M4 - Tier-1 composition: exact warden invoke tools + per-attempt grant for the LLM fallback #6
Labels
No labels
correctness
coverage
milestone:M4
polish
security
tech-debt
No milestone
No project
No assignees
1 participant
Notifications
Due date
No due date set.
Dependencies
No dependencies set
Reference
public/warden#6
Loading…
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Next mesh milestone (0.16.0), in progress now. When Tier-0 can't fix an incident, hand the LLM auto-resolver an EXACT allowlisted
warden invoke <cap> --target <id> --param k=vtoolset composed from the routed actuator's registered manifest (enum product <= N, else decline+escalate - never a prefix/glob), plus a per-attempt HMAC grant (bound to incident+capabilities+exp) so the child drivesPOST /invokeand never holds NATS creds; add acontext_note(Tier-0 failure reason) to the prompt. Pieces:resolve.compose_invoke_tools,autoresolve.attempt(extra_tools=, context_note=)+ClaudeAgentRunner.with_env,remediateTier-1 wiring, docs.Done in v0.16.0 (pushed). Tier-1 composition landed: exact tools from the routed actuator manifest (enum-only, safe-char allowlist, decline+escalate on non-enum/overflow/unsafe), per-attempt HMAC grant bound to incident+registered-caps, child gets only WARDEN_ORCHESTRATOR_URL+WARDEN_INVOKE_GRANT (never NATS creds), context_note outside the evidence fence. A wire-manifest RCE (denylist->allowlist) and a feature-inert bug (composed tools now emit --incident so the grant binding matches, with an e2e compose->cli->/invoke->grant.allows test) were caught+fixed in review. 582 tests.