add_smart_capability

add-smart-capability extends a smart tool that already exists: an agent reads the tool, implements one model-backed capability in its library, exposes it from the CLI, writes the tests and the docs, then runs the tool's own checks (uv run pytest, the conformance kit, and prek run --all-files when prek is installed) and fixes what they report. Nothing is committed; the working tree is left for you to review.

smart-tool-creator add-smart-capability \
  "Given an incident id, fetch its chat transcript and alert timeline from the incident platform and draft a blameless postmortem: summary, impact, contributing factors, and action items with owners" \
  --directory ~/src/incident-postmortem \
  --context "The platform client is src/incident_postmortem/platform.py; fetch through it, never call the API directly" \
  --context "Our postmortem template is at ~/notes/postmortem-template.md; match its headings"

from pathlib import Path

from smart_tool_creator.lib import add_smart_capability

added = add_smart_capability(
    "Given an incident id, fetch its chat transcript and alert timeline from the incident platform "
    "and draft a blameless postmortem: summary, impact, contributing factors, and action items with owners",
    directory=Path("~/src/incident-postmortem").expanduser(),
    context=[
        "The platform client is src/incident_postmortem/platform.py; fetch through it, never call the API directly",
        "Our postmortem template is at ~/notes/postmortem-template.md; match its headings",
    ],
)
added.report, added.checks, added.fix_rounds, added.output_message

Arguments

• REQUEST: the whole brief for one capability: what it does, for whom, and what it takes in and gives back. Required.
• --directory PATH: the smart tool to work in; the current directory when omitted. It must hold a smart-tool.json at its root, and no parent is searched, so a workspace holding several tools is never extended by accident. It must also have an Intelligence interface at src/<package>/intelligence/interface.py for the capability to run through, which a tool scaffolded with --no-agent-providers does not.
• --context TEXT: repeatable free text, usually paths to notes, transcripts, or exemplars the agent should read before it designs anything. It reads the paths itself, so name them rather than pasting their contents.
• --agent-provider: what the agent runs through, {{ agent_providers | join: " or " }}. The first installed, in that order, when omitted.
• --model: the model the agent runs on: a Copilot model id for copilot, <provider>/<model> for amplifier-agent (for instance {{ default_intelligence_models["amplifier-agent"] }}), a Codex model id for codex, a Claude model id for claude. Defaults to {{ default_intelligence_models.copilot }} on copilot, {{ default_intelligence_models["amplifier-agent"] }} on amplifier-agent, {{ default_intelligence_models.codex }} on codex, and {{ default_intelligence_models.claude }} on claude.
• --reasoning-effort: how hard the model thinks before it acts, one of low, medium, high, xhigh, max. Defaults to {{ default_intelligence_reasoning_effort }}. Applies to the copilot, codex, and claude agent providers.
• intelligence, library only: the Intelligence implementation the agent runs through, which wins over agent_provider; resolve_intelligence(agent_provider) when omitted. Tests inject a fake.

Result

AddedCapability carries root, the extended tool's distribution root; report, the agent's final message, which names the capability, the files it touched, the command to try it, and its caveats; checks, one entry per check with its name, command, status (passed, failed, or skipped), and the tail of its output when it failed or the reason when it was skipped; fix_rounds, the extra agent runs spent on failing checks, at most 2; and output_message, all of that for the calling agent, which is what the CLI prints.

The prek check is skipped, never failed, when the tool has no .pre-commit-config.yaml or prek is not on PATH; the conformance check is skipped when the kit itself could not run, and names the failing rules when it fails. A check still failing when the work stops is returned rather than raised and named in the message, and the CLI exits 1; the work stays in the tool's working tree either way, and what it is worth is the caller's call.