| name | mister-smith-runtime-eval |
| description | Use when the user wants a complete Mister Smith live runtime session or workflow evaluation with real execution, durable evidence capture, and honest task/session/autonomy proof instead of mock-only validation. |
Mister Smith Runtime Eval
Use this skill when the ask is to prove what the current Mister Smith runtime actually does on a
live run.
This is a runtime-evaluation skill, not an implementation skill and not a planning-only skill.
When To Use
Trigger this skill when the user wants any of the following:
- a real runtime-backed session or task evaluation
- proof that task, session, and autonomy surfaces agree
- a fresh rerun of an earlier proof note on current
main
- artifact capture for runtime logs, payloads, and result surfaces
- confirmation of success, collapse, or failure-visible behavior on the supported live path
- an honest comparison between deterministic checks and real runtime behavior
Do Not Use
Do not use this skill when:
- the user only wants unit or integration tests
- the task is to write or revise a SpecKit packet
- the task is to implement code changes
- the task is only about Linear, Symphony, or execution-lane workflow state
- mock-only coverage is enough for the request
Required Inputs
- the evaluation target:
- one-shot task path
- session path
- or both
- the proof scope:
- one targeted outcome
- or the full three-case proof matrix
Default runtime assumption unless the user explicitly overrides it:
- provider:
openai_chatgpt
- model:
gpt-5.4
For direct runtime-proof asks, prefer the product/runtime path first. Do not require Linear or
Symphony unless the user explicitly asks for development-workflow tracking alongside the runtime
evaluation.
Start Sequence
Read, in order:
AGENTS.md
docs/current-state.md
docs/plans/2026-03-15-first-live-multi-agent-runtime-proof.md when the ask is about proving
real end-to-end execution
- the active checkpoint or current evaluation note relevant to the ask
- any packet or closure note directly named by the user
If the user is asking for a rerun of a known proof lane, also read the closest existing note under
docs/plans/ before starting the live run.
Use docs/current-state.md to confirm the currently supported live surfaces. Today the supported
autonomy HTTP route is:
GET /api/v1/autonomy/status/{workflow_id}
Do not rely on older session-turn autonomy routes unless current repo truth explicitly restores
them.
Workflow
1. Freeze the evaluation scope
Decide whether the run is:
- a targeted rerun for one known gap
- a bounded two-case evaluation
- or the full three-case proof matrix
Freeze the proof-outcome taxonomy exactly:
graph_formed_and_completed
collapsed_to_sequential
failed_before_graph
Do not invent new proof classes during the evaluation session.
2. Prepare the durable note and artifact lane
Before the live run, choose the note and artifact destination:
docs/plans/YYYY-MM-DD-<slug>.md
docs/plans/artifacts/YYYY-MM-DD-<slug>/
Use the reference playbook at
references/runtime-evaluation-playbook.md for the standard artifact set and note structure.
3. Run deterministic preflight
Run the narrowest honest deterministic bundle for the requested scope before starting the live
runtime. Typical examples:
cargo build --workspace
cargo test -p mister-smith-app
- packet-specific crate tests when the request is tied to a known packet lane
git diff --check
Keep deterministic validation separate from live-proof claims.
4. Launch the isolated live runtime
Start a local runtime with isolated state:
- temporary database
- explicit HTTP port
- runtime log capture
- explicit provider and model
Verify readiness before sending traffic.
Use the actual app path, for example:
target/debug/mister-smith run
- or
cargo run -q -p mister-smith-app -- run
Choose the path that matches current repo practice and the user request.
5. Execute the live proof
For task-based proof:
- submit a real request to
POST /api/v1/tasks
- capture
workflow_id
- inspect task status
- inspect autonomy status by
workflow_id
- cross-check with CLI autonomy status when useful
For session-based proof:
- create or reuse a real session
- send the live turns required by the evaluation
- capture
session_id, turn count, retained assistant result, and the linked workflow ids
Always collect:
- request payloads
- task/session/autonomy responses
- runtime log excerpts
- the runtime execution mode fields
- the exact provider/model used
6. Evaluate honestly
State what happened on the supported live path, not what the packet or docs hoped would happen.
Required distinctions:
- deterministic validation passed or failed
- live runtime proof passed or failed
- success, collapse, and failure-visible cases reached or not reached
- route or contract drift discovered during the run
If a formerly reliable prompt no longer reproduces an outcome, say that explicitly instead of
forcing a false comparison.
7. Clean up
At the end of the run:
- stop the runtime
- drop temporary databases created for the evaluation
- leave the note and artifact bundle in a cold-start-replayable state
If the user did not ask for git closure, stop after honest cleanup of runtime state and a clear
report of any remaining worktree delta created by the evaluation artifacts.
Guardrails
- Do not claim runtime proof from mocks, handler-only tests, or deterministic checks.
- Keep provider and model explicit in the note and final report.
- Use the current workflow-id autonomy route unless repo truth says otherwise.
- Keep the three proof-outcome labels frozen.
- Do not widen into implementation or backlog work unless the user explicitly asks.
- Redact secrets from captured env files, tokens, or delegated-request artifacts before keeping
them.
- If the request is actually about the development workflow rather than the runtime product, route
it with
route_workflow_request, pull state with get_control_plane_snapshot or
get_issue_execution_snapshot, then continue with the Smith workflow-family tools instead of
forcing a runtime-eval flow.
Related Skills
Reference
Read references/runtime-evaluation-playbook.md when you need:
- the standard note layout
- the default artifact bundle checklist
- the proof matrix checklist
- route reminders
- cleanup expectations