4.3 KiB
Orchestrator Learnings (Universal)
Cross-project heuristic adjustments based on observed variance data.
Note: This file contains generic patterns only. Project-specific evidence is stored in each project's
docs/tasks/orchestrator-learnings.json.
Task Type Multipliers
Apply these multipliers to base estimates from ORCHESTRATOR.md:
| Task Type | Base Estimate | Multiplier | Confidence | Samples | Last Updated |
|---|---|---|---|---|---|
| STYLE_FIX | 3-5K | 0.64 | MEDIUM | n=1 | 2026-02-05 |
| BULK_CLEANUP | file_count × 550 | 1.0 | MEDIUM | n=2 | 2026-02-05 |
| GUARD_ADD | 5-8K | 1.0 | LOW | n=0 | - |
| SECURITY_FIX | 8-12K | 2.5 | LOW | n=0 | - |
| AUTH_ADD | 15-25K | 1.0 | HIGH | n=1 | 2026-02-05 |
| REFACTOR | 10-15K | 1.0 | LOW | n=0 | - |
| TEST_ADD | 15-25K | 1.0 | LOW | n=0 | - |
| ERROR_HANDLING | 8-12K | 2.3 | MEDIUM | n=1 | 2026-02-05 |
| CONFIG_DEFAULT_CHANGE | 5-10K | 1.8 | MEDIUM | n=1 | 2026-02-05 |
| INPUT_VALIDATION | 5-8K | 1.7 | MEDIUM | n=1 | 2026-02-05 |
Phase Factors
Apply to all estimates based on task position in milestone:
| Phase Position | Factor | Rationale |
|---|---|---|
| Early (tasks 1-3) | 1.45 | Codebase learning overhead |
| Mid (tasks 4-7) | 1.25 | Pattern recognition phase |
| Late (tasks 8+) | 1.10 | Established patterns |
Estimation Formula
Final Estimate = Base Estimate × Type Multiplier × Phase Factor × TDD Overhead
Where:
- Base Estimate: From ORCHESTRATOR.md task type table
- Type Multiplier: From table above (default 1.0)
- Phase Factor: 1.45 / 1.25 / 1.10 based on position
- TDD Overhead: 1.20 if tests required
Known Patterns
BULK_CLEANUP
Pattern: Multi-file cleanup tasks are severely underestimated.
Why: Iterative testing across many files, cascading fixes, and debugging compound the effort.
Observed: +112% to +276% variance when using fixed estimates.
Recommendation: Use file_count × 550 instead of fixed estimate.
ERROR_HANDLING
Pattern: Error handling changes that modify type interfaces cascade through the codebase.
Why: Adding fields to result types requires updating all callers, error messages, and tests.
Observed: +131% variance.
Multiplier: 2.3x base estimate when type interfaces are modified.
CONFIG_DEFAULT_CHANGE
Pattern: Config default changes require more test coverage than expected.
Why: Security-sensitive defaults need validation tests, warning tests, and edge case coverage.
Observed: +80% variance.
Multiplier: 1.8x when config changes need security validation.
INPUT_VALIDATION
Pattern: Security input validation with allowlists is more complex than simple validation.
Why: Comprehensive allowlists (e.g., OAuth error codes), encoding requirements, and security tests add up.
Observed: +70% variance.
Multiplier: 1.7x when security allowlists are involved.
STYLE_FIX
Pattern: Pure formatting fixes are faster than estimated when isolated.
Observed: -36% variance.
Multiplier: 0.64x for isolated style-only fixes.
Changelog
| Date | Change | Samples | Confidence |
|---|---|---|---|
| 2026-02-05 | Added BULK_CLEANUP category | n=2 | MEDIUM |
| 2026-02-05 | Added STYLE_FIX multiplier 0.64 | n=1 | MEDIUM |
| 2026-02-05 | Confirmed AUTH_ADD heuristic accurate | n=1 | HIGH |
| 2026-02-05 | Added ERROR_HANDLING multiplier 2.3x | n=1 | MEDIUM |
| 2026-02-05 | Added CONFIG_DEFAULT_CHANGE multiplier 1.8x | n=1 | MEDIUM |
| 2026-02-05 | Added INPUT_VALIDATION multiplier 1.7x | n=1 | MEDIUM |
Update Protocol
Graduated Autonomy:
| Phase | Condition | Action |
|---|---|---|
| Now | All proposals | Human review required |
| After 3 milestones | <30% change, n≥3 samples, HIGH confidence | Auto-update allowed |
| Mature | All changes | Auto with notification, revert on regression |
Validation Before Update:
- Minimum 3 samples for same task type
- Standard deviation < 30% of mean
- Outliers (>2σ) excluded
- New formula must not increase variance on historical data
Where to Find Project-Specific Data
- Project learnings:
<project>/docs/tasks/orchestrator-learnings.json - Cross-project metrics:
jarvis-brain/data/orchestrator-metrics.json