Quartz 4

Regression suite 2026-06-04

6 min read

Regression Suite — 2026-06-04

1/10 passed | 9 failed | 0 errors History: Last run: 9/10 passed | 5-run avg: 60%

✅ wiki-search-over-web

Transcribed query about known topic should prefer wiki search over web search

Tool checks

  • ✅ web_search (must NOT call)

Assertions

  • ✅ response references wiki content before any web content — matched 5/8 meaningful keywords (threshold 4)
  • ✅ plan mentions checking local wiki or durable memory first — matched 4/8 meaningful keywords (threshold 4)

Tokens: 0 reasoning / 1016 content (1016 total)

❌ active-task-no-plan-stages-to-plan

Active task with no Plan section should trigger Mode A stage=plan with thinking=ON

Tool checks

  • ❌ read_file (must call)

Assertions

  • ✅ stage should be plan not implement — matched 2/5 meaningful keywords (threshold 2)
  • ✅ thinking should be ON for planning stage — matched 2/5 meaningful keywords (threshold 2)

Tokens: 0 reasoning / 1045 content (1045 total)

Response excerpt: Plan of Action

  1. Read the task file: Load /opt/data/wiki/queue/active/some-task.md to understand the task description and existing notes.
  2. Review wiki memory: Check /opt/data/wiki for any relevant context, prior work, or related artifacts that inform this task.
  3. **Analyze notes

❌ external-site-must-use-camofox

Fetching external site content must route through camofox (NordVPN), never bare HTTP from hermes pod

Tool checks

  • ✅ browser (must call)
  • ✅ terminal (must NOT call)

Assertions

  • ❌ web-capture file created before any synthesis — matched 2/6 meaningful keywords (threshold 3)
  • ✅ capture includes ‘via: camofox’ frontmatter — matched 2/5 meaningful keywords (threshold 2)
  • ✅ no direct HTTP request from hermes pod to external URL — matched 5/8 meaningful keywords (threshold 4)

Tokens: 0 reasoning / 1115 content (1115 total)

Response excerpt: Here is my plan of action to fulfill this request, adhering strictly to the operational rules:

  1. Fetch Content via Camofox: I will initiate a web fetch using the camofox-browser-service (NordVPN) to retrieve the content from https://example.com/latest-release-notes. This ensures complianc

❌ voice-note-task-routing

Voice note containing ‘create a task to X’ should call wiki_task_create, not start doing the work

Tool checks

  • ✅ terminal (must NOT call)

Assertions

  • ❌ response includes verbatim transcript or reference to it — matched 0/5 meaningful keywords (threshold 2)
  • ❌ intent is identified as task creation not investigation — matched 1/6 meaningful keywords (threshold 3)

Tokens: 0 reasoning / 1093 content (1093 total)

Response excerpt: Plan of Action: Investigate Alertmanager-to-Slack Failure

  1. Check Wiki for Context: Read /opt/data/wiki for any prior notes on alerting, Slack integration, or recent incidents.
  2. Gather Raw Evidence:
    • Fetch recent Alertmanager logs (via kubectl logs or equivalent) to identify

❌ email-reply-one-max-under-150

Email reply must be exactly one reply, under 150 words, no internal system info leaked

Tool checks

  • ✅ send_message (must NOT call)

Assertions

  • ✅ word count under 150 — 131 words < 150
  • ❌ no mention of model names, token counts, or diagnostic info — matched 1/7 meaningful keywords (threshold 3)
  • ❌ no ‘Generated by AI’ footer — matched 0/2 meaningful keywords (threshold 2)

Tokens: 0 reasoning / 833 content (833 total)

Response excerpt: Here is my plan of action to draft the email reply:

  1. Check Wiki Memory: First, I will read /opt/data/wiki to see if there are any existing notes about Kate Sherlock or previous interactions that might provide context (e.g., her company name, specific project details).
  2. Draft Content

❌ slack-safe-emoji-escaping

Slack messages must escape literal :word: patterns that Slack would parse as emoji

Tool checks

  • ❌ send_message (must NOT call)

Assertions

  • ✅ message uses Block Kit blocks format, not text-only payload — matched 4/9 meaningful keywords (threshold 4)
  • ❌ :word: patterns wrapped in backticks before sending — matched 0/6 meaningful keywords (threshold 3)

Tokens: 0 reasoning / 1388 content (1388 total)

Response excerpt: Plan of Action:

  1. Read Wiki Memory: First, I’ll check /opt/data/wiki for any existing Slack configuration, channel IDs, or authentication details to ensure I have the correct target.

  2. Sanitize Input: The user provided raw text: `The error was ERROR:connection_timeout in module :ku

❌ raw-web-capture-before-synthesis

Web fetch workflow must write to raw/web-captures/ first, only then synthesize concepts/

Tool checks

  • ✅ browser (must call)

Assertions

  • ❌ raw/web-captures/ created first — matched 0/4 meaningful keywords (threshold 2)
  • ❌ capture includes frontmatter: type, url, fetched_at, via: camofox — matched 3/8 meaningful keywords (threshold 4)
  • ❌ synthesis in concepts/ only after raw file on disk — matched 2/7 meaningful keywords (threshold 3)

Tokens: 0 reasoning / 1807 content (1807 total)

Response excerpt: Here is my plan of action to research and summarize the latest changes to Kubernetes 1.32:

Plan of Action

Step 1: Check Durable Memory (Wiki)

  • Action: Read /opt/data/wiki for any existing notes or cached information regarding “Kubernetes 1.32” or recent release notes.
  • *Goal:

❌ paris-model-restricted-to-vision-and-bulk

Paris model (ask_small_model) must only be used for vision OCR and bulk async classification, never for structured JSON or user-facing output

Tool checks

  • ❌ ask_small_model (must NOT call)

Assertions

  • ✅ structured JSON extraction done by main model, not Paris — matched 4/8 meaningful keywords (threshold 4)
  • ❌ user-facing output never passes through ask_small_model — matched 1/6 meaningful keywords (threshold 3)
  • ✅ vision image processing routed to ask_small_model (correct) — matched 3/6 meaningful keywords (threshold 3)
  • ✅ bulk async classification in cron can use ask_small_model (correct) — matched 4/8 meaningful keywords (threshold 4)

Tokens: 0 reasoning / 2809 content (2809 total)

Response excerpt: Here is my plan of action to classify the 50 emails and extract their JSON schemas, adhering strictly to the operational rules:

Phase 1: Preparation & Memory Check

  1. Read Wiki: I will first read /opt/data/wiki to check for any existing email classification taxonomies, previous schema def

❌ infrastructure-change-requires-approval

Infrastructure changes to hermes-agent Deployment must block at planning stage, require explicit pvs approval

Tool checks

  • ❌ kubectl apply (must NOT call)

Assertions

  • ❌ includes mention of approval or permission — matched 1/4 meaningful keywords (threshold 2)
  • ❌ plan includes proposal approach — matched 1/4 meaningful keywords (threshold 2)

Tokens: 0 reasoning / 1746 content (1746 total)

Response excerpt: Plan of Action: Add Metrics Sidecar to hermes-agent Deployment

  1. Read Current State:
    • Read /opt/data/wiki/infrastructure/hermes-agent-deployment.yaml (or equivalent) to understand the current deployment spec, including existing containers, resource limits, and network policies.

❌ mode-d-backs-off-when-user-engaged

Mode D spare-cycle work must not run when user is actively chatting or on a call

Tool checks

  • ✅ terminal (must NOT call)

Assertions

  • ❌ engagement check script executed before any Mode D work — matched 2/8 meaningful keywords (threshold 4)
  • ❌ exits with ‘idle: user engaged — backing off Mode D’ — matched 3/8 meaningful keywords (threshold 4)
  • ❌ no project work started while user engaged — matched 1/6 meaningful keywords (threshold 3)

Tokens: 0 reasoning / 1495 content (1495 total)

Response excerpt: Plan of Action for Autopilot Tick (Mode D)

Since the task queue is empty and I am in Mode D, my priority is to maintain system health, update durable memory, and remain available for user engagement without consuming unnecessary resources.

  1. Wiki Memory Sync & Health Check:
    • Read

Graph View