Subagent reference

Reference: askit subagents

The seven askit subagents are bounded, read-only or narrowly scoped delegates that parent skills dispatch for specific phases of work. None owns its own conversation; each is invoked by a parent skill via the agent chain contract. Chain permissions are governed by agents/_chain-permitted.yaml - a subagent not listed there for a given parent may not be dispatched by that parent.

Four of the seven have a declared parent; three do not. askit-evaluator, askit-quality-grader, askit-reviewer, and askit-skill-author each have a declared edge. The three general-purpose delegates (askit-explorer, askit-file-ops, askit-file-search) appear nowhere in _chain-permitted.yaml, because only two skills declare a chain: at all today (askit-build-skill and askit-evaluate) and neither names them. S4 (chain-contract) validates the edges that are declared rather than requiring every subagent to have one, so this passes the gate - but it does mean no skill may currently dispatch those three under the contract. They were authored ahead of the skills that will declare them. This page states that gap rather than implying a permission the contract does not grant.

askit-evaluator

Purpose. The delegated assessment role behind askit-evaluate. Runs npx agent-skills-toolkit evaluate <target> --json and reports findings grouped by severity and requirement ID, each with its file path and the remediation the message states. Read-only: it never edits the target.

Parent skill. askit-evaluate (primary); also dispatched by askit-skill-author for post-authoring conformance checks.

Chain permissions (from agents/_chain-permitted.yaml).

Permitted callers: askit-evaluate, askit-skill-author

askit-explorer

Purpose. A bounded, read-only discovery delegate. Surveys a repository broadly and returns a structural map of its components and layout: what component types are present, where they live, what manifests exist, and what conventions the repo follows. It reads excerpts to locate and classify, not whole files, and never edits. The broad counterpart to askit-file-search (which resolves a single known query).

Parent skill. None declared. Its natural caller is a discovery-heavy skill such as askit-migrate (assess mode), where a foreign repo must be surveyed before it can be graded, but askit-migrate declares no chain: today.

Chain permissions (from agents/_chain-permitted.yaml).

No entry. No skill declares an edge to it, so no skill may dispatch it under the chain contract.

askit-file-ops

Purpose. A bounded file-mutation delegate. Carries out a specified set of create and edit operations precisely: the “do the writes” role an authoring skill delegates to once it has decided what to change. It applies exactly the operations given; it does not decide scope, design content, or run commands. Read-before-write is mandatory.

Parent skill. None declared. Its natural callers are authoring skills that have already decided what to write, such as askit-build-skill (improve mode) or askit-build-docs, but neither declares an edge to it: askit-build-skill chains only askit-skill-author, and askit-build-docs declares no chain: at all.

Chain permissions (from agents/_chain-permitted.yaml).

No entry. No skill declares an edge to it, so no skill may dispatch it under the chain contract.

askit-file-search

Purpose. A bounded, read-only search delegate. Answers a precise locate question: which files match a pattern or where a symbol or string lives. Returns paths with the matched lines. The pinpoint counterpart to askit-explorer: explorer maps breadth (what is here), file-search resolves a known query (where is X). Never edits.

Parent skill. None declared. Its natural caller is any skill that needs to locate a specific file, symbol, or text pattern before acting, but no skill declares an edge to it.

Chain permissions (from agents/_chain-permitted.yaml).

No entry. No skill declares an edge to it, so no skill may dispatch it under the chain contract.

askit-quality-grader

Purpose. The behavioral-judge delegate behind askit-evaluate’s behavioral mode. Runs a skill against its eval-set (evals/ triggering and behavior cases) and judges, case by case, whether the skill fires when it should, stays silent when it should not, and produces the expected behavior. Reports a per-case verdict with evidence. This is the LLM-judged layer that sits beside the deterministic gate and cannot change the CI pass/fail result (Design Principle 3). Distinct from askit-evaluator (deterministic conformance) and askit-reviewer (qualitative artifact review).

Parent skill. askit-evaluate (behavioral mode only).

Chain permissions (from agents/_chain-permitted.yaml).

Permitted callers: askit-evaluate

askit-reviewer

Purpose. A bounded, read-only review delegate. Forms judgments a deterministic check cannot: is the change correct, does it honor the Standard’s intent (not just its lettered rules), is the description at the right altitude, is the component warranted rather than a would-be mode of an existing one? Reports findings with severity and a concrete remediation per finding. The qualitative complement to askit-evaluator. Never edits.

Parent skill. askit-evaluate (review mode) and, since v1.7.0, askit-build-skill (improve mode phase 2, the craft pass - see ADR 0037).

Chain permissions (from agents/_chain-permitted.yaml).

Permitted callers: askit-evaluate, askit-build-skill

Under askit-build-skill its findings are partitioned by scripts/lib/craft-review.mjs into a mechanically SAFE subset (a closed allowlist, applied only on explicit consent) and a JUDGMENT remainder (reported, never touched). Its output is advisory in both parents: it renders beside the gate verdict and cannot move it.

askit-skill-author

Purpose. The delegated authoring role behind askit-build-skill. Scaffolds skills/<name>/ from the skill template, writes a conformant SKILL.md (frontmatter and body), emits native manifests for the declared agent targets (gen-manifest --write --target=all), and iterates to zero gate errors. Delegates conformance assessment to askit-evaluator. The one subagent in this set that has both read and write access plus Bash (needed for scaffolding and manifest generation).

Parent skill. askit-build-skill.

Chain permissions (from agents/_chain-permitted.yaml).

Permitted callers: askit-build-skill
May in turn chain to: askit-evaluator (for post-authoring conformance checks)