scode-powerhorse specification
Dependencies: scode-agent-delegation
NOTE: This file is binding on the skill's text. It is deliberately sparse: it records only requirements that have been stated as such, not a description of everything the skill does. Absence of an entry means the behavior is not yet specified, not that it is unspecified on purpose. When the skill and this file disagree, that is a bug in one of them; fix the skill or change this file in the same change, never leave them apart.
Why this skill exists
scode-galaxy-brain is built to save cost: it routes suitable work to cheaper models and escalates when they fail. This skill serves a different objective, finishing a goal sooner at equal quality by running large chunks of work in parallel on strong delegates. Expressing that as overrides on galaxy-brain was tried on paper and rejected: its delegate-or-not test, its routing, its escalation ladder, and its design defaults each pulled against it. This skill is a separate orchestrator so that galaxy-brain's text, behavior, and evals stay untouched.
Requirements
- The session in which the skill is invoked is the supervisor. It owns the decomposition into chunks, the interfaces between chunks, the user's intent, every gate, integration, and all version control. Delegates never commit, branch, push, or open PRs.
- The objective is time to a finished, gated goal at equal quality. Cost is a constraint, never the reason to delegate or not to delegate.
- This skill and scode-galaxy-brain are never active in the same session. While this skill is active, the session does not invoke, load, or follow galaxy-brain or scode-model-routing. Invoking this skill while galaxy-brain is active, directly or through a goal file the user had written, stops galaxy-brain; invoking galaxy-brain later stops this skill. The skill's text names galaxy-brain only to state this exclusion and the difference in objective.
- There is no routing. Chunks of the supervisor's own decomposition run at the worker selection: the session's own model and effort unless the user or a goal file names another (at the named effort, or the session's own effort when none is named), resolved once per run to a concrete model and effort, recorded, and set explicitly at launch. A later session keeps the recorded selection; when it cannot reach it, the user is asked. The supervisor never substitutes another model on its own. Spawns another process defines (a per-PR review, a swarm, a scope review) run at exactly the model, effort, agent type, and skill that process or the user demands, and at the session's own model where it names none. The supervisor delegates nothing else: no helper agents for searches, scans, or monitoring. A chunk delegate may use read-only helpers of its own (the user decided this explicitly), which count against harness capacity; it may not start helpers that edit files.
- The supervisor chooses each launch mechanism itself: native only when it can select the model and effort and, for a
chunk, resume the same agent after a checkpoint; otherwise
external, which leaves the CLI to the shell-out skill, and a resume goes through the CLI recorded at launch. Availability restrictions in the user's~/.scode-model-routing.md, when it exists, are honored, and what it says about the provider or harness for a model is passed on as the launch preference. A demanded model nothing can reach goes to the user, never to a substitute. - A chunk is one coherent outcome with its tests and documentation, roughly a reviewable PR or a subsystem, finishable without another in-flight chunk's unfinished changes. Work smaller than a chunk is never delegated, and chunks are never split only to widen the fan-out. A chunk is delegated when it can overlap other in-flight work or would fill a large share of the supervisor's context; otherwise the supervisor does it itself.
- The supervisor settles intent, shared interfaces and invariants, compatibility, and integration order; the delegate owns the chunk's internal design within those boundaries and proposes it at the first checkpoint.
- Writer width defaults to two chunks in flight (running or stopped at a checkpoint), overridable by the user or a goal file. Launches pause on resource alerts, rate-limit or quota errors, more than one delegate waiting on the supervisor, or ungated results.
- Every chunk runs in a tree the supervisor creates, named by the run id, even when it is the only chunk in flight, with shared out-of-tree state isolated or serialized. While a delegate owns a tree, the supervisor makes no edits and runs no version-control commands in it beyond read-only inspection. Accepted results are integrated serially with checks after each apply. Per-PR reviews the user's process requires run while other chunks continue.
- Every chunk delegation, and every process-defined spawn that does not run natively, goes through
scode-agent-delegation, which generates the run id, owns the run directory, launches, runs the checkpoint protocol, and returns one verdict from its vocabulary. This skill's text carries no copy of the task-spec checklist, the checkpoint protocol, the gate's steps, or the addendum. It acts on every verdict in that vocabulary through its action table for chunks; a verdict with no row is a bug here. A process-defined spawn whose verdict is notacceptedis relaunched once at the same demanded model and effort, then reported. - The deep review of a writer chunk happens at the
AWAITING REVIEWcheckpoint, with findings sent back throughREVIEW.md(includingAlso:items) so the delegate fixes them in its own context. Findings the gate maps tospec defect,blocked on user, orsubstantive failure, structuralreturn that verdict instead. The final gate confirms against the complete change set, untracked file contents included, saved at that stop. - A fixable failure gets one fixup at the worker selection before takeover or replanning. An
unresumablechunk that the chunk table records at the second checkpoint is taken over, not discarded: the supervisor re-runs the checks, applies outstanding answers and review items, gates the current change set, and records a recovery note in place of the missing report. A rate-limit or quota ending is an environmental cause, never a decline, and does not count toward the delegation skill's crash-resume cap; this departure from that skill's classification is stated inSKILL.md. - The supervisor keeps a chunk table in a file of its own, edited in place at every phase change (an append-only resume log records its path and the phase changes, never the table rewritten), and reconciles it against reality before launching anything after compaction or resume.
- Multiple concurrent uses of this skill, or of this skill and galaxy-brain, by different sessions must not conflict through any state the skill itself maintains. Every artifact it creates for a delegation (trees, run directories, prompts, logs, saved change sets) is named by a run id the creating session generated, and no session removes or reinterprets another's. The chunk table is run-wide state at the path the goal file or the user is given, shared only by the sessions that resume the same run one after another, never concurrently.
- The skill's dependency set is the
Dependencies:line above, and the dependency graph between skills is one-way: this skill may refer to a skill it depends on, and a skill it depends on never refers back to it. The shell-out skill is not a dependency of this skill: the delegation skill loads it. - Dependencies use the current harness's authorized skill loader or resource resolver, or, when no dedicated loader
exists, the exact
SKILL.mdlocation its catalog or instructions supplies. On Codex, when no location is supplied, use${CODEX_HOME:-$HOME/.codex}/skills/<name>/SKILL.md(the root verified for Codex 0.152). Known interfaces are examples, not an allowlist of harnesses. Missing filesystem metadata or an unfamiliar harness alone must not block loading or trigger a permission question; actual tool permissions remain binding. Never guess installation paths or URI schemes, assume a sibling dependency directory, or search other skill roots. - Identify the exact requested skill from its returned frontmatter, or the loader's reported identity when frontmatter is not exposed. Missing or conflicting identity fails loading. Read the skill in full and all sidecars the current step needs. Truncated output is recovered through the tool's continuation before acting; content that cannot be fully retrieved stops the affected operation and names the skill and the failing path, URI, or tool. No permission bypass, remembered instructions, alternate copy, or similar skill may replace a failed load.
- The loading text is the marked stanza in
SKILL.md, checked against the canonical template intests/skill_deps.rs. - Invoking the skill activates it for the rest of the session, per "Staying active for the whole session" in
SKILL.md; the dependency, by contrast, activates nothing when loaded. - The skill is meant to work on modern Linux and macOS. Commands, paths, and tools it prescribes must be available on both.