Batch-Driver Skill

Process steps in order. Do not skip ahead.

This skill orchestrates the A1 pipeline across a CSV of companies. The per-company pipeline (discovery → selection → build-and-evaluate) is itself implemented by three other skills; this skill loops over them with a shared run timestamp so the batch is diff-able as a unit. Designed for pre-conference batch runs (Snyk attendee list, ~100 leads) where manual per-company target selection isn't feasible — selection auto-picks the top-confidence candidate and notes the auto-pick in selection.json for traceability.

Step 1 — Parse the CSV and Branch on Yield

Input: a CSV path. Expected columns (case-insensitive header): company (required) and domain (optional). Empty cells, duplicate rows, and rows with non-printable characters are dropped with a warning.

Run skills/batch-driver/scripts/parse-attendees.py <csv-path>. The script emits a normalized JSON list of {slug, company_name, domain} records to stdout and a {ok, dropped_rows, kept_rows} summary to stderr.

Branch by parse outcome:

• Script exits 0 with kept_rows >= 1: proceed to Step 2 with the parsed list cached in working memory.
• Script exits non-zero (or kept_rows == 0): the CSV has no usable companies — every row was dropped (empty, duplicate, or non-printable). Finish here. Output an explanation that surfaces:
  • The CSV path that was processed
  • The stderr summary verbatim ({ok: false, kept_rows: 0, dropped_rows: N, reasons: {...}})
  • An actionable next step (re-export the CSV, verify the company column header is correct, check the source export job) Do NOT proceed to Step 2. Do NOT create a run directory under runs/. Do NOT write a batch-manifest.json. Do NOT invent company names to keep the batch going — fabricating data to satisfy a downstream pipeline is the exact failure mode this gate exists to prevent. Do NOT invoke discovery-produce or any other per-company skill.

Why this gate is load-bearing

A pre-0.1.2 version of this skill had the parse-outcome branch implicit ("Exit non-zero if no usable rows remain"), and the agent ignored it — given an empty CSV, the agent fabricated five companies, generated a run directory, wrote a batch-manifest.json, and ran the full per-company pipeline against the invented data. The published-time eval batch-empty-csv regressed from baseline 0.53 to with-skill 0.10 on exactly this failure mode. The branch is the first action in the step because the agent must commit to halting before any per-company tooling executes.

Step 2 — Initialize the Batch Run Directory

Generate a shared DISCOVERY_RUN_TS for the batch (UTC timestamp with a human-readable suffix, e.g., 2026-05-13T-snyk-batch). Export it as an environment variable so every per-company script invocation in Steps 3–4 writes to the same runs/<DISCOVERY_RUN_TS>/ parent directory.

Write runs/<DISCOVERY_RUN_TS>/batch-manifest.json with the parsed company list and the run timestamp. The manifest is the index every downstream step (summary rendering, post-batch review) reads — do not rely on filesystem-listing alone.

Proceed immediately to Step 3.

Step 3 — Run Per-Company Pipeline

For each company in the manifest, in sequence:

1. Invoke Skill(skill: "discovery-produce") with the company name. The skill writes runs/<DISCOVERY_RUN_TS>/<slug>/discovery.json and returns the path.
2. Branch on the discovery verdict:
   • BUILD → run skills/batch-driver/scripts/auto-select.py <discovery-path> to write a selection.json auto-picking the top-confidence target (no human gate in batch mode), then invoke Skill(skill: "build-and-evaluate") for the rest of the pipeline.
   • SKIP → record the skip reason in the batch index; do not invoke select-target or build-and-evaluate for this company. Move on.
   • AMBIGUOUS → record the sub-brand list; flag for manual re-invocation with parent/sub-brand. Move on.
3. Capture per-company outcome: {slug, verdict, target_id?, aggregate_lift?, report_path?, error?} and append to an in-memory batch results array.

Continue past per-company failures rather than aborting the batch — a single company failing should not block the other 99. If discovery returns a non-JSON or malformed result, log the error and move on.

Proceed immediately to Step 4 after the last company.

Step 4 — Render Batch Summary

Run skills/batch-driver/scripts/summarize-batch.py runs/<DISCOVERY_RUN_TS> to emit runs/<DISCOVERY_RUN_TS>/index.md. The script consumes the batch-manifest and walks the per-company subdirectories to collect verdicts, lift scores, and report paths. Schema of the rendered summary:

• Header: batch ID, run timestamp, total companies
• Verdict counts: BUILD / SKIP / AMBIGUOUS / error
• Per-company table: slug, verdict, top-target title (BUILDs only), aggregate lift, link to report
• Aggregate lift across BUILDs
• Companies flagged for follow-up: SKIPs (review search trail), AMBIGUOUS (re-invoke), errors (re-run individually)

Output the absolute path of index.md and finish here.