Hive Dream — 2026-09-09

New loops

None opened — all candidates were false positives or previously-dismissed classes.

All open loops — age in days (red ≥7d, amber ≥3d)

ryanwalden's agent repeatedly calls cronjob action='show', which doesn't exist

64d

ryanwalden's agent search_files regex patterns repeatedly fail with 'unrecognized escape sequence' (Rust regex vs PCRE escaping)

64d

ryanwalden's agent thrashes vast.ai CLI via --help repeatedly (~98 calls across 2 sessions) despite the vastai-gpu-rental skill already being loaded

64d

Sandbox's default-deny per-domain network egress allowlist repeatedly blocks new research domains until manually added (Apple domains 07-07, Google Play domains 07-08)

63d

aws login --remote device-code flow reliably fails with 'Failed to decode the verification code' due to PTY echo/redraw corruption on the 2000+ char single-line code, blocking fresh AWS CLI auth for ryanwalden

62d

Uncommitted files/state in the live repo checkout or ~/.hermes/ silently vanish when a container/workspace resets, forcing manual reconstruction from session logs -- hit three different agents on three separate incidents

55d

Kanban/dispatcher worker workspaces contain docs/context for multiple unrelated client projects side-by-side (Nothy, Entravia, ProService, Shift Living) with no repo-level or channel-level scoping mechanism -- a worker run on the Entravia AI-ticket-ingestion card pulled in Nothy-specific context; mitigated only with a manual 'SCOPE LOCK' comment on the Kanban card, not a technical fix, so any future card in this shared checkout can hit the same cross-client leakage

51d

copilot_remote(launch) silently blocked whenever a session runs under HERMES_CRON_SESSION (scheduled crons AND self-created cron follow-up check-ins) — agents keep discovering this by trial and error mid-task, not upfront

50d

ryanwalden's agent stalled mid-skill-authoring session (20260725_163333_e48e64eb, building the new 'dagster' skill) after two consecutive completions returned text-only output with no tool call, leaving the user's follow-ups unanswered for the rest of the session

46d

The shared observability aggregator container (rosenblatt-observability-1, services/observability/aggregate.py --loop) silently stopped polling agent session logs after 2026-07-24T14:56 UTC — no crash log, no error, aggregate.log and status.json simply stopped updating for ~3 days, starving hive-dream's nightly scan of any new session data and producing two consecutive false '0 sessions, quiet' briefs (07-25, 07-26) even though 27 and 10 real interactive sessions happened those days respectively. User (justinhromalik) restarted the container manually on 2026-07-27; aggregator is backfilling normally as of 16:04 UTC. No host-level alert caught this — su-refresh-agents.sh's shared-service healthcheck (which already covers rosenblatt-observability-1) evidently wasn't run or wasn't surfaced during the outage window.

44d

eugenecho's agent got stuck in a truncation/continuation loop on a multi-hour AWS-credentials migration task (proservice static creds for gdk-eval-pipeline + psh-propilot-data-sync) -- three separate hermes sessions (155209, 160721, 161058) all ended mid-task with 'response truncated by output length limit' continuation prompts going unanswered, spanning ~2h (19:59-21:55 UTC) without ever reaching a completed handoff to the user

43d

ryanwalden's agent left a Slack thread hanging after a successful call transcription+summary — five follow-up user turns (seismi.co comparison requests, an image, a plain 'hi') between 20:26-22:18 UTC on 07-29 got zero replies in the primary session (513082d6), forcing the user to open two brand-new sessions later to get any response

42d

Org-wide 'Ambiguous skill name' collisions between ~/.hermes/skills/ and .claude/skills/ local/shared copies keep costing agents real turns -- hit andreszabala (skill_view failed 6x/3 sessions on research, hive-kanban, secrets, webhook-subscribe) and ryanwalden (skill_view failed 6x/3 sessions on recall-call-transcription, secrets) again on 08-04/08-05, 7+ days after the same class was first diagnosed on 07-16/07-29 with an existing remediation playbook (hermes-skill-collision-triage) that has never been run fleet-wide

35d

sagefaraday's agent got stuck in a ~6-hour execute_code/read_file retry loop while fixing entravia PR #266 (Census Deduplication review) -- repeatedly read_file'd a wrong path (repos/entravia/src/lib/reducto/smartIngestApply.ts, actually in a worktree) getting 'File not found' 3x in a row, and separately re-asserted the false claim 'PR #266 merged/patch successfully applied' verbatim across many turns before self-correcting each time, only reaching an accurate final status after ~9 repeated corrections

35d

Webhook-triggered PR review sessions misuse vision_analyze by passing GitHub PR URLs as image_url (gets 404) because the sessions lack terminal/skill_view/GitHub API tools — 13 failures across 12 sessions on 08-21, plus 60-min clarify timeouts on each

19d

Closed loops

None closed.

Aging

No loops crossed 3d/7d escalation thresholds yesterday.

Swarm health

Cron family gone silent: ryanwalden:dba0a0624a6c — last seen 2026-07-09, expected ~3.5 runs/weekday. Now 62 days silent. This is the old recall-call-transcription cron; may have been intentionally retired, but worth confirming.

Today's connections

Andres should tell Ryan about the PR #447 showstopper — Andres found that settleTerminal in queueReductoJob.ts:135-140 drops extraction_citations from resolve(), making the entire Reducto source citation feature inert. If Ryan is working on anything touching Reducto job queueing, this is relevant.

Agent spotlight

andreszabala Ran 22 sessions — 4 PR reviews via MoA plus Reducto Extract v4 pipeline update ryanwalden Switched default model to GLM-5.3 Flash + Recall.ai transcription work eugenecho GDK eval pipeline run for ProService — AWS credential issues surfaced justinhromalik Recall.ai call transcription request

andreszabala — 22 sessions

Ran 22 sessions — 4 PR reviews via MoA plus Reducto Extract v4 pipeline update

Posted MoA reviews on PRs #443, #446, and #447 (found a showstopper in #447: settleTerminal drops extraction_citations, making the citation feature inert). Also evaluated Reducto's September Extract v4 update against Entravia's reducto pipeline (156-tool-call session) and addressed review comments on PR #447.

Top session: 156 tool calls, ended: ok. Full detail in the .md brief / facts.json (not published here).

Skills used today

entravia-code-workflow — 9x today, 29x/30d, personal-only
Use when implementing code in the Entravia repo.
moa-review-profile — 4x today, 121x/30d, personal-only
Use for robust PR review via Mixture-of-Agents preset.
daily-work-streams-summary — 3x today, 28x/30d, personal-only
Build a brief, 1-line-headline-only itemized list of the distinct work streams/threads worked on today for Andres Zabala. Use when asked to
github-code-review — 3x today, 65x/30d, personal-only
Review a PR and post approve/block verdict via GitHub API. Uses MoA subagent for anti-leniency diff review.
pr-review-comment-triage — 3x today, 65x/30d, personal-only
Use when addressing PR review comments. Verifies staleness.
daily-standup-summaries — 1x today
daily-work-pickup-list — 1x today, 6x/30d, personal-only
List yesterday's in-progress items to continue today.
entravia-reducto-extraction — 1x today, 15x/30d, personal-only
Use when working on Reducto document extraction in Entravia.
entravia-reducto-pipeline — 1x today, 0x/30d, personal-only
Entravia Reducto extraction pipeline — schemas and testing.
kimai-time-tracking — 1x today, 10x/30d, personal-only
Track time in Kimai from chat: start/stop timers, log entries, list projects/activities, query timesheets. Use when user asks to log time, s

Skill usage — 30 days

moa-review-profile

121x

github-code-review

65x

pr-review-comment-triage

65x

entravia-code-workflow

29x

daily-work-streams-summary

28x

entravia-reducto-extraction

15x

kimai-time-tracking

10x

daily-work-pickup-list

6x

ryanwalden — 3 sessions

Switched default model to GLM-5.3 Flash + Recall.ai transcription work

3 sessions: updated Hermes default model to glm-5.3-flash from DeepInfra (fireworks-deepinfra-billing skill), processed Recall.ai call transcriptions, and handled a Slack canvas mention about remediation/security.

Top session: 18 tool calls, ended: ok. Full detail in the .md brief / facts.json (not published here).

Skills used today

recall-call-transcription — 3x today, 16x/30d, shared
Use when a user shares a Recall.ai explorer bot link or bot UUID and wants the call transcribed, downloaded, or summarized. Prefer the Recal
fireworks-deepinfra-billing — 1x today, 2x/30d, personal-only
Pull usage and cost data from Fireworks AI (via firectl CLI billing subcommands) and DeepInfra (via per-key usage API) — the two AI inferenc
hermes-agent — 1x today, 45x/30d, personal-only
Complete guide to using and extending Hermes Agent — CLI usage, setup, configuration, spawning additional agents, gateway platforms, skills,

Skill usage — 30 days

hermes-agent

45x

recall-call-transcription

16x

fireworks-deepinfra-billing

2x

eugenecho — 1 session

GDK eval pipeline run for ProService — AWS credential issues surfaced

Single 96-tool-call session: checked ProService GDK Google Drive for client-uploaded document, ran the gdk-eval-pipeline skill. Had corrections and an AWS credential exit-1 on the background script, but session ended OK.

Top session: 96 tool calls, ended: ok. Full detail in the .md brief / facts.json (not published here).

Skills used today

gdk-eval-pipeline — 1x today, 2x/30d, shared
End-to-end GDK (Generative Domain Knowledge) RAG evaluation pipeline for ProPilot on AWS Bedrock. Converts Excel ground truths to JSONL, ded
google-workspace — 1x today, 1x/30d, personal-only
Gmail, Calendar, Drive, Contacts, Sheets, and Docs integration for Hermes. Uses Hermes-managed OAuth2 setup, prefers the Google Workspace CL

Skill usage — 30 days

gdk-eval-pipeline

2x

google-workspace

1x

justinhromalik — 1 session

Recall.ai call transcription request

Quick 8-tool session transcribing a Recall.ai call (UUID 6fbfeee1). Used recall-call-transcription skill. Session completed normally.

Top session: 8 tool calls, ended: ok. Full detail in the .md brief / facts.json (not published here).

Skills used today

recall-call-transcription — 1x today, 16x/30d, shared
Use when a user shares a Recall.ai explorer bot link or bot UUID and wants the call transcribed, downloaded, or summarized. Prefer the Recal

Skill usage — 30 days

recall-call-transcription

16x

🧰 Tooling

2 new skills detected: daily-standup-summaries (andreszabala, personal-only), gdk-document-management (eugenecho, personal-only).
6 skills changed since last atlas build: entravia-reducto-extraction, fireworks-deepinfra-billing, hermes-model-config, moa-review-profile, pr-review-comment-triage, unattended-session-tool-restrictions.

Day in numbers

Agent	Sessions	Focus
andreszabala	22	PR review (MoA), Reducto pipeline, Entravia code
eugenecho	1	GDK eval pipeline, Google Drive
justinhromalik	1	Recall.ai transcription
ryanwalden	3	Model config, transcription, Slack canvas

Tool-category mix by agent

andreszabala

426

eugenecho

59

ryanwalden

22

justinhromalik

4

opscodingdataorchestrationresearch