schedule-ads

Files (3)

• SKILL.md
• config.create.example.yaml
• config.example.yaml

SKILL.md (first lines)

> **${var}** selects the flow. Empty/unset = **schedule** (launch ads into existing ad sets). `create` = **create-campaign** (provision Meta campaigns + ad sets). Both are config-driven, PAUSED-by-default, and make the AdManage API calls **in-run** via `./secretcurl` (the `{ADMANAGE_API_KEY}` placeholder keeps the key off the command line), behind fail-closed spend guardrails.

Reads a declarative config, computes what to do, and makes the AdManage.ai API calls **in-run** via `./secretcurl`. The calls are an irreversible outbound side-effect (real ad spend), so they are each branch's **final** actions and run only behind the guardrails below (PAUSED-by-default, `dailySpendCap` circuit-breaker, dry-run). `ADMANAGE_API_KEY` is injected in-run via this skill's `requires:` — always write it as the `{ADMANAGE_API_KEY}` placeholder, never a bare `$ADMANAGE_API_KEY` (the Bash permission layer refuses that).

## Preamble (both branches)

1. Read `memory/MEMORY.md` for context. Read the last ~3 days of `memory/logs/` for recent launch / provisioning activity — don't re-report a signal already logged.
2. Parse `${var}`:
   - empty / unset → run the **Schedule branch** below.
   - `create` → run the **Create branch** below.
   - anything else → log `SCHEDULE_ADS_UNKNOWN_SELECTOR: <value>` and exit cleanly (no notify).
3. Both branches spend real money on ad platforms. The shared safety posture (see each branch) is: PAUSED by default, config-only (never invent campaigns/creative/targeting), dry-run available, and exit silently when there's nothing to do.

---

# Schedule branch (default — empty `${var}`)

Reads `skills/schedule-ads/config.yaml`, picks schedule entries matching today, and launches those ads **in-run** via AdManage.ai (`POST /v1/launch` through `./secretcurl`), behind the spend guardrails below.

## Safety defaults (schedule)

This branch **spends real money on ad platforms**. Guardrails, in priority order:

1. **PAUSED by default.** Every launch request sets the entity to PAUSED. The operator has to resume manually in the AdManage dashboard before spend starts. `launchPaused: false` in config is the explicit opt-out.
2. **Daily spend cap.** Before launching, the branch checks `GET /v1/spend/daily` for today. If spend ≥ `dailySpendCap` in the config, all launches are skipped and a warning is notified. If the spend figure can't be verified (malformed / empty response), **fail closed** — skip and notify, don't launch. This is a circuit breaker, not a budget enforcer — platform budgets still apply.
3. **Dry-run mode.** If `DRY_RUN=true` in env or `dryRun: true` in config, the branch builds the payloads, writes them to `.pending-admanage/dryrun/`, notifies what *would* launch, and exits without calling the API.
4. **Config-only.** The branch does not invent campaigns, creative, or targeting. If there's no schedule for today, it exits cleanly with no API calls.
5. **Single source of truth.** All ads/campaigns/targeting live in `config.yaml`. The branch never generates new creative on the fly.

## Network note (schedule)

Launching ads is an irreversible outbound side-effect (real ad spend), so it is the branch's **final** action and runs only after the guardrails above pass:

- Auth'd calls go through `./secretcurl` with the `{ADMANAGE_API_KEY}` placeholder — never a bare `$ADMANAGE_API_KEY` (the Bash permission layer refuses that). `ADMANAGE_API_KEY` is injected in-run via `requires:`.
- The branch checks the daily spend cap (`GET /v1/spend/daily`), then per batch calls `POST /v1/launch`, polls `GET /v1/batch-status/{id}` to a terminal state, and reports via `./notify`.
- If `ADMANAGE_API_KEY` is unset, or the launch/spend call fails, skip the launch and notify — do not retry blindly. There is no deferred postprocess fallback.

## Steps (schedule)

1. **Load config.** Read `skills/schedule-ads/config.yaml`. If the file doesn't exist, log `SCHEDULE_ADS_NOT_CONFIGURED` and exit cleanly (no notify, no error). The example template lives next to this file as `config.example.yaml`.

2. **Validate config shape.** Required top-level keys: `defaults` (with `adAccountId`, `workspaceId`, `page`), and `schedules` (array). If either is missing, file an issue in `memory/issues/` per the CLAUDE.md issue tracker convention, notify once, and exit.

3. **Pick today's schedule entries.** For each entry in `schedules`, match against today's date:
   - `when.everyDay: true` → always matches.
   - `when.dayOfWeek: monday` (or any weekday name, lowercase) → matches if today is that weekday (UTC).
   - `when.date: "2026-04-25"` → matches only on that exact date.
   - `when.dates: ["2026-04-25", "2026-05-02"]` → matches if today is in the list.
   - `when.cron: "0 8 * * 1"` → (advanced) matches if today satisfies the cron. Optional — skip if it's too much parsing effort.

   If no entries match today, log `SCHEDULE_ADS_NOTHING_TODAY` and exit cleanly (no notify).

4. **Build launch payloads.** For each matching schedule entry, construct the AdManage `POST /v1/launch` body:
   ```json
   {
     "ads": [
       {
         "adName": "<templated from ad.adName, {date} replaced>",
         "adAccountId": "<from defaults or entry override>",
         "workspaceId": "<from defaults or entry override>",
         "title": "<from ad>",
         "description": "<from ad>",
         "cta": "<from ad or defaults.cta>",
         "link": "<from ad>",
         "page": "<from defaults>",
         "insta": "<from defaults, Meta only>",
         "adSets": [ { "value": "<id>", "label": "<name>" } ],
         "media": [ { "url": "<media url>" } ],
         "status": "PAUSED"
       }
     ]
   }
   ```
   Enforce `status: PAUSED` on every ad unless `defaults.launchPaused` is explicitly `false`. Never strip it silently.

   Template substitutions inside string fields:
   - `{date}` → today's ISO date (YYYY-MM-DD)
   - `{dateHuman}` → "April 21, 2026" style

5. **Pre-flight validation.** For each payload:
   - `media[*].url` must be an absolute `https://` URL. Reject entries with local paths or obviously broken URLs.
   - `adSets[*].value` must be a non-empty string. If missing, skip the entry with a warning in the log.
   - For Meta entries (`adAccountId` starts with `act_`): `page` and `insta` must be set. TikTok/Snapchat/etc. have their own requirements — don't block on Meta-specific fields for other platforms.
   - `title` and `description` must be non-empty.

   Drop invalid entries, keep going. Log which ones were skipped and why.

6. **Handle dry-run.** If `DRY_RUN=true` or `config.dryRun: true`:
   - Write payloads to `.pending-admanage/dryrun/{schedule-name}-{timestamp}.json`.
   - Notify a preview (see step 9) but with `[DRY RUN]` prefix.
   - Skip step 7.
   - This mode exists for the operator to sanity-check before arming real launches.

7. **Launch in-run.** This is the branch's final action — spends real money, so run the guardrails first. Only `./secretcurl`, `jq`, `date`, `echo`, `mkdir`, `grep`, `python3`, and the `Write` tool are available.

   a. **Config check.** `[ -n "${ADMANAGE_API_KEY:+x}" ]` (the `${VAR:+x}` form — a bare `$ADMANAGE_API_KEY` trips the secret-expansion analyzer and reads as unset). If unset → notify "ads computed but ADMANAGE_API_KEY missing — nothing launched" and stop.

   b. **Daily spend circuit-breaker (once).** Take the strictest `dailySpendCap` (`CAP`) across today's payloads. If set, read today's spend and **fail closed** unless it's a clean number *below* the cap:
      ```bash
      SPEND=$(./secretcurl -sS --max-time 30 -H "Authorization: Bearer {ADMANAGE_API_KEY}" \
        "https://api.admanage.ai/v1/spend/daily?startDate=$TODAY&endDate=$TODAY" | jq -r '.metadata.totalSpend // ""')
      echo "$SPEND" | grep -qE '^[0-9]+(\.[0-9]+)?$' || { echo "spend unverifiable — fail closed"; exit 0; }
      echo "$CAP"   | grep -qE '^[0-9]+(\.[0-9]+)?$' || { echo "dailySpendCap not numeric — fail closed"; exit 0; }
      python3 -c "import sys; sys.exit(0 if float(sys.argv[1])>=float(sys.argv[2]) else 1)" "$SPEND" "$CAP" \
        && { echo "daily spend cap tripped (today=$SPEND cap=$CAP) — launching nothing"; exit 0; }
      ```
   c. **Per batch: launch, then poll.** For each payload `{ ads: [ ... ] }`:
      ```bash
      RESP=$(./secretcurl -sS --max-time 60 -w 'http=%{http_code}\n' -X POST "https://api.admanage.ai/v1/launch" \
        -H "Authorization: Bearer {ADMANAGE_API_KEY}" -H "Content-Type: application/json" -d "$PAYLOAD")
      # success => .success==true and .adBatchId set; else record FAILED (.message/.error) and continue.
      # Poll GET /v1/batch-status/$BATCH_ID (~90s, 5s interval) until .summaryStatus is success|error.
      ```
      Record each batch's outcome (ok / error / still-running-after-timeout) for the notify. Ads launch **PAUSED** (the payload sets it) unless `launchPaused: false`.

8. **Write artifact to `output/.chains/schedule-ads.md`** so downstream chain consumers can read what was queued. Format:
   ```markdown
   # Schedule Ads — ${today}

From aeonfun/aeon, MIT.

More from aeonfun/aeon