Adapter Capability Probe

adapter-capability-probe

Lifecycle stage: VERIFY

Trigger

Use when Agent Brain is being run through an adapter, CLI runtime, subagent runtime, approval-gated runtime, skill runtime, or markdown-only harness and any claim depends on what the runtime can actually do.

Use before upgrading a run from read-only smoke to full validation, before accepting native /brain-* command support, or before trusting that blocked commands were reported honestly.

When not to use

Do not use for ordinary product verification where runtime capability is irrelevant. Do not use it to bypass /brain-verify; load it with runtime-smoke when capability evidence is part of the verification claim.

Inputs

• Selected command and command file.
• Adapter path, or unknown when no adapter is documented.
• Runtime name, version, Python executable, and exact invocation command.
• Sandbox/write mode and approval state.
• Claimed /brain-* command mode: native commands, markdown specs, mixed, or unknown.
• Candidate loaded skills.
• Probe results for read files, write files, shell commands, approval requests, network access, schema artifact emission, native command routing, and blocked-command reporting.
• Transcript path and redaction status when a runtime transcript exists.

Procedure

1. Inspect the selected command file and list the skills it requires.
2. Record whether /brain-* entries are native commands or markdown specs; treat them as markdown specs unless the runtime proves native support.
3. Probe one capability at a time with the least-privileged safe check: read access, temp-dir write access, shell access, approval flow, network policy, artifact emission, and blocked-command reporting.
4. Mark each capability as yes, no, blocked, or unknown; never infer support from a confident model response.
5. Record the exact evidence source for each capability: command output, runtime setting, adapter docs, transcript line, or unknown.
6. If the probe is read-only or blocked, keep run_scope as read_only_smoke and list the blockers instead of claiming full validation.
7. Feed the capability matrix into runtime-smoke and validate the resulting artifact against schemas/runtime-smoke.schema.json.

Anti-Rationalization

• Do not claim full validation when dependency install, shell execution, writes, or tests were blocked.
• Do not treat a runtime's natural-language statement as proof of native command support.
• Do not hide blocked commands; blocked-command reporting is itself a capability.
• Do not include secrets, private absolute paths, or unredacted sensitive transcript content in smoke artifacts.

Verification

• Confirm the selected command, loaded skills, adapter path, sandbox/write mode, brain command mode, blocked commands, run scope, and capability matrix appear in the runtime-smoke artifact.
• Validate JSON evidence with python scripts/runtime_smoke.py ... or python scripts/validate_repo.py when using checked-in examples.
• For repository changes, run rm -rf scripts/__pycache__ tests/__pycache__, python -m pytest -q, python scripts/validate_repo.py, git diff --check, and a targeted exact-name scrub before handoff.

Output Artifact

Runtime Smoke

Use templates/runtime-smoke.md and schemas/runtime-smoke.schema.json, plus QA evidence when /brain-verify owns the overall verification. Include blockers and next action so another agent can resume without inferring capability status from prose.

Failure Modes

Stop if the runtime cannot read the repository, the adapter path is undocumented, the command route is ambiguous, write-capable checks lack approval, transcript content cannot be redacted, or the runtime claims capabilities that cannot be backed by concrete evidence.

Example

Trigger: /brain-verify is checking whether Agent Brain works in a new approval-gated runtime. Action: open commands/brain-verify.md, list qa-evidence, runtime-smoke, adapter-capability-probe, ci-recovery, and agent-output-verifier as loaded skills, then probe read access, temp-dir write access, shell access, approval flow, command routing mode, and blocked-command reporting. Output artifact: templates/runtime-smoke.md with run_scope: read_only_smoke if shell or writes are blocked. Verification: validate the runtime-smoke JSON and include the exact command, runtime version, adapter path, sandbox/write mode, blocked commands, and capability evidence before handoff.