AutoSci: Skill for Claude Code

/hiera-experiment

Purpose and activation boundary

Use this skill only after the user explicitly selects the optional Hiera path for an existing AutoSci experiment. Hiera is an isolated search sidecar:

• It reads the linked wiki experiment/design and runtime contract as inputs.
• It stores all Hiera state below runs/hiera///.
• It searches immutable candidate snapshots and records evidence in the sidecar ledger.
• It never creates wiki entities, changes wiki schema or experiment status, rewrites wiki/log.md, or writes search results back to the wiki.
• It does not replace /exp-design, /exp-run, /exp-eval, or /research. A confirmatory result must return to the normal, user-gated workflow.

Before creating or changing a run, read:

1. README.md for the user-facing sequence and command examples.
2. references/interface.md for contract fields, source layouts, and remote receipts.
3. The relevant runtime contract files when a field or permission is unclear.

Never infer a method, metric direction, resource boundary, seed policy, or remote environment from a vague request. Ask for the missing decision or ask the user to provide a text constraint.

State model and command ownership

Treat the lifecycle as a state machine, not as one automatic command:

init
  -> inspect/complete semantic_space and task_contract
  -> propose semantic points
  -> author drafts and candidate specs
  -> admit immutable candidates
  -> explicit human source/contract review
  -> approve frozen contract
  -> preflight
  -> run(screen)
  -> choose one iteration branch:
       propose a new semantic direction
       tune reviewed config_schema children
       loop existing candidates at deep depth
       status/recover pending remote attempts
  -> finalize sidecar report
  -> optional formal confirmation through /exp-run and /exp-eval

The CLI subcommand and the agent action are different:

Stage	CLI command	Agent-owned action
Initialize	init	Create the sidecar and show readiness errors.
Complete inputs	none	Edit only user-confirmed semantic or contract inputs; validate them.
Propose	propose	Show point IDs and metadata; do not create source.
Author	none	Write reviewed source/spec under this run's drafts; do not call propose again.
Admit	admit	Import exactly the reviewed draft into an immutable snapshot.
Approve	approve	Run only after explicit human approval of the execution boundary.
Preflight	preflight	Check a candidate without reserving evaluation budget.
Evaluate	run	Execute one immutable snapshot at screen or deep depth.
Tune	tune	Create and evaluate bounded immutable config children.
Loop	loop	Schedule deep bouts among existing eligible candidates only.
Observe/recover	status / recover	Poll or reconcile an existing remote attempt; never resubmit it.
Finalize	finalize	Write the exploratory sidecar report only.