skilld
All skills
simota avatar

/rally

@c805268
by shingo imotasimota/agent-skills85 stars
15

Orchestrating multi-session parallel execution via Claude Code Agent Teams API and Codex CLI Subagents — launch, manage, coordinate concurrent tasks. Use when parallel work is needed.

Ask your Agent

Use this Skill: https://skilld.dev/gh/simota/agent-skills/rally

This session only. Nothing lands on disk.

referenceparallel-learning.md

≈2.2k tokens on demand. Your agent reads this file only when SKILL.md points to it.

Parallel Learning (HARMONIZE)

Purpose: Read this after team execution when evaluating performance, adapting defaults, or rolling back a bad adaptation.

Table of Contents

  1. HARMONIZE Workflow
  2. Learning Triggers
  3. Team Effectiveness Score
  4. Team Design Matrix
  5. Adaptation Rules
  6. Safety Guardrails and Rollback
  7. Integration and Templates

HARMONIZE Workflow

COLLECT -> EVALUATE -> EXTRACT -> ADAPT -> SAFEGUARD -> RECORD
Phase Purpose Required output
COLLECT capture execution facts EXECUTION_DATA
EVALUATE compute performance TES and grade
EXTRACT identify patterns composition, size, model, or ownership insight
ADAPT propose default changes bounded proposal with rationale and rollback
SAFEGUARD verify safety invariant checks, evidence check, snapshot
RECORD persist learning journal entry, Lore share, Judge data

EXECUTION_DATA

Capture after every team session:

EXECUTION_DATA:
  session_id: <unique id>
  task_type: <feature | bugfix | refactor | migration | test>
  team_pattern: <Frontend+Backend | Feature_Parallel | Pipeline | Specialist_Team | Code+Test+Docs>
  team_size: <int>
  subagent_types:
    - name: <teammate name>
      type: <general-purpose | Explore | Plan>
      model: <sonnet | opus | haiku>
      role: <description>
  task_count: <int>
  completed_tasks: <int>
  retried_tasks: <int>
  ownership_conflicts: <int>
  total_file_ops: <int>
  timing:
    total_seconds: <int>
    estimated_sequential_seconds: <int>
  integration_checks:
    build_pass: <bool>
    tests_pass: <bool>
    lint_pass: <bool>
  user_interventions:
    - type: <team_size_override | composition_override | task_reassign | manual_fix | abort>
      detail: <description>
  final_status: <SUCCESS | PARTIAL | BLOCKED | FAILED>

Learning Triggers

ID Condition Scope Actions
RY-01 every team execution completes Lightweight COLLECT + EVALUATE
RY-02 same team pattern fails or conflicts 3+ times Full all 6 phases
RY-03 user overrides team size or composition Full all 6 phases
RY-04 Judge sends quality feedback Medium COLLECT + EVALUATE + EXTRACT + RECORD
RY-05 Lore sends a parallel pattern update Medium COLLECT + EVALUATE + EXTRACT + RECORD
RY-06 30+ days since the last full HARMONIZE cycle Full all 6 phases

Trigger Priority

  1. RY-03
  2. RY-02
  3. RY-04
  4. RY-05
  5. RY-06
  6. RY-01

Team Effectiveness Score

TES = Parallel_Efficiency x 0.30
    + Task_Economy x 0.20
    + Conflict_Prevention x 0.20
    + Integration_Quality x 0.20
    + User_Autonomy x 0.10

Components

Component Weight Definition
Parallel_Efficiency 0.30 estimated_sequential_time / actual_parallel_time, capped at 1.0
Task_Economy 0.20 completed_tasks / total_tasks x (1 - retry_rate)
Conflict_Prevention 0.20 1 - (conflicts / total_file_ops)
Integration_Quality 0.20 (build + tests + lint) / total_checks
User_Autonomy 0.10 1 - (user_overrides / total_decisions)

Grading Scale

Grade TES range Interpretation
A >= 0.90 excellent
B >= 0.80 good, protect with approval
C >= 0.70 acceptable adaptation target
D >= 0.60 below standard
F < 0.60 full review required

Data Requirements

  • TES requires >= 3 data points for the same task-type and team-pattern combination.
  • Grades are provisional until >= 5 data points.
  • Trend analysis requires >= 10 data points.
  • Below the threshold, report INSUFFICIENT_DATA.

Team Design Matrix

Use cumulative TES by task type:

                 | feature | bugfix | refactor | migration | test |
Frontend+Backend |    —    |   —    |    —     |     —     |  —   |
Feature Parallel |    —    |   —    |    —     |     —     |  —   |
Pipeline         |    —    |   —    |    —     |     —     |  —   |
Specialist Team  |    —    |   —    |    —     |     —     |  —   |
Code+Test+Docs   |    —    |   —    |    —     |     —     |  —   |

— means INSUFFICIENT_DATA.

Update Rules

  1. Record the team-pattern and task-type outcome after each RY-01.
  2. When a cell reaches >= 3 data points, calculate the TES average and grade.
  3. Prefer higher-graded patterns at design time.
  4. If every pattern is INSUFFICIENT_DATA, fall back to reference/team-design-patterns.md.

Adaptation Rules

Allowed Scope by Grade

Current grade Allowed adaptations Approval
A no auto-change human required
B no auto-change human required
C team-size defaults, subagent_type preference tuning auto with snapshot
D size defaults, model heuristics, pattern recommendations auto with snapshot
F full review including pattern, size, model, and ownership reassessment auto with snapshot

Adaptation Types

Type Example Max per session
Team size default change default to 3 for bugfix work 2
subagent_type preference prefer Explore for refactor investigation 1
Model heuristic use opus for complex migration design 1
Pattern recommendation recommend Pipeline for migration tasks 1
Ownership refinement split shared test directories 1

Total maximum: 3 parameter default changes per session.

Application Process

  1. Propose the change with rationale, expected TES impact, and rollback plan.
  2. Run SAFEGUARD.
  3. Save a snapshot.
  4. Apply the change.
  5. Monitor the next 3 executions.

Safety Guardrails and Rollback

Mechanism Rule
Change volume limit maximum 3 parameter default changes per session
Auto-adaptation ceiling TES >= B requires human approval
Rollback snapshot save state before every adaptation
Diminishing returns 3 consecutive TES improvements < 0.02 -> pause HARMONIZE
Lore sync extracted patterns must be sent to Lore
Minimum evidence no adaptation with < 3 execution data points
File ownership invariant never modify the exclusive_write or shared_read protocol

Rollback Protocol

  1. Detect regression: TES drops >= 0.10 after adaptation, measured across the next 3 executions.
  2. Load the pre-adaptation snapshot.
  3. Restore previous defaults and heuristic state.
  4. Record the rollback in .agents/rally.md with root cause analysis.
  5. Share the rollback event to Lore as a negative pattern.

Integration and Templates

Integration Points

Partner Direction Data exchanged
Lore Rally -> Lore team composition patterns, TES trends, rollback events
Lore Lore -> Rally validated parallel patterns
Judge Rally -> Judge TES scores and completion quality
Judge Judge -> Rally QUALITY_FEEDBACK

Lore Sync Protocol

RALLY_TO_LORE_PATTERN:
  type: PARALLEL_EXECUTION_PATTERN
  pattern_name: <descriptive name>
  source_data:
    task_type: <task type>
    team_pattern: <team pattern>
    team_size: <int>
    sample_size: <int>
    tes_impact: <float>
  pattern_detail: <description>
  confidence: <HIGH | MEDIUM | LOW>
  actionable: <bool>
LORE_TO_RALLY_PATTERN:
  type: PARALLEL_PATTERN_UPDATE
  pattern_name: <descriptive name>
  source_agents: <list of contributing agents>
  recommendation: <action to consider>
  priority: <HIGH | MEDIUM | LOW>

Session Feedback Record

## Team Session Feedback — <session_id>

**Date:** YYYY-MM-DD
**Task Type:** <feature | bugfix | refactor | migration | test>
**Team Pattern:** <Frontend+Backend | Feature_Parallel | Pipeline | Specialist_Team | Code+Test+Docs>
**Team Size:** <int> | **Subagent Types:** <type list> | **Models:** <model list>
**Final Status:** <SUCCESS | PARTIAL | BLOCKED | FAILED>
**TES:** <score> (Grade: <A-F>) | **Previous TES:** <score> (Grade: <A-F>)

Team Profile Update

## Team Profile Update — <date>

**Team Pattern:** <pattern name>
**Task Type:** <task type>
**Previous Grade:** <grade or INSUFFICIENT_DATA>
**New Grade:** <grade>
**Data Points:** <count>

Adaptation Log

## Adaptation Log — <date>

**Trigger:** <RY-xx>
**Current TES:** <score> (Grade: <grade>)
**Target TES:** <expected score>

### Safeguard Check
- File ownership invariant: PASS / FAIL
- Evidence threshold (>= 3): PASS / FAIL
- Grade ceiling check: PASS / FAIL
- Snapshot saved: <snapshot identifier>

### Outcome
- TES after 3 executions: <score>
- Rollback needed: YES / NO

Source: SKILL.md on GitHub

Third-party checks

1 warning13d5 checks · Risk SAFE
  • Gen Agent Trust Hub13d

    The skill orchestrates parallel agent tasks and is generally safe. However, it has an indirect prompt injection surface because it passes untrusted task descriptions from upstream sources to subagents without sanitization. It also supports highly autonomous execution modes that bypass user approval for tool use.

  • Socket13d

    No alerts

  • Snyk13d

    Risk: LOW · No issues

  • Runlayer6mo

    1/12 files flagged

  • ZeroLeaks5mo

    Score: 93/100 · 2 sections analyzed

Provenance

Signed by skilld at c805268. This ties the file your Agent reads to that commit on GitHub. It does not review the instructions.

Last checked against GitHub 2 days ago.

Activeupdated 2 weeks ago

README badge

README badge for simota/agent-skills/rally