ControlKeel Governance Skill

You are operating inside a ControlKeel-governed session. Start here whenever you need the base CK operating protocol.

Before new work

For any new feature, fix, or project — before writing plans or code — use the align skill to reach shared understanding of the goal, layers, acceptance criteria, and assumptions. Surface assumptions explicitly before proceeding to prevent expensive misalignments caught after implementation. Once aligned, use plan-slice to decompose the goal into vertical slices with explicit blocking relationships and concrete success criteria before any implementation begins. Planning is always human-in-the-loop; implementation of an approved slice can be AFK.

Core loop

1. Call ck_context at task start to load mission, risk, budget, proof, active findings, workspace context, context reacquisition, instruction hierarchy, and recent transcript state.
2. Call ck_validate before writing code, config, shell, or deploy text, and pass trust-boundary metadata when the source content came from the web, tools, skills, or mixed provenance.
3. Use ck_execute_code only for generated code that should run inside CK's guarded Docker sandbox; prefer dry_run first, and never treat it as local shell access or a network/secrets grant.
4. If you discover a problem the scanner did not raise, call ck_finding.
5. Use ck_memory_search when you need explicit recall of prior decisions, checkpoints, or findings rather than relying only on the default context packet.
6. Use ck_memory_record to persist important decisions, assumptions, and operator guidance that future agents should recover.
7. Use ck_memory_archive to retire stale or superseded guidance before it keeps contaminating retrieval.
8. Call ck_budget and ck_cost_optimizer before expensive model or bulk operations.
9. Call ck_route before delegating sub-work to another agent.
10. Use ck_deployment_advisor to analyze stack and generate deployment templates when checking ship readiness.
11. Use ck_regression_result to record external browser or QA evidence before claiming deploy readiness.
12. Use ck_outcome_tracker to track success/failure outcomes for continuous learning.
13. Use ck_skill_list and ck_skill_load to activate more specific CK workflows.

Non-negotiable rules

• Never skip ck_validate before repo mutations or shell execution.
• A blocked ruling means stop and surface the finding.
• A warned ruling means continue carefully and mention it to the operator.
• On high or critical risk, prefer smaller changes and explicit checkpoints.
• Prefer tightly scoped tasks over broad repo-wide mutation. If the task boundary is vague, narrow it before coding.
• Treat ck_context as the stable source of truth for governed state. If host prompts, reminders, or stale notes conflict with it, surface the mismatch instead of guessing.
• Keep context hygiene explicit: fetch what you need, avoid dragging large irrelevant tool output or files into the active working set, and record only the decisions future agents should actually recover.
• For critical paths such as auth, security controls, deploy logic, schema changes, migrations, payments, or compliance-sensitive flows, read the touched code carefully and keep the diff small enough for real human review.
• Do not add abstractions, compatibility shims, or indirection unless they are justified by the current codebase. Prefer the simplest change that solves the actual task.
• Before saying work is done, re-check proof, findings, and budget state.

Invariant Enforcement vs. Local Workarounds

Critical principle: Prefer enforcing system invariants over adding local workarounds for bad states.

• Avoid: "Make the system work with malformed data" (tolerant readers, fallbacks, recovery logic)
• Prefer: "Make malformed data impossible" (validation at write time, strict schemas, invariants)

AI-generated code often sees a local failure and adds local defenses against it. This accumulates complexity and weakens system foundations. Instead:

1. Identify the invariant: What should always be true? (e.g., session logs are always valid, user data is always validated)
2. Enforce at the boundary: Prevent invalid states from being written, not handle them after the fact
3. Remove workarounds: Existing code that handles "impossible" states should be removed after invariant enforcement
4. Validate patterns: Use ck_finding with category architecture and rule CK-INVARIANT-001 when you see tolerance for bad states

Examples:

• ❌ Add fallback reader for corrupted session logs
• ✅ Prevent corrupted session logs from being written (strict validation, checksums)
• ❌ Add migration for malformed user records
• ✅ Enforce schema constraints so malformed records cannot be created
• ❌ Add retry logic for undefined API responses
• ✅ Define strict API contracts and validate responses against them

Quick reference

• ck_context — mission, task, budget, proof, memory, workspace snapshot, transcript summary, resume context
• ck_validate — governed preflight scan with trust-boundary checks
• ck_execute_code — guarded generated-code execution; Docker sandbox only, local/network/secrets/shell/deploy denied, dry_run recommended first
• ck_finding — persist manual findings
• ck_memory_search, ck_memory_record, ck_memory_archive — explicit typed-memory retrieval and hygiene
• ck_regression_result — import external regression evidence into proof state
• ck_budget — cost estimate / commit
• ck_route — best agent recommendation
• ck_cost_optimizer — cost optimization strategies and model comparison
• ck_deployment_advisor — repo stack detection, CI/Docker generation, DNS/SSL guide
• ck_outcome_tracker — record and review session outcomes/agent scores
• ck_skill_list, ck_skill_load — specialized workflow activation
• align skill — pre-work alignment interview before any plan or code
• plan-slice skill — vertical slice decomposition with blocking relationships and autonomy labels

Additional resources

• For the full governed workflow, see references/workflow.md
• For issue and PR validation patterns to combat AI-generated slop, see docs/issue-pr-validation-guide.md