CLI Exercise

Exercise the caller's selected executable and return local evidence. Accept precise cases, a broad goal, or a mixture. Preserve explicit requirements while letting the agent choose useful investigations where the caller left room.

By default, produce one chaptered session video, even for one case. Each case covers one behavior and has a very short natural-language behavior/outcome title. Actually execute and record that case's preparation, one exercise invocation, then validation before starting the next case. Approved cleanup can follow validation. Never batch all setups or validations or reorder historical clips to imply that execution order. Exact supplied steps and narrower coverage remain authoritative: ask about conflicts rather than silently adding, splitting, or reordering a caller's case.

Deliver media only by default. Infer clear output intent: a video exercising a PR calls for an MP4, while a request for an interactive review report calls for media plus HTML. Ask only when the requested deliverable is ambiguous. Generate HTML with --html only when wanted; do not create or open it for video-only requests. Keep raw captures and JSON results as internal evidence.

The internal SKILL.md files below are bundled instruction modules. Read the needed module from this package and follow it in the current agent context. They do not require new agents, sessions, or separately installed sibling skills.

Load only what is needed

This entry file is the routing overview. Load the selected flow, not every internal module. Execute the Go helper as a tool; read its source only when diagnosing a problem or reviewing the implementation.

Go owns prerequisite checks, execution policy, controllers, capture, verification, and rendering. Node is only a mechanical adapter to Tuistory's PTY API. The skill does not require another scripting-language environment.

These direct links keep resources one hop from the entry point. For an exact request, do not load the exploration flow unless the caller also supplied an exploratory case. Do not load controller details for a live-only run.

1. Preserve the request and choose the fidelity policy

Classify each case, not just the entire request:

• Exact: the caller supplied cases or required steps. Preserve commands, arguments, input values, order, assertions, and timing/output constraints. Do not add exploratory steps or silently substitute an equivalent-looking operation for a specified key sequence.
• Explore: the caller supplied a goal or expected outcome. Choose checks and interactions within the agreed context and permitted effects. Separate supplied expectations from your hypotheses.

Input fidelity is independent of execution strategy. Either policy can use a bounded controller, live observation/action calls, or both. Switching strategy never grants permission to change an exact case.

Resolve only missing details that affect correctness or authority:

• The executable or build to exercise, with its actual path and identity.
• Cases or goal, supplied context, expected outcomes, and initial state.
• Permitted files, hosts, accounts, repositories, and other effects.
• Desired media format and constraints. Honor an explicit format; use GIF for an unspecified terminal animation and MP4 when the caller requests video.

Default to condensed presentation, with static pauses shortened. Honor explicit real-time requests and retain the original capture timestamps. Keep timing adjustments in the recording metadata, not repeated on-screen labels.

A request such as "upload an attachment" is not permission to choose an arbitrary account or publish to the current repository. Ask for the missing target/authorization before performing a write. Prefer synthetic data and isolated configuration.

Command discovery is a separate case property. A case may provide the command and test its use, or explicitly require finding it. Do not silently turn a known-command case into a discoverability test.

Preserve the non-secret original request alongside normalized cases. Never copy credentials into the contract or recording; use approved secret references or report blocked if the requested literal recording would disclose them. Use the run contract; unsupported or ambiguous requirements must be clarified or marked blocked, not discarded.

2. Establish readiness and authorization

Read the preflight module. It checks dependencies. If setup is needed, propose commands using the pinned requirements, obtain caller approval, and recheck afterward.

Installation approval and permission to modify test resources are independent. Record the executable path, version/revision, checksum, accepted case contract, and approved effect scope before starting the subject.

Use a secure private workspace outside the repository. Do not put recordings, credentials, runtime caches, or operator configuration in this skill directory. No secret values belong in the contract or a tool request.

3. Load the appropriate internal flow

• Supplied cases: record exact cases.
• Broad goals: explore a goal.

Both use the shared run module. It provides real terminal observation, input, controller execution, recording, and bounded process ownership. Read only the internal modules needed. Use a recording session to keep case and phase order, carry approved state between related commands, and assemble their actual recordings with evidence --session, not the per-invocation --run-dir path. Keep validation commands in that recording. The session overview uses the task-specific title, derived fidelity, and case results, with 10 seconds per readable page. Progress spans the entire final video.

4. Verify, render, and clean up

Always finish through the evidence module, including when the case is blocked or the subject fails.

Return the case outcome, meaningful observed behavior, requested media paths, executable identity, inspection coverage, and limitations. Keep case failure, capture failure, and render failure distinct. A video is evidence of an observed flow, not proof of general correctness or a replacement for tests.

Include a report link only when the caller requested the report. Do not leave a host, terminal, recorder, or playback server running. Publishing the evidence is a separate, explicit action.