| name | xvision-autooptimizer-ops |
| description | Operate the xvision autooptimizer: distill Observations into candidate Patterns, run the gate against today and the untouched test period, record blind findings, activate or retire Patterns, and preserve the lineage evidence for audit. |
xvision autooptimizer ops
Use this skill for offline Pattern distillation work. AutoOptimizer
commands are offline-only; do not run them inside a live trading
decision process.
When to use
Run this skill when you are offline, outside a live trading decision, and need
to distill candidate Patterns from Observations, run the gate, inspect a run,
activate a passing run, or retire a Pattern.
When NOT to use
- Inspecting Pattern or Observation inventory → use
xvision-memory-ops
- Reading flywheel velocity or overall health → use
xvision-flywheel-ops
- Any task involving a live trading cycle or real-time data → no skill applies
Trigger examples
- Distill a new Pattern from last week's Observations for agent abc123
- Run the autoresearch gate on run_id xyz with a Sharpe baseline of 0.8
- Activate run xyz after the gate passed
- Retire Pattern p_001 — soft delete only
- Inspect autoresearch run xyz and show me the gate output JSON
- Write a blind finding for this autoresearch run before I see the scores
- Check whether Pattern p_002's training cutoff is correct
Standard flow
Note: xvn optimizer is deprecated. All sub-commands print a deprecation
notice and delegate to xvn optimize. Use xvn optimize directly in new scripts.
xvn optimize run --strategy <strategy_id> --cycles 3 --json
xvn optimize run-cycle --strategy <strategy_id> \
--budget 100 --provider openrouter --model kimi-k2 \
--objective sharpe --experiments-per-cycle 5 \
--json
xvn optimize inspect <run_id> --json
The gate, activate, and retire operations are now managed via the dashboard
Optimizer panel. Use xvn optimize inspect <run_id> to fetch run state for audit.
Invariants
- Never activate from a single Observation cohort.
- The gate's numeric decision (Kept / Dropped) must pass independently
before activation.
- The finding is qualitative context, not the verdict.
- A Pattern's training cutoff must come from the latest source
Observation's bar timestamp, not wall clock time.
- Retiring is soft-delete first; hard delete requires explicit operator
confirmation through the memory janitor path.
Evidence to capture
xvn optimize inspect <run_id> --json
- Gate input/output JSON, including the blind finding fields.
- Activated Pattern row and contributing Observation ids.
- Look-ahead-protection regression output before declaring an
activation path change done.
Gotchas
Single-cohort activation: Never activate from one Observation cohort — the
invariants section states this. The gate must pass first.
Wall-clock training cutoff: The cutoff must come from the latest source
Observation's bar timestamp, not the current date. Using wall clock time
introduces look-ahead contamination.
Finding written after scores visible: The blind finding must be composed
before the numeric gate scores are read. Running xvn optimize run-cycle and
then writing the finding retrospectively breaks the audit trail.
Hard delete without janitor path: Retiring is soft-delete. Hard delete
requires the explicit memory janitor confirmation flow — skipping it destroys
lineage evidence.
Wrong skill for inventory inspection: xvn optimize inspect shows one
run. Use xvision-memory-ops to list all Patterns or Observations across an
agent.
Owner
autoresearch-ops track (team/contracts/autoresearch-ops.md)