Updates & self-healing
How agents verify skills and toolchain updates, request authorization for impacting changes, and autonomously heal minor regressions.
When an AI agent operates within Marrow, it is responsible for maintaining environmental integrity, checking for skill/toolchain updates, and autonomously repairing minor issues introduced by upstream changes.
Update Check Workflow
During execution, the agent should periodically check for toolchain and skill updates:
npx skills check # Check installed skills against their upstream sources
npm outdated # Check workspace dependencies (optional)The agent categorizes every detected update into one of two tiers:
| Update Tier | Definition | Action |
|---|---|---|
| Tier 1: Safe / Patch | Bug fixes, non-breaking skill improvements, internal optimizations, or added documentation that do not alter existing schema contracts or workspace configurations. | Update automatically without touching user configuration or unrelated project code. |
| Tier 2: Impacting / Breaking | Changes that modify YAML schema validation, deprecate root or device keys, alter layout placement geometry, change icon identifiers, or affect existing output HTML files. | Pause and alert the user. Detail the exact impact and wait for explicit user authorization before updating. |
Impact Notification & Authorization Protocol
If a Tier 2 update is detected, the agent must not apply the update silently. It must prompt the user with an explicit impact brief:
- What is updating: Package or skill name and versions (
v1.2.0 → v2.0.0). - What will be affected: Affected topology files in
input/, configuration keys inmarrow.config.yaml, or output behaviors. - Potential risks: Deprecated keys requiring migration, visual layout shifts, or required rebuilds.
- Authorization request: Ask the user directly whether to proceed with the update and migration.
Only proceed with the update once the user gives explicit consent.
Autonomous Self-Healing Loop
Upstream updates may occasionally cause minor regressions: renamed YAML attributes, newly required fields, deprecated icon IDs, or compiler warnings. When this occurs, the agent does not stop or request user intervention for trivial fixes. Instead, it executes the autonomous self-healing loop:
Autonomous Self-Healing Loop
Self-contained diagnostic and automated remediation feedback loop for AI agents
1. Detect & Diagnose
Run the validator with --json to obtain machine-readable diagnostics:
npm run marrow -- validate input/<file>.yaml --jsonDiagnostics return structured fields:
code: Stable error code (e.g.MRW-E001,MRW-E010).line&column: Exact source location.hint: Actionable remediation guidance.suggestions: Ranked replacements for renamed keys or missing icon IDs.
2. Autonomous Remediation
The agent patches the issue directly:
- Deprecated or renamed keys: Update the YAML keys according to the new schema (e.g., migrating legacy annotations).
- Missing or reorganized icons: Query
npm run marrow -- icons search <keyword>to identify the verified replacement ID and update the device node. - Uncompiled runtime or styles: Run
npm run compileto synchronize the runtime engine with updated dependencies. - Icon pack index: Run
npm run icons:indexif vendor icon packs were updated.
3. Verification
Re-run validation and headless visual snapshotting:
npm run marrow -- validate input/<file>.yaml
npm run marrow -- snapshot input/<file>.yaml -o /tmp/verify.pngValidation must exit with code 0 before the agent considers the self-healing process complete.
Escalation Boundaries
The agent heals minor regressions automatically, but must escalate to the user under the following conditions:
- The update requires deleting or fundamentally altering user network data (e.g., removing unmapped subnets or links).
- The update breaks custom integration scripts outside Marrow's core compiler.
- Self-healing cannot reach exit code
0after iterative attempts. - The update involves sensitive security credentials or changes external repository remotes.