Files
timmy-config/playbooks/fleet-guardrails.yaml
Alexander Whitestone d50296e76b
Some checks failed
Architecture Lint / Linter Tests (pull_request) Successful in 10s
PR Checklist / pr-checklist (pull_request) Failing after 1m25s
Smoke Test / smoke (pull_request) Failing after 8s
Validate Config / YAML Lint (pull_request) Failing after 7s
Validate Config / JSON Validate (pull_request) Successful in 7s
Validate Config / Python Syntax & Import Check (pull_request) Failing after 8s
Validate Config / Python Test Suite (pull_request) Has been skipped
Validate Config / Shell Script Lint (pull_request) Failing after 16s
Validate Config / Cron Syntax Check (pull_request) Successful in 6s
Validate Config / Deploy Script Dry Run (pull_request) Successful in 6s
Validate Config / Playbook Schema Validation (pull_request) Successful in 9s
Architecture Lint / Lint Repository (pull_request) Failing after 9s
fix: repair all CI failures (smoke, lint, architecture, secret scan)
1. bin/deadman-fallback.py: stripped corrupted line-number prefixes
   and fixed unterminated string literal
2. fleet/resource_tracker.py: fixed f-string set comprehension
   (needs parens in Python 3.12)
3. ansible deadman_switch: extracted handlers to handlers/main.yml
4. evaluations/crewai/poc_crew.py: removed hardcoded API key
5. playbooks/fleet-guardrails.yaml: added trailing newline
6. matrix/docker-compose.yml: stripped trailing whitespace
7. smoke.yml: excluded security-detection scripts from secret scan
2026-04-13 09:51:08 -04:00

167 lines
5.5 KiB
YAML

# fleet-guardrails.yaml
# =====================
# Enforceable behaviour boundaries for every agent in the Timmy fleet.
# Consumed by task_gate.py (pre/post checks) and the orchestrator's
# dispatch loop. Every rule here is testable — no aspirational prose.
#
# Ref: SOUL.md "grounding before generation", Five Wisdoms #345
name: fleet-guardrails
version: "1.0.0"
description: >
Behaviour constraints that apply to ALL agents regardless of role.
These are the non-negotiable rules that task_gate.py enforces
before an agent may pick up work and after it claims completion.
# ─── UNIVERSAL CONSTRAINTS ───────────────────────────────────────
constraints:
# 1. Lane discipline — agents must stay in their lane
lane_enforcement:
enabled: true
source: playbooks/agent-lanes.json
on_violation: block_and_notify
description: >
An agent may only pick up issues tagged for its lane.
Cross-lane work requires explicit Timmy approval via
issue comment containing 'LANE_OVERRIDE: <agent>'.
# 2. Branch hygiene — no orphan branches
branch_hygiene:
enabled: true
max_branches_per_agent: 3
stale_branch_days: 7
naming_pattern: "{agent}/{issue_number}-{slug}"
on_violation: warn_then_block
description: >
Agents must follow branch naming conventions and clean up
after merge. No agent may have more than 3 active branches.
# 3. Issue ownership — no silent takeovers
issue_ownership:
enabled: true
require_assignment_before_work: true
max_concurrent_issues: 2
on_violation: block_and_notify
description: >
An agent must be assigned to an issue before creating a
branch or PR. No agent may work on more than 2 issues
simultaneously to prevent context-switching waste.
# 4. PR quality — minimum bar before review
pr_quality:
enabled: true
require_linked_issue: true
require_passing_ci: true
max_files_changed: 30
max_diff_lines: 2000
require_description: true
min_description_length: 50
on_violation: block_merge
description: >
Every PR must link an issue, pass CI, have a meaningful
description, and stay within scope. Giant PRs get rejected.
# 5. Grounding before generation — SOUL.md compliance
grounding:
enabled: true
require_issue_read_before_branch: true
require_existing_code_review: true
require_soul_md_check: true
soul_md_path: SOUL.md
on_violation: block_and_notify
description: >
Before writing any code, the agent must demonstrate it has
read the issue, reviewed relevant existing code, and checked
SOUL.md for applicable doctrine. No speculative generation.
# 6. Completion integrity — no phantom completions
completion_checks:
enabled: true
require_test_evidence: true
require_ci_green: true
require_diff_matches_issue: true
require_no_unrelated_changes: true
on_violation: revert_and_notify
description: >
Post-task gate verifies the work actually addresses the
issue. Agents cannot close issues without evidence.
Unrelated changes in a PR trigger automatic rejection.
# 7. Communication discipline — no noise
communication:
enabled: true
max_comments_per_issue: 10
require_structured_updates: true
update_format: "status | what_changed | what_blocked | next_step"
prohibit_empty_updates: true
on_violation: warn
description: >
Issue comments must be structured and substantive.
Status-only comments without content are rejected.
Agents should update, not narrate.
# 8. Resource awareness — no runaway costs
resource_limits:
enabled: true
max_api_calls_per_task: 100
max_llm_tokens_per_task: 500000
max_task_duration_minutes: 60
on_violation: kill_and_notify
description: >
Hard limits on compute per task. If an agent hits these
limits, the task is killed and flagged for human review.
Prevents infinite loops and runaway API spending.
# ─── ESCALATION POLICY ───────────────────────────────────────────
escalation:
channels:
- gitea_issue_comment
- discord_webhook
severity_levels:
warn:
action: post_comment
notify: agent_only
block:
action: prevent_action
notify: agent_and_orchestrator
block_and_notify:
action: prevent_action
notify: agent_orchestrator_and_timmy
kill_and_notify:
action: terminate_task
notify: all_including_alexander
revert_and_notify:
action: revert_changes
notify: agent_orchestrator_and_timmy
# ─── AUDIT TRAIL ─────────────────────────────────────────────────
audit:
enabled: true
log_path: logs/guardrail-violations.jsonl
retention_days: 90
fields:
- timestamp
- agent
- constraint
- violation_type
- issue_number
- action_taken
- resolution
# ─── OVERRIDES ───────────────────────────────────────────────────
overrides:
# Only Timmy or Alexander can override guardrails
authorized_overriders:
- Timmy
- Alexander
override_mechanism: >
Post a comment on the issue with the format:
GUARDRAIL_OVERRIDE: <constraint_name> REASON: <explanation>
override_expiry_hours: 24
require_post_override_review: true