free-will-pivot - SKILL.md Agent Skill

name: free-will-pivot description: "When a free-will block discovers its assigned task is already done, pivot to autonomous high-value discovery instead of re-running tests and exiting." activation: keywords: - free will - free-will - exploration block - already done - pivot - autonomous discovery - unstructured block - self-directed when: - A free-will or exploration shift block starts and the intended task is already complete - The agent detects prior work covers the planned scope - No specific task was assigned and the block is open-ended tool_doctrine: discovery: workflow: - scan_tasks_queue - check_system_health - audit_state_files - run_overdue_maintenance - pursue_novatrade_research validation: workflow: - verify_prior_work_complete - confirm_no_regressions - assess_highest_value_pivot output_contract: required: - pivot_reason - discovery_source - work_performed - summary - files_changed - verification - confidence

Free-Will Pivot: Autonomous Discovery on Task Completion

When a free-will or exploration block discovers its intended work is already done, the agent must not waste the block by re-running tests and exiting. Instead, it pivots to the highest-value autonomous work available.

When To Use

A free-will shift block starts and the prior session already completed the planned work.
The agent's initial assessment reveals nothing new to do on the original topic.
An unstructured block has no specific assignment.

Step 1: Confirm Prior Work (< 2 minutes)

Before pivoting, verify the prior work is genuinely complete:

Read the most recent OUTPUT/ file for the same shift block stem.
Confirm the output has a valid CONTRACT section with confidence: medium or higher.
Spot-check one or two claims (e.g., run the specific tests mentioned, confirm file changes exist).
If prior work is incomplete or broken, resume it instead of pivoting.

Do NOT run the entire test suite as a confirmation step — that wastes the block.

Step 2: Discovery Cascade

Execute this priority-ordered cascade. Stop at the first category that yields actionable work:

Priority 1: Queued Tasks

ls TASKS/*.md  # pending tasks (not .done, .failed, .inprogress)

If any pending tasks exist, execute the highest-priority one using the task-execution skill.

Priority 2: System Health Issues

cat HEARTBEAT.md  # check for any non-healthy items
cat STATE/progress_history.json | tail -5  # recent scores and trends

Look for:

Any heartbeat check showing degraded/failing status
Progress dimensions marked degrading or below 70
Metric anomalies (failure rates, stale data, unbounded growth)
Service issues (high memory, stale PIDs, log bloat)

Priority 3: Overdue Maintenance

Scan for systemic hygiene issues:

STATE/ file sizes — anything over 5MB needs rotation or pruning
Log retention — run scripts/log_retention.py if logs are large
Test pollution — check if test runs are contaminating production state files
Dead code / unused files — stale .inprogress files, orphaned configs
Dependency health — outdated packages, security advisories

Priority 4: Code Quality

Run ruff or linting on recently changed files
Fix any test failures in the suite (run targeted, not full suite)
Address TODO/FIXME/HACK comments in actively maintained files

Priority 5: NovaTrade Research & Advancement

If the system is healthy and no maintenance is needed, invest in NovaTrade:

Review recent trade journal entries for patterns
Research strategy improvements based on rejection reasons
Investigate execution pipeline gaps
Advance any open NovaTrade implementation phases

Priority 6: Infrastructure Improvement

If everything above is clean:

Improve monitoring coverage
Add missing test cases for edge cases discovered in prior sessions
Enhance error messages or logging in problem areas
Document architectural decisions that lack ADRs

Step 3: Execute the Pivot Work

Once a target is identified:

State your pivot reason — one sentence explaining why you chose this work over alternatives.
Execute with full rigor — treat the pivot work with the same discipline as an assigned task. Use implementation-team patterns for non-trivial changes.
Write tests if modifying production code.
Write output to OUTPUT/<original_stem>__<timestamp>.md preserving the shift block lineage.

Step 4: Output Contract

Every pivot execution must end with:

## CONTRACT
pivot_reason: <why this work was chosen over alternatives>
discovery_source: <which cascade priority yielded the work — tasks/health/maintenance/quality/novatrade/infra>
work_performed: <one-line description>
summary: <2-3 sentences of what was done and why it matters>
files_changed: <comma-separated paths, or "none">
verification: <how correctness was confirmed>
confidence: <low | medium | high>

Anti-Patterns to Avoid

Anti-Pattern	Why It's Wrong	Do This Instead
Re-run full test suite and exit	Wastes the entire exploration block	Spot-check 1-2 things, then pivot
Repeat work from a prior session	Duplicates effort with zero new value	Read OUTPUT/ first, resume only if incomplete
Pick trivially easy work	Low-value use of exploration time	Follow the priority cascade honestly
Skip verification on pivot work	Pivot work is still production code	Apply same test/verify standards
Touch > 3 unrelated areas	Unfocused scatter leads to half-done changes	Pick one thing, finish it completely

Examples of Good Pivots

From shift_20260331_15_free_will_pm sessions:

Trade journal rotation: Discovered 54K-line unbounded file growth, implemented rotation with atomic writes and auto-trigger — systemic fix, not a band-aid.
Metrics pollution cleanup: Found 93% of "contract failures" were test artifacts, fixed 5 tests and cleaned 1,312 phantom entries — restored metric trustworthiness.
Progress history cap: STATE file growing 4MB/day, added MAX_HISTORY_ENTRIES enforcement — prevented future performance degradation.
Risk engine data source fix: Collector reading stale mock data instead of real drawdown state — wired to actual novatrade_risk_state.json.

Each pivot was discovered autonomously, completed fully, and produced measurable improvement.