Approval volume for routine file.write/shell.exec creates a pending pile-up

[vikasprogrammer]

Surfaced by the fleet-insights skill. The default policy routes a lot of routine ops to human approval, and many are never resolved.

Evidence (45-day window)

• file.write → approve: 52 (instapods) + 43 (instawp); file.write [pending]: 7 (instapods) + 6 (instawp).
• shell.exec → approve: 36 (instapods) + 63 (instawp); shell.exec [pending]: 21 (instawp).
• Related dead-end: unresolved approvals block agents the same way unanswered questions do (Unattended runs (cron/Slack/automation) strand on blocking ask/approval #138).

The tension

This is deliberate governance — the whole point is mediation — so this is a tuning decision, not a blind loosen. But a 21-deep pending shell.exec queue means the current defaults create more human load than gets serviced, so agents stall.

Options

• Revisit config/policy defaults for low-risk file.write/shell.exec (e.g. narrow the "→ ask" rule to riskier paths/commands; keep the never guardrails).
• Lean on the existing "Always approve" affordance + per-agent scoping so a human curbs repeat approvals once.
• Surface an "N pending approvals are blocking agents" nudge so the queue doesn't silently grow.

Files: config/policy, src/governance/policy.ts, Inbox approval UI.