governed-dev
Evidence-first development with fail-closed behavior
When & Why to Use This Skill
The Governed Development skill enforces a rigorous, evidence-first framework for AI-driven software engineering. It ensures operational integrity by implementing 'fail-closed' behavior, requiring verifiable evidence for every claim, and restricting the agent to a secure, pre-approved set of tools and scripts.
Use Cases
- Automated Quality Gates: Ensuring that smoke tests and build processes stop immediately upon failure, preventing the agent from proceeding with unverified code.
- Compliance and Auditing: Maintaining a persistent audit trail by mandating that every development action produces and links to specific artifacts (logs, summaries, or reports).
- Secure Environment Control: Restricting AI actions to authorized wrapper scripts while blocking high-risk operations like arbitrary shell commands or external network requests (curl/wget).
- Reliable Progress Reporting: Eliminating 'hallucinated' success by marking any claim without captured output or physical artifacts as UNVERIFIED.
| name | governed-dev |
|---|---|
| description | Evidence-first development with fail-closed behavior |
| user-invocable | false |
Governed Development Skill
This skill enforces evidence-first development practices for Interlock.
Core Principles
1. Evidence-First
Every claim must be backed by evidence:
- If a command was not run: Mark the claim as
UNVERIFIED - If output was not captured: Do not claim success
- If artifacts were not produced: Do not claim completion
## Claim: Smoke tests pass
**Evidence**: `artifacts/claude/20260110T120000Z/smoke/summary.md`
**Status**: VERIFIED - Exit code 0, all steps passed
vs.
## Claim: Smoke tests pass
**Evidence**: None
**Status**: UNVERIFIED - Command not executed
2. Fail-Closed
Never "limp past" failures:
- Any non-zero exit: Stop and report
- Missing artifacts: Stop and report
- Partial success: Report as failure
# WRONG - Ignoring failures
./scripts/claude/smoke.sh || true
# RIGHT - Respecting failures
./scripts/claude/smoke.sh
if [ $? -ne 0 ]; then
echo "GATE FAILED"
exit 1
fi
3. Artifacts Are Deliverables
Every operation should produce artifacts:
- Link to artifacts in reports
- Preserve artifacts for audit
- Never overwrite without archiving
See: `artifacts/claude/20260110T120000Z/smoke/summary.md`
Verification Rules
Before Making Claims
- Run the relevant wrapper script
- Check exit code
- Read the summary.md artifact
- Only then make claims about results
After Failures
- Report the failure immediately
- Link to error artifacts (stderr.log)
- Do not attempt to "fix and continue" without explicit approval
- Mark all downstream claims as BLOCKED
Allowed Operations
| Operation | Tool | Purpose |
|---|---|---|
| Read files | Read | Inspect code and artifacts |
| Search files | Grep, Glob | Find relevant code |
| Run wrappers | Bash(./scripts/claude/*) | Execute verified scripts |
Prohibited Operations
- Arbitrary shell commands
- Network requests (curl, wget)
- Modifying production code without planning
- Claiming success without evidence