## Summary Harden unattended `squad watch --execute` workflows so untrusted work-item content is separated from privileged execution and merge authority. ## Architecture 1. **Ingestion boundary** — retrieve issue content through a low-privilege reader and produce an immutable, provenance-bound snapshot. 2. **Execution boundary** — verify sandbox and tool-governance coverage, minimize credentials, and fail closed when requested protection is unavailable. 3. **Lifecycle boundary** — enforce admission, limits, draft/human gates, and independently protected auto-complete. ## Principles - `--yolo` remains the default approval posture for admitted trusted runs, but is not treated as a trust boundary. - Normalization and hostile-content screening may reduce exposure or hold work; they never establish that natural-language input is safe. - Trusted repositories may retain unattended auto-complete when GitHub required checks and independent server-side controls enforce it. - Public issues must remain disclosure-safe and must not contain private reproduction details. ## Non-goals - Inventing a Squad-specific sandbox or policy language. - Treating prompt wording or sanitization as the sole security boundary. - Claiming coverage for Copilot, Agency, or MCP paths before it is verified. ## Completion criteria - Trigger provenance is enforced before unattended execution. - Privileged execution is bound to an immutable snapshot. - Protected modes stop rather than silently weakening their guarantees. - Lifecycle modes and trusted auto-complete gates are explicit and testable. - Adversarial regression coverage and operator diagnostics are available. Implementation issues are attached as sub-issues.
Activity
github-actions commentedon Jul 13, 2026
🏷️ Issue type updated →
type:epic