Sub-Agent Output Templating

Overview

This skill provides standardized templates for sub-agent OUTPUT formatting. It complements subagent-prompting (P0.1) which defines INPUT structure:

Use this skill when:

• Defining output requirements for sub-agent invocations
• Parsing sub-agent results in pipeline stages
• Ensuring consistent log formats across all agents

Log File Format

File Location

logs/{agent-name}-{YYYYMMDD-HHMMSS}.yaml

Example: logs/bulwark-code-auditor-20260111-143022.yaml

YAML Schema

# Required: Metadata block
metadata:
  agent: {agent-name}           # e.g., bulwark-code-auditor
  timestamp: {ISO-8601}         # e.g., 2026-01-11T14:30:22Z
  model: {model-used}           # sonnet, haiku, or opus
  task_id: {unique-identifier}  # For tracking across pipeline stages
  duration_ms: {execution-time} # Execution duration in milliseconds

# Required: Goal from the prompt (for traceability)
goal: "{GOAL from 4-part prompt}"

# Required: Completion report
completion:
  why:
    problem: "{What was broken/missing}"
    root_cause: "{Why it happened}"
    solution: "{What was implemented}"

  what:
    - file: {path}
      lines: "{range}"
      change: "{description}"

  trade_offs:
    gained:
      - "{benefit 1}"
    cost:
      - "{drawback 1}"

  risks:
    - risk: "{description}"
      mitigation: "{how addressed}"
      severity: {low|medium|high|critical}

  next_steps:
    - "{action item 1}"

# Required for code-writing agents (omit for read-only agents):
# Pipeline suggestions from implementer-quality.sh output
pipeline_suggestions:
  - pipeline: "{recommended pipeline name}"
    target_files:
      - "{file path}"
    reason: "{why this pipeline is recommended}"

# Required: Summary for main thread (100-300 tokens)
summary: |
  {Concise summary for main thread consumption}

# Required: Diagnostic output
diagnostics:
  model_requested: {model}
  model_actual: {model}
  context_type: {main|forked}
  parent_vars_accessible: {true|false}
  hooks_fired:
    - {hook-name}
  execution_time_ms: {duration}
  completion_status: {success|error|timeout}

Task Completion Report (WHY/WHAT/TRADE-OFFS/RISKS)

Every sub-agent MUST conclude with this structured report. This enables explicit decision documentation rather than implicit code changes.

WHY Section

Document the problem and solution rationale.

why:
  problem: "Authentication bypass vulnerability in refresh token path"
  root_cause: "Token validation skips expiry check on refresh"
  solution: "Added isExpired() check to refresh token handler"

Guidelines:

• problem: What was broken, missing, or needs improvement
• root_cause: The underlying reason (not just symptoms)
• solution: What was done to address it

WHAT Section

List all changes made with file locations.

what:
  - file: src/auth/token.ts
    lines: "45-52"
    change: "Added isExpired() check before token refresh"
  - file: src/auth/token.test.ts
    lines: "120-145"
    change: "Added test for expired refresh token rejection"

Guidelines:

• One entry per file modified
• Include line ranges for precise location
• Describe the change, not the code

TRADE-OFFS Section

Acknowledge explicit compromises made.

trade_offs:
  gained:
    - "Security: Expired tokens now properly rejected"
    - "Compliance: Meets OWASP session management requirements"
  cost:
    - "Performance: Additional DB lookup on refresh (negligible)"
    - "Complexity: New error handling path for expired tokens"

Guidelines:

• Be honest about costs
• Quantify impact where possible
• Include both technical and business trade-offs

RISKS Section

Document forward-looking concerns.

risks:
  - risk: "Existing sessions with expired refresh tokens will fail"
    mitigation: "Grace period of 24h for migration"
    severity: medium
  - risk: "Grace period could be exploited"
    mitigation: "Monitor for unusual refresh patterns"
    severity: low

Severity Levels:

NEXT STEPS Section

List follow-up actions for pipeline or human.

next_steps:
  - "Monitor refresh failure rate for 24h"
  - "Remove grace period after migration window"
  - "Update documentation for new error codes"

Guidelines:

• Actionable items only
• Include owner if known (e.g., "DevOps: Update monitoring dashboard")
• Order by priority

Summary Format for Main Thread

Purpose

The summary is returned to the main thread for pipeline decision-making. It should enable the orchestrator to:

1. Understand key findings without reading full log
2. Decide next pipeline stage
3. Report status to user

Token Budget

Summary Template

Found [N] [severity] issue(s): [brief description].
[Action taken / recommendation].
[Key risk or follow-up if any].

Examples

Simple (120 tokens):

Found 1 critical vulnerability: refresh tokens not validated for expiry.
Fixed by adding isExpired() check in token.ts:45-52. Added regression test.
Risk: existing sessions may fail during 24h migration window.

Moderate (200 tokens):

Found 3 issues in authentication module:
- 1 critical: token expiry bypass (fixed)
- 1 medium: weak password hashing (fixed, migration needed)
- 1 low: verbose error messages (fixed)

All issues addressed with tests added. Migration script created for password re-hashing.
Next: run migration in staging, monitor for 48h before production.

What to Include

• Finding count and severity
• Actions taken
• Key risks or blockers
• Recommended next steps

What to Exclude

• Full reasoning or analysis
• Code snippets
• Verbose explanations
• Duplicate information from log

Pipeline Suggestions in Summary (Code-Writing Agents)

Code-writing agents (e.g., bulwark-implementer) that invoke implementer-quality.sh and receive pipeline suggestions MUST include them in the summary with MANDATORY language. This ensures the orchestrator sees and acts on them per SA6.

MANDATORY FOLLOW-UP (SA6): Run the following pipeline(s):
  - {pipeline} on {target_files} ({reason})
Orchestrator MUST evaluate each suggestion and either execute or document deferral per SA6.

Read-only agents (reviewers, auditors) omit this section.

Diagnostic Output

Purpose

Enable automated behavioral testing without mocking. Diagnostics verify:

• Correct model was used
• Context isolation worked (for context: fork agents)
• Hooks fired as expected
• Execution completed successfully

Location

logs/diagnostics/{agent-name}-{YYYYMMDD-HHMMSS}.yaml

Format

skill: subagent-output-templating
timestamp: 2026-01-11T14:30:22Z
diagnostics:
  model_requested: sonnet
  model_actual: sonnet
  context_type: forked      # main or forked
  parent_vars_accessible: false  # Should be false for forked
  hooks_fired:
    - Stop
  execution_time_ms: 4520
  completion_status: success  # success, error, timeout
notes: "Task completed successfully"

Diagnostic Fields

Quick Reference

Minimal Log Template

metadata:
  agent: {name}
  timestamp: {ISO-8601}
  model: sonnet
  task_id: "{id}"
  duration_ms: 0

goal: "{goal}"

completion:
  why:
    problem: "{problem}"
    root_cause: "{cause}"
    solution: "{solution}"
  what:
    - file: {path}
      lines: "{range}"
      change: "{description}"
  trade_offs:
    gained: ["{benefit}"]
    cost: ["{cost}"]
  risks:
    - risk: "{risk}"
      mitigation: "{mitigation}"
      severity: medium
  next_steps:
    - "{action}"

# Include for code-writing agents only (omit for read-only agents):
pipeline_suggestions:
  - pipeline: "{pipeline name}"
    target_files: ["{path}"]
    reason: "{reason}"

summary: |
  {100-300 token summary}

diagnostics:
  model_requested: sonnet
  model_actual: sonnet
  context_type: forked
  parent_vars_accessible: false
  hooks_fired: []
  execution_time_ms: 0
  completion_status: success

Summary Checklist

[ ] Count and severity of findings stated
[ ] Actions taken described
[ ] Key risks mentioned
[ ] Next steps listed
[ ] Under 300 tokens
[ ] Pipeline suggestions with MANDATORY language (code-writing agents only)

Output Location Checklist

[ ] Main log: logs/{agent}-{YYYYMMDD-HHMMSS}.yaml
[ ] Diagnostics: logs/diagnostics/{agent}-{YYYYMMDD-HHMMSS}.yaml

Timestamp Formats

Why two formats?

• File names: No colons (filesystem-safe on Windows), compact, lexically sortable
• YAML fields: Standard ISO-8601 for parsing and interoperability

Important: Always use {YYYYMMDD-HHMMSS} in file paths, never {timestamp} or {ts}.

Related Skills

• subagent-prompting (P0.1): Defines INPUT structure (GOAL/CONSTRAINTS/CONTEXT/OUTPUT)
• pipeline-templates (P0.3): Pre-defined workflows that consume this output format

References

For extended examples and edge cases, see references/examples.md.