Files
TalkEdit/docs/ai-policy.md
2026-04-15 16:36:21 -06:00

2.4 KiB

AI Execution Policy

Purpose: define what autonomous AI can do in this repository without explicit human approval.

Default Mode

  • AI may implement and debug within approved scope.
  • AI must run validation commands after code changes.
  • AI must stop and escalate when blocked by policy or ambiguity.

Allowed Autonomous Actions

  1. Edit frontend, backend, shared schema, docs, and scripts.
  2. Add/modify tests related to the task.
  3. Run non-destructive validation commands.
  4. Update project docs and Copilot instructions when behavior changes.

Restricted Actions (Require Approval)

  1. Security/privacy-sensitive logic changes.
  2. Data migrations or destructive file operations.
  3. Credential handling changes or secrets management changes.
  4. Breaking API/schema changes.
  5. Build/release signing, packaging, and deployment automation changes.

Prohibited Actions

  1. Destructive git commands (git reset --hard, force pushing protected branches).
  2. Deleting user project/media data.
  3. Bypassing required checks in CI.

Required Validation Workflow

For each autonomous task:

  1. Implement smallest safe change set.
  2. Run lint/type/test/build checks for impacted scope.
  3. Inspect errors and fix with bounded retries.
  4. Re-run checks until green or escalated.
  5. Produce concise summary with risks and assumptions.

Escalation Triggers

AI must ask a human when:

  1. Requirements are ambiguous and affect user-visible behavior.
  2. Multiple product choices are plausible with no clear preference.
  3. Potential legal, security, or compliance impact exists.
  4. CI remains failing after 3 repair attempts in the same area.
  5. A requested operation conflicts with this policy.

Required Artifacts In AI PR/Change Summary

  1. What changed.
  2. Why it changed.
  3. Validation commands and outcome.
  4. Residual risks.
  5. Follow-up tasks.

Risk Levels

  • Low: docs, styling, isolated refactors, non-critical bugfixes.
  • Medium: feature additions with contract-stable behavior.
  • High: API/schema/security/export pipeline/transcription pipeline changes.

High-risk changes require explicit human review before merge.

TalkEdit-Specific Rules

  1. Preserve compatibility for desktop bridge contracts unless explicitly approved.
  2. Keep routers thin and business logic in backend services.
  3. Export/transcription pipeline changes must include regression tests.
  4. Linux WebKit startup behavior and media URL consistency are mandatory regression targets.