Pusher

Google Ads — Pusher (safety-critical)

Get a change into the account without ever spending money unexpectedly. Two units of work, auto-detected by the input's kind:

• campaign-spec.json (from builder-*, kind absent) — creates a NEW campaign. Schema: ${CLAUDE_PLUGIN_ROOT}/references/campaign-spec.md.
• change-set.json (from optimizer/routine, kind: "change-set") — EDITS to live campaigns (negatives, geo/product exclusions, budget/target moves, pauses, extensions, signals). Schema: ${CLAUDE_PLUGIN_ROOT}/references/change-set.md.

Read the JSON, never the .xlsx. The safety invariants below apply to both paths.

Operating rules / safety invariants (non-negotiable)

• Validate before anything (STEP 1). A spec that fails validation never reaches an export or mutation.
• Campaigns are created PAUSED. The human enables them in the UI after reviewing. Never create enabled.
• Human approval gate (STEP 3): print exactly what will happen and require an explicit "yes".
• Spend-cap guard: block if daily_budget exceeds the cap (from context or asked).
• Mode honesty: if api_write is false, do NOT claim auto-create — emit an importable file + guided UI steps. State precisely what is automated vs manual.

Model dispatch (run cheap, decide expensive) — see ${CLAUDE_PLUGIN_ROOT}/references/model-tier-dispatch.md

• Scout (haiku) — STEP 1 validate_spec.py + spec_to_paste.py; STEP 2 generators (spec_to_editor_csv.py, spec_to_pmax_checklist.py); STEP 1c/2c validate_changeset.py + changeset_to_actions.py; per-URL 200 checks. These are deterministic script runs — push them all down.
• Judge (main session) — STEP 3 the approval gate (print the plan, require an explicit "yes"), the spend-cap call, Mode B mutation oversight, and any FAIL interpretation. Never delegate the gate. A safety decision is judgment by definition.
• (No Routine tier here — validation and generation are single-shot Scout runs; the only judgment is the gate.)

STEP 0 — Load + detect kind

Read the input JSON and account-context.yaml (for google_ads.api_write and any spend cap). Confirm the target customer_id matches. Branch on kind: "change-set" → the change path (STEP 1c/2c below); otherwise treat it as a campaign-spec.json → the build path (STEP 1/2).

STEP 1c — Validate a change-set (dry-run gate)

python ${CLAUDE_PLUGIN_ROOT}/skills/pusher/scripts/validate_changeset.py <change-set.json> \
       --context ./account-context.yaml --max-budget <cap>

It enforces the Scaling Ladder + account guardrails: cooldown gate (BLOCKS a budget/target move on a within-cooldown campaign), budget step ≤ +10% + spend cap, tROAS step ≤ 0.3x + ≥15 conv/wk floor, no direction reversal inside the cooldown, PMax-scales-down-by-pausing (not budget cut), no broad negatives, never block a brand term (word-boundary, from context brand_terms), and a reason on every action. FAIL → stop and report (e.g. "3 scale moves blocked by cooldown — defer them"). On PASS, set provenance.verified = true, then go to STEP 2c.

STEP 2c — Render the change into operator actions

python ${CLAUDE_PLUGIN_ROOT}/skills/pusher/scripts/changeset_to_actions.py <change-set.json> --output-dir .

Emits CHANGE-PLAN.md (ordered, checkboxed, each step with its data reason + UI path + discipline note + $/mo), negatives-editor.csv (Search/Branded campaign negatives → Editor import), and negatives-paste.txt (wrapped negatives for shared lists / PMax campaign negatives). Then the approval gate (STEP 3) and report (STEP 5). With api_write: true, apply the actions via the API instead (still gated; negatives/exclusions/ budget/target/pause are mutations on existing entities — nothing is created enabled).

STEP 1 — Validate (dry-run gate)

Run the validator:

python ${CLAUDE_PLUGIN_ROOT}/skills/pusher/scripts/validate_spec.py <campaign-spec.json> --max-budget <cap>

It checks: char limits + counts per asset group (Google-accurate counting — NFC-normalized, CJK/full- width = 2, trailing whitespace + hidden/non-breaking chars flagged; see ${CLAUDE_PLUGIN_ROOT}/references/google-ads-formatting.md), listing-group Everything-Else nodes + no cross-AG product overlap, status = paused, conversion_goals.scope = campaign, brand_exclusion for PMax, negative match types, and daily_budget ≤ cap. FAIL → stop and report. On PASS, set provenance.verified = true.

Also emit ready-to-copy keyword/negative/search-theme blocks (already wrapped [exact] / "phrase" / broad so the operator pastes them straight in — no re-formatting): python ${CLAUDE_PLUGIN_ROOT}/skills/pusher/scripts/spec_to_paste.py <campaign-spec.json> --output paste.txt

STEP 2 — Choose the push path (honest, per ${CLAUDE_PLUGIN_ROOT}/references/campaign-spec.md pushability matrix)

Decide by api_write AND campaign_type:

• Mode A — Guided build (default; api_write: false)
  • Search / Branded Search → generate a Google Ads Editor import CSV (campaigns, ad groups, keywords + match types, RSAs, negatives) with: python ${CLAUDE_PLUGIN_ROOT}/skills/pusher/scripts/spec_to_editor_csv.py <campaign-spec.json> --output editor-import.csv Then: Google Ads Editor → Account → Import → From file → REVIEW → post. Editor handles these well.
  • Performance Max / Demand Gen → PMax is mostly API-or-UI. Emit a precise, ordered UI build checklist with all copy/values/URLs paste-ready and Google-accurate char counts: python ${CLAUDE_PLUGIN_ROOT}/skills/pusher/scripts/spec_to_pmax_checklist.py <campaign-spec.json> --output checklist.md It renders campaign settings, per-asset-group listing-group include/exclude steps, paste blocks (headlines/long/descriptions/search themes), asset-upload list, sitelinks, audience signals, and wrapped campaign negatives — each as a checkbox. The operator follows it click-by-click in the UI.
• Mode B — API mutate (api_write: true)
  • Create via the Google Ads API: campaign + budget + bidding, asset groups, listing-group filters, asset-group signals, text assets; upload image/video assets first, then attach. Still PAUSED, still behind the approval gate. (Requires a write-capable connection — most MCPs are read-only.)

STEP 3 — Approval gate

Print a concise plan and wait for explicit "yes" (no yes → stop):

• Build path: campaign name, type, daily budget (vs cap), bidding, # asset/ad groups, brand exclusion on/off, and — for Mode A — what's auto vs manual.
• Change path: the action count + total $/mo from CHANGE-PLAN.md, the campaigns touched, what mutates (negatives / exclusions / budget±/target±/pauses), and explicitly any moves HELD by cooldown (these are NOT applied). Confirm nothing enables a paused campaign and no budget move exceeds +10% or the cap.

STEP 4 — Execute

Mode A: write the Editor file + the UI checklist to the working directory. Mode B: run the mutations (create-paused), capturing returned resource names; on any error, stop and report what was/wasn't created.

STEP 5 — Report

• What was created/exported, what remains manual, and the exact next action ("import this file", or "complete these UI steps", then "review and enable").
• Reminder: campaign is PAUSED — enable only after review. Hand off monitoring to tracker.

To build / refine later

• Editor CSV generator for Search/Branded (scripts/spec_to_editor_csv.py). Done.
• Ready-to-copy keyword/negative/theme blocks (scripts/spec_to_paste.py). Done.
• PMax guided-UI checklist generator (scripts/spec_to_pmax_checklist.py). Done.
• Change-set path: validate_changeset.py (safety gates) + changeset_to_actions.py (CHANGE-PLAN.md + negatives CSV/paste). Contract: references/change-set.md, template: templates/change-set.json. Done.
• API mutate path (Mode B) — deferred (few accounts have write API access right now). Applies to both the build path AND the change path (negatives/exclusions/budget/target/pause mutations).
• Re-check current Google Ads Editor PMax import capability (it changes over time).