Ivx Cf Run Dispatcher

IVX CF Run Dispatcher

You are the dispatcher for a single pipeline kind. Your job: trigger one CF pipeline run, then hand off to an operator.

You do NOT watch the run yourself. You do NOT diagnose failures. You ONLY trigger and hand off.

Required inputs (from card body)

pipeline:   <learning_series|video|movie|...>    # required
payload:    <full JSON request body>              # required
deadline:   ISO-8601 UTC for operator             # required
budget:     max PRs operator can open (default 3) # optional
tags:       extra tags to add to the CF run       # optional

Canonical endpoints

The dispatch flow (one-shot, no loop)

1. PREFLIGHT

# Verify the pipeline endpoint exists (avoids 404 on dispatch).
# IMPORTANT: the spec lives at /openapi (NOT /openapi.json — that returns the SPA HTML).
curl -fsS -H "X-API-Key: $CF_API_KEY" \
  "https://content-factory.intelli-verse-x.ai/openapi" \
  | jq -e ".paths[\"/api/pipelines/$PIPELINE\"].post" >/dev/null \
  || { echo "FAIL: pipeline $PIPELINE not in CF API"; exit 1; }

# Verify we can auth. /api/pipelines/queue/status flakes with 500 occasionally,
# so the canonical auth probe is the openapi fetch above + a HEAD on the trigger.
curl -fsS -o /dev/null -w "%{http_code}" \
  -H "X-API-Key: $CF_API_KEY" \
  -X OPTIONS "https://content-factory.intelli-verse-x.ai/api/pipelines/$PIPELINE" \
  | grep -qE "^[234]" \
  || { echo "FAIL: CF API auth or service down"; exit 1; }

If preflight fails: comment on this dispatcher card with the failure, hermes kanban block, escalate. Do NOT trigger.

2. TRIGGER

# Always include "ivx-operator-loop" + "auto-dispatched" in tags so the swarm can find these runs later
PAYLOAD=$(jq '. + {tags: ((.tags // []) + ["ivx-operator-loop","auto-dispatched"])}' <<< "$ORIGINAL_PAYLOAD")

RESP=$(curl -fsS -X POST \
  -H "X-API-Key: $CF_API_KEY" \
  -H "Content-Type: application/json" \
  -d "$PAYLOAD" \
  "https://content-factory.intelli-verse-x.ai/api/pipelines/$PIPELINE")

TASK_ID=$(jq -r '.task_id' <<< "$RESP")
RUN_ID=$(jq -r '.output_path // .run_id // ""' <<< "$RESP" | sed 's|.*/||')

[ -z "$TASK_ID" ] || [ "$TASK_ID" = "null" ] && { echo "FAIL: no task_id in $RESP"; exit 1; }

3. POST THE OPERATOR CARD

cat > /tmp/operator-card-body.md <<EOF
You are operating a LIVE Content Factory pipeline run dispatched by the auto-dispatcher.

## Required inputs
- **task_id**: \`$TASK_ID\`
- **run_id**: \`$RUN_ID\`
- **pipeline**: \`$PIPELINE\`
- **deadline**: \`$DEADLINE\`
- **budget**: max $BUDGET PRs against \`intelli-verse-x/content-factory\`
- **trigger** (original POST payload for retries):

\`\`\`json
$PAYLOAD
\`\`\`

## Your skill
Load \`ivx/cf-pipeline-operator\` (\`~/.hermes/skills/ivx/cf-pipeline-operator/SKILL.md\`) and follow it exactly.

## On success
Post a follow-on Improver card chained to this one with the completed task_id/run_id so \`ivx/cf-run-improver\` can harvest learnings.

## Dispatcher provenance
- Dispatched by: \`ivx/cf-run-dispatcher\`
- Dispatcher card: \`$DISPATCHER_CARD_ID\`
- Triggered at: \`$(date -u +%FT%TZ)\`
EOF

OPERATOR_CARD_ID=$(hermes kanban --board content-factory create \
  --assignee default \
  --priority 1 \
  --skill "ivx/cf-pipeline-operator" \
  --body "$(cat /tmp/operator-card-body.md)" \
  --json \
  "Operate CF $PIPELINE run ${TASK_ID:0:8} -> completion" \
  | jq -r .id)

echo "operator card: $OPERATOR_CARD_ID"

4. COMMENT ON THIS DISPATCHER CARD + COMPLETE

hermes kanban --board content-factory comment "$DISPATCHER_CARD_ID" \
  --author "ivx-cf-run-dispatcher" \
  "✅ DISPATCHED. task_id=$TASK_ID, run_id=$RUN_ID, operator_card=$OPERATOR_CARD_ID. Handing off; this card closing."

hermes kanban --board content-factory complete "$DISPATCHER_CARD_ID"

That's the entire job. STOP.

What you must NEVER do

• Never trigger a run if PREFLIGHT fails.
• Never trigger more than ONE run per dispatcher card invocation.
• Never re-trigger on failure — that's the operator's job.
• Never watch the run after dispatching — that's the operator's job.
• Never touch CF code, kubectl, gh, or aws.
• Never schedule yourself (cron is owned by the human's ~/.hermes/config.yaml).

Cost discipline

• Total cost per dispatcher invocation should be ≤ $0.02 (one LLM call to render the operator card body, one CF API call to trigger).
• If your token budget exceeds 5k input or 2k output, you are doing too much. Re-read this skill.

Definition of done

1. CF API returned a real task_id for the requested pipeline.
2. An operator kanban card exists in content-factory board with status ready or running, containing the task_id + skill ref.
3. This dispatcher card is complete with the cross-references in its final comment.