Skill Usage

Skill Usage

Use this skill when a workflow needs a compact, deterministic record of a named skill invocation and its validation or failure handling.

Contract

Prereqs:

• Keep workflow judgment in the calling skill; use this CLI only for structured evidence capture.
• Keep observed-friction triage in HEURISTIC_SYSTEM.md and the calling

workflow skill; this CLI does not detect or classify every command failure.

• Choose an explicit output directory from project policy, or use a project-scoped agent-out project --topic skill-usage --mkdir

directory when the project has no evidence path.

• Required PATH usage: skill-usage available on PATH from nils-cli 0.8.5 or newer.
• Released boundary: Homebrew nils-cli 0.8.5 includes the expanded nils-agent-workflow-primitives binary surface with skill-usage.

Verify the actual binary with skill-usage --version.

• Local checkout fallback: Rust/Cargo plus a validated local nils-cli checkout that builds nils-agent-workflow-primitives, used only

when the PATH binary is absent or reports a version older than the release that includes skill-usage.

• Do not run multiple write commands against the same --out directory

concurrently. Serialize writes for one record directory and verify after the final update.

Inputs:

• init: required --out DIR, --skill TEXT, --intent TEXT, and --user-request-summary TEXT; optional `--trigger

user-explicit|agent-selected|project-policy|other, repeatable --referenced-file PATH, repeatable --external-source TEXT, --cwd DIR, --started-at TEXT, --validation-waiver TEXT, --force, and --format text|json`.

• link-record: required --out DIR, --type TEXT, and --path PATH; optional --format text|json.
• record-failure: required --out DIR, --phase preflight|execution|validation|cleanup|delivery, `--classification

skill-contract|script-bug|missing-dependency|external-service|project-state|user-scope|unknown, --symptom TEXT, --diagnosis TEXT, --handling TEXT, and --result fixed|worked-around|blocked|accepted-risk; optional --command TEXT, --exit-code CODE, repeatable --artifact PATH, and --format text|json`.

• record-validation: required --out DIR, --command TEXT, --status pass|fail|skipped, and --summary TEXT; optional `--artifact

PATH and --format text|json`.

• record-outcome: required --out DIR, --status pass|fail|blocked|worked-around|accepted-risk|skipped, and --summary TEXT;

optional --ended-at TEXT, repeatable --artifact PATH, repeatable --follow-up TEXT, and --format text|json.

• verify / show: required --out DIR; optional --format text|json.

Outputs:

• Writes skill-usage.record.json under --out DIR.
• JSON stdout uses versioned schema values such as cli.skill-usage.verify.v1.
• The record schema is skill-usage.record.v1.
• Recorded text is redacted for secret-like tokens before persistence.

Exit codes:

• 0: command succeeded; for verify, the skill usage record is complete.
• 1: runtime failure or incomplete/invalid skill usage evidence.
• 64: usage error.

Failure modes:

• skill-usage is unavailable on PATH and no validated local checkout invocation is being used.
• Evidence directory cannot be created or written.
• Concurrent writes target the same output directory and risk corrupting

skill-usage.record.json; serialize them or split the evidence into separate record directories.

• verify finds missing required fields, missing final validation without waiver, missing failure records for non-pass outcomes, or

secret-like values.

• Caller treats the record as policy judgment; record retention, promotion, and compression decisions remain in the workflow/runbook.
• Caller treats any non-zero exit as automatically durable; ordinary command

friction must be triaged before retention unless it is skill-bound and belongs in an already-required usage record.

Setup

Required released PATH boundary:

skill-usage --version

Use the PATH command only when it resolves to skill-usage 0.8.5 or newer.

Local checkout fallback boundary:

cargo run --locked --manifest-path /path/to/nils-cli/Cargo.toml \
  -p nils-agent-workflow-primitives --bin skill-usage -- --version

Run the Cargo form from the workflow's target directory. It is only a fallback transport for a validated local checkout when the released PATH binary is absent or older than 0.8.5. Do not mix PATH and local checkout evidence claims without stating which source was used.

Commands

Required released PATH command:

skill-usage init --out  --skill  --intent  --user-request-summary  [--format json]
skill-usage link-record --out  --type  --path  [--format json]
skill-usage record-failure --out  --phase preflight|execution|validation|cleanup|delivery --classification skill-contract|script-bug|missing-dependency|external-service|project-state|user-scope|unknown --symptom  --diagnosis  --handling  --result fixed|worked-around|blocked|accepted-risk [--format json]
skill-usage record-validation --out  --command  --status pass|fail|skipped --summary  [--format json]
skill-usage record-outcome --out  --status pass|fail|blocked|worked-around|accepted-risk|skipped --summary  [--format json]
skill-usage verify --out  [--format json]
skill-usage show --out  [--format json]
skill-usage completion 

Local checkout fallback command:

cargo run --locked --manifest-path /path/to/nils-cli/Cargo.toml \
  -p nils-agent-workflow-primitives --bin skill-usage --  ...

Workflow

1. Create or choose a retained evidence directory only when the workflow requires a skill usage record.
2. Initialize the record near the start of the skill workflow:

skill-usage init --out --skill --intent --user-request-summary --format json

1. Run write commands for that --out directory serially; do not parallelize

link-record, record-failure, record-validation, or record-outcome against the same record.

1. Link typed child evidence records instead of copying them into this envelope:

skill-usage link-record --out --type review-evidence --path

1. Record validation or an explicit validation waiver, then record the final outcome.
2. Before citing the record, run:

skill-usage verify --out --format json

Guardrails

• Do not hand-edit skill-usage.record.json or duplicate redaction, completeness, or JSON envelope logic in skill-local scripts.
• Do not use agent-side parallelism for writes to one record directory. If

workflow policy is not enough, route primitive-level locking or atomic updates to nils-cli instead of implementing a second writer here.

• Do not commit raw runtime records by default; commit only curated review, incident, audit, fixture, or compressed durable records.
• Do not use this record as a substitute for typed child evidence such as review-evidence.json, test-first-evidence.json, or

browser-session.json.

• Keep promotion/compression policy in docs/runbooks/skills/SKILL_USAGE_RECORDING_V1.md; the CLI owns deterministic writing and

verification only.

• Keep observed-friction promotion policy in HEURISTIC_SYSTEM.md and the

calling workflow. Use failures for skill-bound command, script, dependency, validation, delivery, or external-service failures only after a usage record is already required.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: graysurf
• Source: graysurf/agent-kit
• License: MIT

Install and usage instructions live in the source repository linked above.