subtree spend -- the run's own steps plus every descendant run chained under it -- so a spawning node's guard trips on money its children spent. fractal node activity's cost column is the row's own-node step cost only (run and iteration end rows sum just that node's steps), so activity totals read below the guard figures on a spawning node by design -- the gap is the children's spend, itemized by cost breakdown. And every figure is per-run: runs are isolated, and no cost command reports a lifetime rollup -- reconcile across runs by scoping each run (cost spent --run <id>) and summing.

[!WARNING] A small --max-cost on a child running an expensive --model is the combination most likely to blow the budget by a large percentage. The run-level cap is soft and only checked between steps, so one pricey step can be a big fraction of -- or exceed -- the child's whole budget before the next check. Give expensive-model children a budget large enough that a single step is a small slice, or hand a small budget to a cheaper --model. The sizing floor: never set --max-cost (or a remaining grant) within ~2x the model's single-turn cost -- a cap inside that band can be overshot by a large fraction in one turn, and that overshoot is documented, accepted behavior: no enforcement absorbs it.

For well-specified single-mission leaf work, a cheaper model at the same dollar cap (e.g. --model=sonnet) is a first-class choice -- cost-per-point favors it on numeric, single-task work -- but give it iter-cap headroom: micro caps bind on step granularity, and a fast drafter can pack a full iteration into one large step. Keep frontier models for manager, audit, long-horizon, and judgment-call roles: a cheaper model can hold process hygiene and mechanical output while its judgment quality collapses -- at no cost saving, since under a binding budget pool a cheaper token rate buys more steps, not a lower bill.

Spawn

1. Init: run fractal node init --help to see available options, then run fractal node init <name> --path="$PROJECT_DIR" [...]. --agent is optional (currently claude, codex, grok, opencode, or omp): when omitted, the child automatically inherits your agent -- pass it only to give a child a different agent. --provider is optional the same way (e.g. openrouter routes claude or codex through OpenRouter; it inherits, and agents without routes ignore it). --inherit=<surfaces> (comma-separated: steps, scripts, skills, config, or all) seeds those surfaces from YOUR node dir instead of the package seed -- steps/scripts/skills copy your live files, config snapshots preference keys (model, effort, sync, detached, iter/step timing, retries, pacing) and never budget caps; agent config always inherits. <name> uses letters, digits, and _ only (no -). All run parameters (budget, depth/children, iters, timing) are set here and stored in the child's config.json -- editable before launch. Capture the output for the child's project/node dirs. Init is lockfile-serialized -- run calls sequentially. A worked example of spawn hygiene -- every cap, the model, and the step timeout explicit at init (leaf caps shown -- give a manager child depth/children/descendants; fill every placeholder deliberately -- price the <usd> caps per Cost above, and when the tree's NODE.md names exact model recipes, they override this generic guidance), then a registry verify before any configuration; the init call is wrapped to capture the child worktree path from the init output's Initialized <path> line -- an unset $child_worktree silently reads YOUR OWN config (--path= resolves to the current directory; --path on config get names the child's worktree, not the project root):

   child_worktree="$(fractal node init sub_a --path="$PROJECT_DIR" \
       --agent=<agent> --model=<explicit_model> \
       --max-cost=<usd> --max-iter-cost=<usd> \
       --max-iters=<n> --step-timeout=<duration> \
       --max-depth=0 --max-children=0 --max-descendants=0 \
       | sed -n 's/^Initialized //p')"
   fractal node list                                # registry row: sub_a present
   fractal node config get max_cost --path="$child_worktree"   # <usd> -- caps landed
   fractal node config get model --path="$child_worktree"      # <explicit_model>
   

2. Configure: edit the child's NODE.md (instructions + completion requirements), steps/, setup.sh/test.sh/lint.sh, and skills/ -- invest here, it's the highest-leverage work (you can still steer after launch; see Configure below).

   For a well-specified single-mission leaf, consider the leaf step profile: before the config commit, trim the child's steps/ to PLAN + EXECUTE + COMMIT (delete the PREPARE and REVIEW files -- the loop runs whatever steps/ contains, in digit-prefix order). Loop overhead dominates single-mission leaf cost, and a trimmed profile cuts it -- expect the overhead cut, not an outcome guarantee. Two caveats: PREPARE is where a node merges its parent and children -- keep it for any node that spawns, and for leaves whose upstream moves mid-run -- when you cannot rule that out at spawn time, keep PREPARE; REVIEW is where memory folds -- when you trim it, relocate its duties into a surviving step (e.g. append the fold to EXECUTE's tail) or, minimally, tell the child in its NODE.md to fold memory before COMMIT. Managers keep the full profile. When your own steps/ already fits the child, skip the hand-trim and pass --inherit=steps at init.

3. Commit the config so the child starts from a committed baseline (a continue refuses over uncommitted project files -- commit them or pass --clean to discard them). Run the commit from the child's worktree -- a bare commit acts on the current directory's worktree (yours), not the child's (or pass --path=<child worktree>):

   cd <child worktree>  # .worktrees/<branch>
   fractal commit "configure <name>" --init
   

   Material you copy into a child's tree passes through the repo's commit hooks (if any) at this init commit (CRLF->LF, formatter reflow) -- so a byte-frozen store and its delivered copies diverge by construction on the first hooked commit; do not read that divergence as tamper. If you freeze reference bytes for audit, freeze the post-hook form (commit once, then re-pin) or compare normalized (line endings/whitespace) rather than raw.

4. Launch:

   fractal node start <branch>
   

   start takes no config arguments -- all run parameters come from config.json (set at init; adjust a value with fractal node config set <key>=<value>, read one with fractal node config get <key>, or edit the file directly). A configured max_cost must be positive; a missing max_cost launches uncapped with a loud warning. Add --continue only to continue a stopped/exited child. Starting is its own turn: when a spawn gate (child/descendant census, budget arithmetic) decides the launch, read it in one command and start in a separate one -- a chained start commits before you can see the read's output. The init gate re-checks census and budget at start, so treat a rejected start as the gate working; re-read before retrying.

Configure

Node configuration is the highest-leverage work you do -- a well-configured node runs autonomously for hours; a poorly configured one burns budget and creates entropy. Invest real time here, and commit the baseline before launch -- a continue refuses over uncommitted project files until they are committed or discarded with --clean. You can still steer a running child by editing its NODE.md, steps, or scripts (the loop re-reads them each iteration), but a strong baseline is fundamental.

For complex tasks, configure nodes as a manager: provide sufficient resources for them to spawn their own children, and write detailed NODE.md instructions to direct decomposition.

• NODE.md (<child_node_dir>/NODE.md): The child reads this fresh each step, so it is both the initial brief and the live steering channel. Write clear instructions, completion requirements, and add any relevant constraints. You can edit it after launch to redirect the child mid-run.
• Steps (steps/): Don't change your own steps, but when configuring a child you may add or replace step files (the loop re-discovers them each iteration) to fit the task -- e.g. adversarial plan/review/critic steps, dedicated research or test/debugging steps, or multi-pass execution, etc. If sync is enabled (the default), it runs automatically before each step to handle radio communication. The first and last steps (PREPARE and COMMIT in the stock set) are structurally important to the lifecycle (merging parent changes, committing work) -- do not remove or fundamentally alter them (one scoped exception: the leaf step profile in Spawn step 2 drops PREPARE for a childless leaf whose upstream won't move mid-run). Middle steps can be freely renamed, added, or replaced. A step file may begin with YAML frontmatter: agent: <command> runs it on a different agent (each agent keeps its own woven session across the steps it runs), provider: <route> overrides the provider route (agents without routes ignore it), model: <name> overrides the model, effort: <level> overrides the reasoning effort, timeout: <duration> overrides the node-global step_timeout for that step alone, detached: true isolates that step in its own session within a continuous node, and requires_approval: true holds the loop after the step completes until you approve it (fractal node pending/approve). Size step timeouts where the step is defined -- give a slow test or research step its own timeout: ceiling rather than inflating the global -- and retune the global mid-run with fractal node update --step-timeout (it lands at the next iteration top). Set a generous global step_timeout too, as a resilience floor: without one a hung or stalled invocation (a stuck API call, a wedged tool) blocks the step and the whole node indefinitely, where a ceiling lets the loop abort and retry. Pass --no-sync at init to disable sync for lightweight leaf nodes.
• Scripts (setup.sh/test.sh/lint.sh): setup.sh runs every iteration (keep it idempotent) -- add dependency installs, env setup, or data seeding here. lint.sh is invoked by fractal commit and test.sh during EXECUTE; extend them to match the child's scope (e.g. narrower test paths, additional linters). A parent's tuned scripts propagate with --inherit=scripts at init.
• Skills (skills/): add or customize skill files to give the child domain-specific capabilities beyond the defaults.

The seed lives in the node data directory (.fractal/), which is tracked by git and captured by your fractal commit --init; the same goes for the project wiki (wiki/). Neither belongs in .gitignore. Fractal ignores its own runtime artifacts (worktrees, the central database, status, agent logs) via the repo-local .git/info/exclude, which it writes automatically.

Meta nodes

A meta node configures another node's seed instead of doing project work directly. Use --meta=<target_branch> at init to create one -- this sets --base to the target's branch and --scope to its seed directory (.fractal/<target-branch>), so the meta node can only edit the target's seed files (NODE.md, steps, scripts, skills, etc.). $META_MODE is true when running as a meta node, and $META_TARGET is the target node's branch.

Use a meta node when a child's configuration is complex enough to warrant its own iteration cycle. The meta node studies the project and writes a high-quality seed; once done, merge it and launch the target.

Monitor and control

• Status: fractal node list, fractal node status <branch>. A node whose run ends on its --max-cost reports exited (exit 0 -- a designed landing, not a crash) -- a deliberate under-claim (a budget-aborted node must never read completed). The one exception: a node whose requirements verified and that ran a deliberate goal-met finish books completed even when the drain crossed the cap, with the overshoot recorded on the run row. So treat a capped node's exited as "inspect the work", not "failed": check its memory/plans and merge if the work is done. Verify a child's executable artifacts -- checkers, scripts, anything you will run -- from a clean checkout of its committed bytes (tmp=$(mktemp -d) && git archive <branch> | tar -xf - -C "$tmp"), never in-place in a working tree: artifacts that depend on uncommitted or path-relative state pass in-place and fail everywhere else. (Textual claims need only git show <branch>:<path>.)
• Stop: fractal node finish <branch> (after iteration), fractal node stop <branch> (after step; waits for the in-flight step to complete), fractal node kill <branch> (immediately). All three cascade over the target's entire subtree, children first.
• Pause: fractal node pause <branch> freezes the child's subtree in place (aborts in-flight agent turns; loops park with their runs open), fractal node resume <branch> relaunches it exactly there -- same budgets, same iteration count. A paused child still holds its spawn slot and blocks your finish-drain, so resume or kill it before finishing. Only resume, kill, and chat act on a paused node; start --continue is for stopped/exited nodes (fresh run, restored worktree -- uncommitted project files need --clean, and a budget-ended run needs an explicit --max-cost), never for paused ones.
• Clean up: fractal node merge <branch>. A conflicted merge restores your worktree exactly as it was and leaves the resolution to you: redo the squash by hand (git merge --squash <branch>), resolve and stage the conflicts, then finish with fractal node merge <branch> --continue rather than committing yourself -- the continue runs the rest of the merge (seed strip, wiki index refresh, commit, merge-base advance) that a hand-rolled finish would miss. It names every file where your resolution kept your content over the child's -- the child still carries its own version there, so land the resolution on the child (or retire it) or a later re-merge silently re-stages it. The continue's own failures leave your staged resolution in place to fix and re-run. Deleting after the merge (fractal node delete <branch>) is OPTIONAL hygiene, never automatic -- a merged node's branch and records keep audit value, so keep them unless clutter demands otherwise. Delete is destructive -- it force-removes the worktree and force-deletes the branch (and the whole subtree) regardless of merge state, discarding any unmerged work, so confirm the merge succeeded first.

Continue mode

When $CONTINUE_MODE is true, decide per child whether to propagate by assessing its memory/plans:

• continue (fractal node start <branch> --continue) if its work was in progress and still relevant -- uncommitted project files in its worktree refuse the launch, so commit them or add --clean to discard them;
• reset (fractal node init <name> --path="$PROJECT_DIR" --reset, then start) if the direction was wrong but the task stands -- --reset wipes the node to a stock empty node (memory, plans, steps, skills, config all cleared), so you must re-author its NODE.md, steps, and skills before starting;
• delete (fractal node delete <branch>) if the task is no longer needed; or
• leave (merge its work if completed and not yet merged).

Radio

Every node has a radio (auto-initialized) for live inter-node messaging -- channels public, private (owner-only), inbox (others write), outbox (you write); you auto-subscribe to the readable channels (public and outbox) of your parent and each direct child. If sync is enabled, radio is checked before every step, so messages and directives are picked up automatically. It is the live coordination path (the project wiki only syncs at merge). Messages have a priority (0-10, higher = more urgent). Run fractal radio --help (and fractal radio <command> --help) to explore the CLI. See the radio skill for messaging conventions.

Commands act on the current directory's node, so run them from your worktree -- you never pass a path for yourself. Radio verbs -- row-writing (send, post, reply, react, unsend, save, unsave, sub, unsub, channel create/delete) and listings alike -- act as the loop's exported node even from another directory, so a send and the sent listing that checks it always agree (read-your-writes); --path selects another mailbox. Name another node's branch positionally to act on it (e.g. fractal node status <branch>); --path is only for running from outside a worktree. fractal node init is the exception: <name> plus the project root via --path (e.g. $PROJECT_DIR).