Operate the DURABLE initiatives board — the Postgres store, the sync, LLM recaps, the workbench viewer, the signal->initiative router and the /api/ask assistant. Use for: the initiatives viewer/board, initiatives.current/latest, the initiative store/sync, initiative recaps, the vllm-recap model, the initiatives assistant/chat, "what's on my board" as a durable page. The one-off scan is `initiative-scan`.
Operate the DURABLE initiatives board — the Postgres store, the sync, LLM recaps, the workbench viewer, the signal->initiative router and the /api/ask assistant. Use for: the initiatives viewer/board, initiatives.current/latest, the initiative store/sync, initiative recaps, the vllm-recap model, the initiatives assistant/chat, "what's on my board" as a durable page. The one-off scan is `initiative-scan`.
initiatives subsystem operations
The durable, queryable counterpart to the on-demand /initiative-scan scan, and the
successor to the retired agent-ops TUI's momentum panel. Four layers store → sync → viewer → recaps + a router (signal→initiative)
a read-only assistant. Code: ~/workspace/devrc/scripts/initiatives/ (sync.py, viewer.py,
route.py, assistant.py, recap.py, run-sync.sh, run-viewer.sh); engine =
scripts/session-analysis/initiative-scan.py. Everything is workbench-only,
serverMode-gated — the laptop (no ~/.server-mode) correctly runs none of it.
Arc: devrc/claudedocs/handoff-initiatives-nextstep-dispatch-shipped-2026-07-26.md +
handoff-initiatives-consolidation-2026-07-22.md (read first).
Reference files (~/.claude/skills/initiatives/reference/, source
~/workspace/devrc/claude/skills/initiatives/reference/) — read on demand:
viewer-board.md — board layout, states/chips/search, per-card actions, archive lifecycle. Read when changing the viewer UI or debugging a card's lane/state.
clawgate-tasks.md — the initiative:<slug> tag join, dispatch guard, fetch shape + the fetch worker's termination bounds (body phase fixed 2026-08-11; connect/headers still unbounded, and the single-flight guard does NOT hold it to one thread). Read when a dispatched task doesn't link, when touching tagging, or before claiming that fetch cannot leak threads.
agent-devpod.md — devpod deploy, kubeclaw 0.7.x hardening values, image 2026.6.11-py + web-search-off, identity pin, least-priv DB role. Read when deploying/bumping the agent pod or debugging a crash-loop.
How a card comes to exist (discovery model)
SESSION/ACTIVITY-FIRST. Initiatives are (repo, topic) groups mined from Claude SESSIONS
activity telemetry. topic = the transcript ai-title (~/.claude/projects/*/<sid>.jsonl,
{"type":"ai-title"}); recency = last user-turn ts. Noise floor MIN_SESSION_TURNS=1,
MIN_TOPIC_TOKENS=2 (constants in initiative-scan.py).
SESSIONS-ONLY. A card is ALWAYS session-backed (source ∈ {session, both}). A handoff doc
of ANY filename can ANCHOR a session group (best_title_match → source=both, adopting its
slug/summary/next_step/investigations for ENRICHMENT) but NEVER creates a standalone card
(pure source=doc is DROPPED in combine_docs_and_groups).
Timing is session-only — last_touch/momentum come from the session last-user-turn
(+ git/telemetry/live signals); doc_touch_epoch does NOT feed last_touch.
undocumented=true tags a session-only (no doc-anchor) card.
Title-drift MERGE is OFF (ENABLE_TITLE_DRIFT_MERGE=False): only Pass-1 EXACT-title
grouping, so distinct sessions don't over-merge into token-salad slugs. Recognizability comes
from the card's start ›/you › lines + the client-side fuzzy matchQ search.
Store is v6: source/undocumented @v4; opening_message genesis prompt @v5;
search_text (full per-session user-turn text, capped 6000, search-only-never-rendered) @v6.
Fusion (git + telemetry + live tmux → momentum/last-touch/next-step) persists to Postgres.
Key facts (verify against live state before asserting)
Thing
Value
Code
~/workspace/devrc/scripts/initiatives/ + the engine scripts/session-analysis/initiative-scan.py
Store
initiatives schema in the homelab mailbox Postgres (SAME DB as mail_actions; ns mailbox, mailbox-postgres-0). KUBECONFIG=$KC_HOMELAB
Store reach
via scripts/mail-actions/_db.py (MailDB), loaded by explicit importlib path — NOT sys.path (its sibling llm.py shadows repo-cos's). Default = kubectl port-forward; direct in-cluster mode (PR #156) via MAILBOX_PG_HOST (+MAILBOX_PG_PORT) or MAILBOX_PG_DIRECT=1
Tables
snapshots (one row/run: id,captured_at,host,days_window,telemetry_available); initiative_snapshot (append-only, one row/initiative/run: slug/repo/title/momentum/last_touch/commits/merged_prs/open_prs/session_count/telem_*/summary/current_doc/open_investigations/docs + recent_messages/recent_commits jsonb); recaps (standalone: repo,slug + identity/identity_hash + status/status_hash + legacy recap); assistant_log (standalone audit); archived (standalone: repo,slug PK + title/reason/archived_at; in NO view → no VIEW_VERSION bump)
Views
latest = rows from max(snapshot_id) (ghost-free; the viewer reads this). current = DISTINCT ON (repo,slug) newest across ALL history → includes aged-out "ghosts" (the router reads this — matching dormant initiatives is desirable). Both DROP+CREATE-guarded by a version marker (/, both )
status
export KUBECONFIG=$KC_HOMELAB# the store is only reachable via port-forward here
systemctl --user status initiatives-sync.timer initiatives-viewer.service --no-pager | head -30
systemctl --user list-timers | grep initiatives
journalctl --user -u initiatives-sync.service -n 30 --no-pager # grep -i telemetry for on/off
curl -sf http://192.168.50.250:8899/healthz; echo# viewer health
PSQL='kubectl -n mailbox exec mailbox-postgres-0 -- psql -U mailbox -d mailbox -c'$PSQL"select id,captured_at,host,days_window,telemetry_available from initiatives.snapshots order by id desc limit 5;"$PSQL"select count(*) from initiatives.latest;"# viewer's set (ghost-free)$PSQL"select count(*) from initiatives.current;"# router's set (latest + N ghosts)$PSQL"select repo,slug,momentum,last_touch from initiatives.latest order by last_touch desc limit 15;"$PSQL"select repo,slug,left(identity,60),left(status,50) from initiatives.recaps order by slug limit 10;"
operate
# run a sync NOW (both routes go through run-sync.sh → telemetry-on):
KUBECONFIG=$KC_HOMELAB systemctl --user start initiatives-sync.service
# confirm: journalctl --user -u initiatives-sync.service | grep -i "telemetry\|recap"# expect "reader creds provisioned — telemetry-on" and a "recap N new/M cached" line.# dry-run the transform without writing (needs the DB port-forward for the read only):
nix-shell -p 'python3.withPackages(p:[p.psycopg2 p.requests])' --run \
'python ~/workspace/devrc/scripts/initiatives/sync.py --dry-run --days 4'# router — which initiative does a free-text signal belong to (read-only, suggests):
nix-shell -p 'python3.withPackages(p:[p.psycopg2 p.requests])' --run \
'python ~/workspace/devrc/scripts/initiatives/route.py "polish the clawgate chat scroll"'# assistant PRIMARY path = the OpenClaw agent devpod (ask via the viewer):
curl -s -X POST http://192.168.50.250:8899/api/ask -H 'Content-Type: application/json' \
-d '{"question":"whats stalled and waiting on me"}' | python3 -m json.tool # intent=agent when the devpod answered# agent devpod status/logs (homelab):
KUBECONFIG=$KC_HOMELAB kubectl -n devpod-initiatives get pods,helmrelease -A 2>/dev/null; \
KUBECONFIG=$KC_HOMELAB kubectl -n flux-system get helmrelease initiatives-agent
POD=$(KUBECONFIG=$KC_HOMELAB kubectl -n devpod-initiatives get pods -l app.kubernetes.io/instance=initiatives -o name | tail -1)
KUBECONFIG=$KC_HOMELAB kubectl -n devpod-initiatives logs $POD -c agent --tail=50 # agent loop# run a skill-tool IN-POD (grounded JSON; the agent calls these):
KUBECONFIG=$KC_HOMELAB kubectl -n devpod-initiatives exec$POD -c agent -- \
python3 /data/repos/devrc/scripts/initiatives/skills/query.py blocked_on_me
# hit the gateway directly (bypass the viewer): token = sha256("gw-"+HOOKS_TOKEN), model openclaw/initiatives
KUBECONFIG= kubectl -n mailbox mailbox-postgres-0 -- psql -U mailbox -d mailbox -c \
nix-shell -p --run \
curl -sf -X POST http://192.168.50.250:8899/refresh;
curl -s -X POST http://192.168.50.250:8899/api/dispatch -H \
-d
Grounded next step (nextstep.py, read-only): every card carries one — documented cards
show their parsed next_step; EMERGING cards show a distinct next (suggested) › line
inferred from a real field (open-PR/investigation/last-prompt/status/stalled) — NEVER
invented. Chat: "what should I do next on <initiative>" → the recommend_next_step tool
(agent + regex fallback).
deploy a change
# scripts (sync.py/route.py/assistant.py/recap.py + wrappers) ship via home-manager:
home-manager switch --flake ~/workspace/devrc --impure # workbench; or scripts/ship.sh after merge# after a viewer.py change (X-Restart-Triggers covers viewer.py+run-viewer.sh on switch;# restart explicitly if you edited out of band):
systemctl --user restart initiatives-viewer.service
# regenerate the store immediately (e.g. after a sync.py or recap change):
KUBECONFIG=$KC_HOMELAB systemctl --user start initiatives-sync.service
# tests: scripts/initiatives/tests/ via nix-shell pytest (pure fixtures; query/agent_client/assistant/sync)# agent devpod deploy → ~/.claude/skills/initiatives/reference/agent-devpod.md
⚠ gotchas (each cost real time)
🔴 X-Restart-Triggers does NOT cover every long-lived sibling — a "shipped" change can be silently INERT.
The viewer importlib-caches its siblings at startup and nix/home.nix only triggers on
viewer.py+run-viewer.sh. So a change to initiative-scan.py (the viewer's attach_tmux
reuses scan.match_tmux_to_initiatives verbatim for the live overlay) or assistant.py switches
cleanly and does nothing until you systemctl --user restart initiatives-viewer.service.
A fresh sync alone won't do it — the live badge is a render-time overlay, not stored.
Running run-viewer.sh from an INTERACTIVE shell needs env -u shellHook.homelab-talos/.envrc
is use flake, and that flake's shellHook exports a relativeKUBECONFIG=homelab-kubeconfig.
The wrapper only defaults (${KUBECONFIG:-<absolute>}) so it defers to the relative one, and the
inner nix-shell -p … re-executes the inherited $shellHook, re-exporting it even if you fixed
KUBECONFIG first — the store read then fails silently from any cwd that isn't the repo root.
Use env -u shellHook ./run-viewer.sh. Systemd is unaffected (minimal env, no direnv).
Don't pin absolute slug counts in docs or tests. The live store drifted 139→144 over ~3 days;
every doc or assertion carrying a total goes stale within days. Assert/state properties (floors,
set-equality, ratios) instead — this already rotted repo-cos's corpus tests three days running.
Schema migration = DROP+CREATE the views, never CREATE OR REPLACE. Adding a column to
initiative_snapshot reorders the latest/current views' columns; CREATE OR REPLACE VIEW
REJECTS a column-name/order change (cannot change name of view column …) → froze the whole store
on the v1→v2 deploy. Fix: bump the VIEW_VERSION/LATEST_VIEW_VERSION marker so _ensure_view
does DROP VIEW IF EXISTS + CREATE (ACCESS EXCLUSIVE, momentary). recaps + are
(in no view) → adding columns there needs NO view-marker bump. — the unit fixtures only assert SQL
strings, they can't catch this.
COMMENT ON VIEW
VIEW_VERSION
LATEST_VIEW_VERSION
v6
Sync
initiatives-sync.timer (systemd --user, workbench, OnUnitActiveSec=15min) → run-sync.sh → sync.py --days 5 → initiative-scan.py --json. Lookback = 5d (Zach's call — a tight recent board; 14d was too cluttered; override via INITIATIVES_SYNC_DAYS). Momentum buckets are absolute last-touch ages (<2d active / 2-7d slowing / ≥7d stalled), so at 5d "stalled" (≥7d) is structurally EMPTY and "cooling"/slowing (2-5d) is the going-quiet triage signal; widen the window if deep-stalled work is wanted. Double-gated serverMode && enableInitiativesSync. Timeout 600s (headroom for a cold-recap batch; warm runs ~6-15s). 90-day retention prune each run
Sync creds
run-sync.shruntime sops-decrypts the ClickHouse reader password from homelab-talos trunk (clusters/homelab/apps/activity/secrets.enc.yaml, host age key) → CLICKHOUSE_* exported → scan runs telemetry-on. Every step guarded → missing key/repo/sops ⇒ telemetry-off (still writes a useful handoff+git snapshot). --input-type yaml is REQUIRED (mktemp has no .yaml ext)
Viewer
initiatives-viewer.service (systemd --user, workbench, serverMode-gated, Restart=on-failure) → run-viewer.sh → viewer.py (stdlib http.server). LIVE at http://192.168.50.250:8899/
Viewer bind
⚠ binds the workbench eth1 LAN IP 192.168.50.250 — NOT 192.168.50.94, which is a HOMELAB node (kube-apiserver/NodePorts/ClickHouse); binding .94 → OSError: Cannot assign requested address crash-loop (bit us in #140→#141), and a 127.0.0.1 smoke test will NOT catch it. See [[workbench-lan-ip]]. Internal work data — deliberately NOT wired to the public gateway
vllm-recap — homelab vLLM, ns promptver, svc vllm-recap:8000, served model recap = Qwen2.5-7B-Instruct-AWQ. Wired via the unit env (INITIATIVES_RECAP_ENABLED=1, RECAP_NAMESPACE=promptver, RECAP_SERVICE=svc/vllm-recap, RECAP_SERVICE_PORT=8000, RECAP_MODEL=recap) — recap.py's in-code defaults are PLACEHOLDERs
Router
route.py — route(signal,repo,limit) / rank_matches() / classify(). Reads initiatives.current; scoring single-sourced from the scan's best_title_match (word-equality, no stemming). Read-only, suggests-never-acts. Wired into repo-cos digests (#138) + mail-actions (#139, adds related_initiative col)
Assistant (Phase 1 agent)
PRIMARY /api/ask = a model-driven OpenClaw devpod (ns devpod-initiatives, svc initiatives-devpod:18789, openclaw/initiatives, DeepSeek V4 Pro via OpenRouter). The MODEL selects which deterministic skill-tool(s) to run (incl. MULTIPLE for compound Qs) — the tools are scripts/initiatives/skills/query.py (reuses assistant.py's run_tool/build_facts/sources_of), reached via _db.pydirect in-cluster mode (#156) with a least-priv initiatives_agent PG role (SELECT-only on initiatives.*). Viewer's agent_client.py proxies via kubectl port-forward + gateway token sha256("gw-"+HOOKS_TOKEN), streams the answer (SSE /api/ask/stream) and renders markdown; graceful FALLBACK to the deterministic regex assistant.py if the devpod is down. Every ask audit-logged to initiatives.assistant_log (intent=agent). Phase 2 (write/dispatch) stays deferred behind a structural server-side gate. Deploy/hardening → ~/.claude/skills/initiatives/reference/agent-devpod.md
# audit log (BOTH paths; intent=agent vs the regex intents):
$KC_HOMELAB
exec
"select id,intent,model,jsonb_array_length(sources) nsrc,latency_ms,left(question,40) from initiatives.assistant_log order by id desc limit 8;"
# FALLBACK deterministic assistant (used when the devpod is down; also the CLI):
'python3.withPackages(p:[p.psycopg2 p.requests])'
'python ~/workspace/devrc/scripts/initiatives/assistant.py "whats stalled and waiting on me"'
# force an out-of-band viewer refresh (same as the ↻ button):
echo
# ONE-TAP DISPATCH → a clawgate Task card (Phase 2a, origin/main 835fe0c). The VIEWER holds
# ~/.claude/clawgate.env; the in-cluster devpod does NOT. Two human gates: this creates a
# card; you still tap Dispatch in clawgate.
'Content-Type: application/json'
'{"repo":"<repo full path OR repo_name>","slug":"<slug>"}'
# 200 {ok,task_id} / 400 / 404 / 502
# dispatch.py mirrors repo-cos/clawgate.py: POST clawgate /api/tasks {directory(=label),body,repo?,tags} —
# model OMITTED → clawgate default deepseek. Confirm the card: GET {CLAWGATE_API_URL}/api/tasks (Bearer).
# Linked-task join, dispatch guard, fetch shape + what still leaks → ~/.claude/skills/initiatives/reference/clawgate-tasks.md
assistant_log
standalone tables
Validate a
migration by replaying it on a throwaway Postgres first
Repo-name tokens are STRIPPED before a live tmux pane → initiative match (2026-07-28).
best_title_match's title_overlap >= 2 gate was clearing on the REPO-NAME tokens alone
(civitai-manager → {civitai, manager}) — which every session AND initiative in the repo shares —
so a generic pane ("Continue civitai-manager development work") got badged ● live on an unrelated
card while that session did other work. Fix in match_tmux_to_initiatives:
ptoks -= set(text_tokens(os.path.basename(repo))) so a pane must share a DISTINGUISHING word;
a pane with none falls to live_unmatched (an honest miss — "a wrong tag costs more than a miss").
Momentum times from the last genuine USER-turn timestamp, NOT the transcript file mtime (PR #149).
Claude Code rewrites session .jsonl files in place (title/mode metadata) → mtime-driven momentum
read 18-day-idle sessions as "active ~47m ago". initiative-scan.py parses turns[].ts for the
last user turn; mtime is a rare fallback only (mirrors doc_touch_epoch).
Recap VRAM contention — vllm-recap shares a single 5080 (16.3GB, time-sliced ×4); to free VRAM,
vllm-joycaption + comfyui are scaled to replicas: 0 in homelab git. Scale them back for a
captioning/comfy pass — but VRAM then contends with recap. Recap is best-effort: model down → the
sync/store are unaffected and the card falls back to the deterministic summary.
Recap identity vs status (PR #154): two independently-cached fields — identity ("what it is",
from the handoff's durable head, hash-keyed on the HANDOFF so it's stable across prompt churn) +
status ("current", hash-keyed on recent activity). Two vllm-recap calls. Viewer renders identity
primary (fallback identity→recap→summary), status secondary.
_db.py importlib load — never add mail-actions/ to sys.path (its llm.py shadows
repo-cos's and breaks synthesis). route/assistant/sync all load it by explicit path.
Assistant is read-only by design — no write/dispatch/devpod/MCP; worst prompt-injection case is a
skewed but grounded answer, never an action. The Phase-2 write/dispatch path is DEFERRED behind a
structural server-side write-gate (NOT the voluntary agent_checkpoint) + a least-privilege DB role.
