auto-skill-discovery

Run the A1 pipeline (produce-mode discovery → auto-selection → build-and-evaluate) across a CSV of company names. Drives the full pipeline unattended for booth-scale runs (~100 leads), sharing a single DISCOVERY_RUN_TS so all per-company outputs land in one diffable run directory. After completion, emits a per-batch `index.md` summarizing outcomes (BUILD/SKIP/AMBIGUOUS counts, aggregate lift across BUILDs, links to per-company reports). Trigger phrases — "run the batch driver on a CSV", "process the attendee list", "batch-pipeline the snyk leads", "A1 batch run on a CSV of company names".

Steps 4–10 of the workflow — takes a selection.json produced by `select-target` and runs the full skill-build + eval + report pipeline. Scaffolds a skill via `tessl skill new`, authors its body for the selected target, generates eval scenarios via `tessl scenario generate`, audits them for bleeding/leaking, runs `tessl skill review --threshold 85` as a quality gate, runs `tessl eval run --variant without-context --variant with-context` to get baseline + with-skill scores in one call, parses the result, computes per-scenario lift, analyzes gaps, and writes a markdown report alongside the discovery / selection artifacts. End-to-end output for the A1 MVP cell of the operating-modes 2x2 in `SPEC.md`. Trigger phrases — "build the skill", "run phases 4-10", "build-and-evaluate", "generate and evaluate the skill for a given company slug".

Triage a raw list of company names (conference attendees, prospect lists, sales lists) before running the discovery pipeline. Classifies each company into seven buckets and returns a focused candidate list of companies whose discovery is likely to yield a usable skill target. Use whenever the user provides a list of more than a handful of company names and needs to decide which deserve discovery effort. Trigger phrases — "filter this list", "triage these companies", "pre-discovery filter", "narrow the conference list".

Produce-mode automated source discovery for a single company (steps 1–2 of the workflow, A1 cell of the operating-modes 2x2 in `SPEC.md`). Takes a company name (optionally `parent/sub-brand` for MEGA_CORP cases that have been pre-scoped) and produces a discovery JSON conforming to `discovery-output-contract.md` — schema_version 3, mode=produce. Targets the company external/public API/SDK surface — what an outside developer would consume — rather than the internal dogfood surfaces consume-mode targets. Output is the input to step 3 (human-gated target selection). Trigger phrases: "run produce-mode discovery on X", "discover external API surface for X", "A1 discovery on X", "produce a produce-mode discovery JSON for X".

Automated source discovery and structure extraction for a single company (steps 1–2 of the workflow in `SPEC.md`). Takes a company name (optionally `parent/sub-brand` for MEGA_CORP cases that have been pre-scoped), produces a JSON artifact conforming to `discovery-output-contract.md`, and writes it to a versioned per-run directory. Output is the input to step 3 (human-gated target selection). Trigger phrases — "run discovery on X", "discover sources for X", "produce a discovery JSON for X".

Step 3 of the workflow — human-gated target selection from a discovery.json. Reads a discovery output, presents ranked skill_targets to the human (sorted by booth-aha score for consume-mode v2 and v3+consume, by raw confidence for v1 and produce-mode v3), accepts a single pick, and persists a selection.json alongside the discovery file. The only manual step in the pipeline per `SPEC.md`. Trigger phrases — "pick a target", "select skill target", "human-gate selection", "step 3 selection", "run selection on a company slug".