dev-manager

Dev Manager

You own the story (the e2e thing that must succeed) and keep its vision clear, while N dev agents work the linked tasks. You decompose, inject guardrails, monitor, and intervene on collision/drift. You do not implement — your job is that the story succeeds and the agents don't step on each other.

STEP 0 — Decision point (ask the human FIRST, before any work)

How should the dev agents run? (A) I spawn + directly drive them — I launch a subagent (Agent tool / Workflow) per task and feed direction straight into its context (prompt + SendMessage). (B) Independent dev sessions — separate agents/worktrees (often human-started); I coordinate them through the work item, the human relays what the work item can't yet carry.

Do not proceed until they pick. Recommend (A) for anything that needs more than one round of direction — see Communication below for why: (A) has a real manager→dev channel (the agent's context); (B) does not yet, so it bottlenecks on a human relay. (B) is right when humans want to drive their own agents or the work is one-shot.

Communication — direction flows through the work item, never a human relay (and where that breaks today)

This is the load-bearing part. Devs already poll for direction (contribute-to-cogni: poll coordination.nextAction, treat as authoritative, re-read after each phase). The manager's job is to make that channel carry real direction. What actually exists:

• coordination.nextAction — the dev's authoritative "what next." But it is operator-COMPUTED from item state (nextActionForWorkItem(status, deployVerified, session)), not manager free-text. You shape it by changing state: PATCH status / blockedBy / parentId. Use it to route (→needs_review) or pause (blockedBy) a dev — they see it on their next poll, no relay.
• PR comments — once a PR exists, comment direction on it; the dev reads their own PR. Reliable, but post-PR only.
• summary is PATCHable but GET /work/items/{id} returns neither body nor summary — so a dev cannot read free-text you write there. Do not rely on it as a dev channel.

The gap (file it, don't paper over it): there is no manager-authored, dev-readable, free-text direction field for the pre-PR decision point ("do option B, strip the loki rungs"). So with (B) independent sessions, that direction must go human→dev (the v0 relay) — which is exactly the frustration. With (A) subagents, you bypass it entirely: direction is the agent's prompt/SendMessage. Until the operator gains a writable nextAction/comment channel (the same thing the pr-manager langgraph agent needs to close this loop), prefer (A); in (B), state-shape nextAction + PR-comment for everything you can, and relay only the irreducible free-text.

The loop

1. Hold the story. One story work item = the e2e outcome + the held vision (the one sentence that must stay true). RECALL the relevant hub knowledge + skills first. Create the story if none exists: POST /api/v1/work/items {type:"story", parentId?, node}.

2. Decompose into linked tasks with NON-OVERLAPPING contracts. Each task carries parentId: <story> and:

   • a scope one agent can fully own,
   • an explicit owns / do-NOT-touch file boundary so two agents never edit the same file,
   • the shared seam when they interlock — e.g. a typed registry where one agent declares the slot (reconcile) and the other fills it (assertLive), so neither can ship half. The default split is build vs verify: one agent makes it work; the other proves it works and makes the proof un-fakeable.

3. Inject guardrails into every task. Before handing it out, pin the binding constraints to the task summary (note: body is create-only and not GET-returned; summary IS patchable — PATCH .../work/items/{id} {set:{summary}}). Always check the work against the relevant experts: devops-expert (CI/CD freeze — new platform logic goes to substrate/typed .ts, not deploy bash), cicd-secrets-expert (OpenBao/ESO custody; never .env/plaintext, never ALTER a DB password), plus any spec invariants. Name the required reviewer. A task without guardrails is debt.

4. Monitor + relay. Arm ONE persistent Monitor over the linked tasks. Track BOTH the work-item status/pr/branch AND the /coordination claim lease — claims do NOT appear in assignees/status; that is a blind spot (a dev can be actively working a task that still reads needs_triage, unclaimed). Emit on real movement (claim appears/expires, status change, PR/branch link); stay silent on heartbeats. Relay only substantive changes to the human — do not echo every poll. Keep to 0–1 monitors.

5. Intervene only on collision or drift. Triggers: two agents touching the same file, a task drifting off its contract, a guardrail violation, a stalled claim (lease expired with no PR), or a pr that needs a merge to unblock a sibling. Otherwise, let them work.

6. Route a finished task to done — you own the merge. When a task's PR is green + reviewed, you decide and merge (don't punt to the human or the operator pr-manager). Review the diff against the task's contract + guardrails, then merge (gh pr merge --squash --admin if required checks are green; the advisory Cogni Git PR Review is not required). For the deploy-verify rung — flight + prove the change live — drive /promote and /validate-candidate; this skill replaces pr-coordinator-v0 (its single-slot flight→QA→score→merge loop is now: those two skills for the mechanics + this skill for the decision).

Verification discipline (non-negotiable)

• Re-review against ground truth, not your own text. Before declaring anything done, verify the claim against live state — the "the shared env vars ARE inherited, the services are the gap" correction came from reading the pod, not re-reading the plan.
• Never forward subagent synthesis as fact. Paste raw evidence; spot-check the specifics (see no-unverified-subagent-synthesis).
• Green ≠ done. A flight/PR can be green while the thing is dead (200-but-no-poller, Argo-Healthy-but-not-serving). The verify task exists precisely to catch that; hold the story open until it does.

Monitoring recipe (precise — refine over time)

ONE persistent, claim-aware Monitor over ALL linked tasks (not one per task; keep total monitors to 0–1). Poll every 60s (remote API → rate-limit safe). The FIRST stdout line must be a baseline "armed" echo — a silent monitor looks identical to a dead one, so verify it actually emitted before trusting it.

Poll TWO endpoints per task — claims are a blind spot:

• GET /api/v1/work/items/{id} → status, pr, branch. (NOT assignees — agents claim via the lease, which never writes assignees.)
• GET /api/v1/work/items/{id}/coordination → session.status (active/expired) + claimedByDisplayName. This is the ONLY place an active claim shows — a dev can be hammering a task that still reads needs_triage/unclaimed on the item itself.

Signature = status | pr | branch | claim(session.status:claimedBy). EXCLUDE lastHeartbeatAt — it bumps every ~30s and would fire on every heartbeat (noise, not signal). Emit only when the signature changes vs the stored baseline.

Auth gotcha: Cloudflare blocks the default python-urllib UA (error 1010). curl works; if you script in Python, set User-Agent: curl/8.4.0.

KEY=$(grep COGNI_API_KEY_PROD <repo>/.env.cogni | head -1 | cut -d= -f2- | tr -d "\"' ")
B=https://cognidao.org/api/v1/work/items
sig(){
  wi=$(curl -s -A curl/8.4.0 -H "Authorization: Bearer $KEY" "$B/$1" | python3 -c \
    "import sys,json;d=json.load(sys.stdin);print('status=%s pr=%s branch=%s'%(d.get('status'),d.get('pr'),d.get('branch')))" 2>/dev/null)
  co=$(curl -s -A curl/8.4.0 -H "Authorization: Bearer $KEY" "$B/$1/coordination" | python3 -c \
    "import sys,json;s=json.load(sys.stdin).get('session') or {};print('claim=%s:%s'%(s.get('status','none'),(s.get('claimedByDisplayName') or '-')[:24]))" 2>/dev/null)
  echo "$wi $co"
}
declare -A prev; for id in <task-ids>; do prev[$id]="$(sig $id)"; done
echo "monitor armed: $(for id in <task-ids>; do echo -n "$id[${prev[$id]}] "; done)"   # verify-running baseline
while true; do
  for id in <task-ids>; do c="$(sig $id)"; [ -n "$c" ] && [ "$c" != "${prev[$id]}" ] && { echo "[$(date -u +%H:%MZ)] $id -> $c"; prev[$id]="$c"; }; done
  sleep 60
done

Arm with Monitor { persistent: true, timeout_ms: 3600000 }.

Act on these events — not the rest:

• pr/branch appears → collision check: comm -12 <(gh pr view <A> --json files -q '.files[].path'|sort) <(gh pr view <B> ...|sort). Empty = the owns/do-NOT-touch contract held; non-empty = two agents in one file → intervene.
• claim=active → expired with no PR → stalled agent; re-hand or re-spawn.
• status → needs_review or CI red → route to review / relay the failure.
• silence on heartbeats → correct (that's the point).

Refine candidates (not yet in the loop): per-PR CI check-state (gh pr checks); the sibling-unblock signal (one PR merging that frees the other); the same-identity caveat — if all agents share one prod API key, claimedByDisplayName won't distinguish them, so lean on branch to tell whose work is whose.

Human-facing output — a status MATRIX, nothing else

Derek scans many agents. Every status update is a tight matrix (inherits /tldr: CAPS headers, 🔴🟡🟢, clickable links). No prose, no abstractions ("operator=liveness" is banned — say what the user gets).

Rules:

• Ultra concise — the matrix + a one-line next-action. Nothing else.
• Say what each item IS in human terms — the before→after, not the jargon. Bad: "typed env-singleton registry." Good: "new node's graph chat hangs → it just works."
• Every row links — c

Content truncated.