Index of every page on this site: llms.txt
All skills
scode avatar

/scode-agent-delegation

@03a9f7f
Run it yourself

How to work with a delegate once a caller has decided to use one: the self-contained task spec, the artifact-file rule, the no-VCS rule, the run id and run directory, the two-stop checkpoint protocol for delegates that write code, crash and resume handling, and the gate that judges the result and returns one of a fixed set of verdicts. Loaded by orchestration skills at the moment they delegate; inert on its own, never active for a session, and not meant to be invoked by the user directly.

  • 8 files
  • 72.8 KB
  • Updated 2 days ago
  • GitHub
Use this Skill: https://skilld.dev/gh/scode/dotfiles/scode-agent-delegation

Nothing lands on disk. Nothing to clean up.

Fork this Skill

Edit a local copy. It keeps the author and licence.

SPEC.md

≈2k tokens on demand. Your agent reads this file only when SKILL.md points to it.

scode-agent-delegation specification

Dependencies: scode-harness-shellout

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.

Requirements

  • The skill's dependency set is the Dependencies: line above, and the dependency graph between skills is one-way: this skill may refer to scode-harness-shellout, and never to a skill that consumes it. No text here points a reader upward into a consumer ("escalate per the caller's rules", a named orchestration skill or one of its files); where the text needs a fact only a caller has — the tree, whether isolation is allowed, the recorded base, the model and mechanism, what to do with a verdict — it names that fact as an input the caller supplies and stops there.
  • The dependency uses the current harness's authorized skill loader or resource resolver, or, when no dedicated loader exists, the exact SKILL.md location 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, using the resolver for that same skill, its reported base, or the directory of its supplied SKILL.md path. A filesystem base is unnecessary when a resolver addresses required resources or no sidecars are needed.
  • Truncated or elided output is incomplete delivery, not a terminal failure by itself. Recover omitted content through the tool's continuation, range reads, or full-output artifact tied to the same resource or result before acting. Unknown skills, denied access, invalid identity, unresolved required resources, or content that cannot be fully retrieved stop the affected operation and name the skill and 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 in tests/skill_deps.rs. On omp, read resolves skill://<name> and skill://<name>/<relative-path>; these satisfy the same capability-based contract as other authorized loaders. The dependency is loaded only for foreign-harness delegation, never for native delegation.
  • A same-named project-local skill may shadow the installed dependency when the harness selects it. This is accepted: the name check proves identity, not revision. Codex's fixed-path fallback does not discover project-local overrides.
  • Loading this skill is side-effect free. Invoking it activates nothing for the session, writes nothing, and claims nothing about later spawns; its SKILL.md says so in its first paragraph and its description says it is loaded by other skills and inert alone. No harness-level switch turns off description-based selection (Codex's allow_implicit_invocation: false makes Muse Code refuse to load the skill for the model at all, verified on Muse Code 1.0.2); the guarantee is inertness.
  • SKILL.md is the public surface and is kept lean, measured as what a session loads before its first delegation: the task-spec rules, the three-step start, when the dependency is loaded, the resumability rule, the crash classification in summary, and the verdict vocabulary stay in SKILL.md; the checkpoint protocol, the gate procedure, and isolated integration live in sidecars read on demand, each with its trigger named in SKILL.md.
  • Multiple concurrent orchestrators must not conflict through anything this skill puts on disk. This skill generates the run id as the first step of every delegation, creates and owns <tree>/.agent-delegation/<run-id>/, names every artifact it prescribes by that id, never removes or reinterprets a run directory it did not create, and moves its own run directory to the session's private scratch space only once the caller reports having acted on the verdict.
  • The gate returns exactly one of the verdicts listed in SKILL.md — accepted, accepted with local fixes, spec defect, substantive failure, fixable, substantive failure, structural, execution-path failure, misclassified, inconclusive, unresumable, blocked on user — with the payload the table names, and never a verdict outside that list. What the caller does with a verdict is the caller's; this skill never escalates, reroutes, relaunches, or removes changes from the tree.
  • Task specifications preserve the caller-supplied user intent, later decisions, binding repository constraints, and implementation outline when available, separately from proposed mechanisms. The gate checks unnecessary obligations as well as omitted behavior, at assumptions review, implementation review, and before further repair after repeated corrective reviews. It considers the aggregate design without demanding a periodic review loop or treating novelty as a defect. spec defect covers both directions, including a defect found before implementation; its payload names the defect and the smallest correction preserving explicit requirements and repository constraints. A simplification that instead needs a user decision returns blocked on user; the gate does not authorize scope changes.
  • REVIEW.md may end with numbered Also: items: changes the caller requires that no DECISIONS.md entry covers, typically defects found by reading the diff at the AWAITING REVIEW stop. They bind the delegate like the per-entry changes, and they are optional: a caller that writes none gets the protocol exactly as it was before they existed. This adds no stop and no file.
  • The existing AWAITING REVIEW checkpoint permits an explicitly incomplete scope exception when a substantial mechanism outside caller-supplied boundaries is needed. The delegate leaves that portion unimplemented, finishes independent in-scope work, and records the dependency, alternatives, unfinished behavior, and actual checks in DECISIONS.md. The caller gates the dependency before authorizing more work. This adds no stop type, does not claim completion or passing checks, and does not excuse ordinary implementation defects or test failures.
  • The skill is meant to work on modern Linux and macOS. Commands, paths, and tools it prescribes must be available on both; nothing may rely on one without an equivalent for the other. No other platform is of concern, and the skill's text need not accommodate one.
  • Native delegation uses available automatic completion delivery or a blocking completion call before falling back to bounded waits. Waits respect tool limits and the caller's responsiveness, resource checks, and deadlines; they avoid repeated empty status turns and batch checks across outstanding delegates. Prefer monitoring without model turns; a small-context watcher is allowed when it can notify the caller without being polled and is expected to reduce total cost, including launch overhead and polling usage. Notification support is not assumed across harnesses. Wait expiry alone is not delegation failure.

Source: SKILL.md on GitHub

No rule matched.

skilld matched fixed text patterns in SKILL.md and file names. Patterns miss obfuscated code.

skilld run checks every file with the same patterns. It asks for approval before it loads a Skill with a behavior marked Needs approval.

No third-party reports yet.

Signed by skilld at 03a9f7f. This ties the file your Agent reads to that commit on GitHub. It does not review the instructions.

Last checked against GitHub yesterday.

Activeupdated 2 days ago

README badge

README badge for scode/dotfiles/scode-agent-delegation