Per Bernd's call: the guardrails prevent security harm but not LLM content errors, so the
worker should triage + draft, not auto-send, until reply quality is proven (matches the
build-stage/recoverability posture).
run_conservative triages NEW messages into a reviewed digest (state_dir/worker-digest.md)
with drafted replies, posts ONE progress note, tracks seen message ids (schedule-safe
dedup), and sends NOTHING to other agents / marks nothing read. `warden worker run
--execute` now runs this conservative tier; `--full-auto` opts into the auto-send path.
Live-verified with the LLM brain on the real inbox: produced a high-quality draft reply to
a secrets-engine coordination message and correctly flagged the llm-connect custody request
as NEEDS YOU. Conservative mode is safe to schedule (T4). 244 tests, lint clean.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
HubClient gains writes (mark_read, send_reply, add_progress). execute_plan/execute_plans
run the safe, allowlisted actions autonomously: route_answer (reply with the computed
answer + auto mark-read), reply (LLM-drafted body), progress_note, mark_read. Escalated
plans and non-auto-executable kinds are left for a human; every action is metadata-only
(no secret value read/sent/logged).
Deliberate guardrail: propose_catalog_diff and any code/routing change is NOT auto-executed
even under full-auto — a bad catalog commit could misroute credentials, so it goes to human
review (recoverability over convenience). AUTO_EXECUTABLE is the messaging/hub tier only.
`warden worker run --execute` runs the executor (dry-run still default). 7 executor tests
(reply+mark, with/without body, escalated skip, catalog-diff-left-for-human, progress,
failure-without-crash); 243 pass, lint clean. First live --execute shakedown is the
operator's (staged rollout); T4 schedules it.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
llm-connect is operational (operator set OPENROUTER_API_KEY). Contract discovered from
the running service: POST /execute {"prompt":...} -> {"content":...}.
LlmConnectBrain embeds the fixed charter + the inbox message as untrusted data, calls
/execute, and parses a JSON action plan (_extract_json tolerates fences/prose), escalating
defensively on malformed/empty/transport errors. The build_plans guardrail still enforces
the allowlist + no-secret invariant on whatever the model returns — the LLM cannot widen
ops-warden's authority. `warden worker run --brain rule|llm` selects the planner.
Live-verified on the real inbox: the LLM brain planned a sensible reply+mark_read for a
secrets-engine coordination message and correctly escalated a secret-custody request as
out-of-lane — better classification than the deterministic RuleBrain.
6 new tests, 236 pass, lint clean. T3 (guarded executor) and T4 (scheduling) remain.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Foundation for an autonomous worker that handles ops-warden's State Hub coordination
lane via llm-connect (Bernd's call: full-auto in-scope + scheduled, staged dry-run ->
manual -> scheduled). T1 is the llm-connect-independent, safe slice:
src/warden/worker.py — HubClient (read unread to_agent=ops-warden), Brain protocol,
deterministic RuleBrain (answers clear routing questions, escalates the rest),
PlannedAction/WorkerPlan model, guardrail allowlist + validate_action enforced
brain-agnostically (no-secret invariant + prod-config + off-allowlist all escalate),
render_plans dry-run output. `warden worker run --dry-run` (default); --execute refused
(exit 2) until the guarded executor (T3) lands.
Guardrails are load-bearing because full-auto has no human in the loop: message content
is untrusted data, the allowlist is enforced regardless of what the brain proposes.
Hard dependency flagged in the workplan: the brain is llm-connect, which needs its
provider key (OPENROUTER_API_KEY, deferred CCR-2026-0003) before it can run.
18 worker tests; 229 pass, lint clean. Live dry-run against the real hub verified.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
