finding-batch

name	finding-batch
description	Use when the user hands over a list, CSV, or directory of source content describing many findings at once and asks to create models for them in bulk. Includes triage, per-finding drafting, mechanical + quality review, human TUI sign-off, and (when applicable) CSV writeback. For one-at-a-time interactive work, use finding-author. For walking through existing model files to propose improvements, use finding-review.
allowed-tools	Bash, Read, Grep, Glob, Write, Edit, Agent

finding-batch

Bulk authoring from a list, CSV, or directory. Cross-finding context isolation is mandatory: each finding is drafted and reviewed by a fresh sub-agent with no exposure to its neighbors. The main skill orchestrates; sub-agents do the per-finding work.

All rules, CLI cheatsheets, and procedures live in prompts/fragments/ and prompts/defaults.yml. This file is pure orchestration.

Accepted input shapes

• CSV with finding names (+ optional description / synonym columns). Structure varies — inspect before parsing.
• Directory of markdown stubs, one finding per file.
• Plain pasted list of finding names in the chat.

Hood JSON and other legacy inputs are not supported.

Orientation (read first, once per session)

• prompts/fragments/core_concept.md — what a finding model is.
• prompts/fragments/session_defaults.md + prompts/defaults.yml — confirm source / contributors / tags with the user at session start.

Step 1 — Parse the input

• CSV: head -1 <csv>, head -3 <csv>. Confirm column layout with the user (see prompts/fragments/csv_writeback.md for column inspection). Identify the row-ID column and whether an OIFM column exists.
• Directory: list the files, confirm they're the ones to process.
• Pasted list: confirm the names back to the user, one per line.

Produce an in-memory list of incoming findings, each tagged with its source row ID (for CSVs) or source file path.

Step 2 — Triage pass (main skill, whole batch visible)

Load prompts/fragments/search_and_triage.md. For each incoming finding, generate 2-3 search targets and run them with small --limit. Pool results per finding and classify:

• Direct match — finding already exists (record the OIFM ID for CSV writeback).
• True ambiguity — needs user triage.
• No match — needs a new model.

Look for cross-row duplicates in the incoming list and surface them to the user before drafting.

Present a triage table and get user decisions on ambiguous rows.

Step 3 — Draft pass (fan-out, sub-agents, parallel groups of 5)

For each row that will become a new model, spawn one sub-agent via the Agent tool with:

• Instructions: "Draft the finding model inputs for this one finding. Read prompts/fragments/core_concept.md, prompts/fragments/naming.md, prompts/fragments/synonym_rules.md, prompts/fragments/scope_and_specificity.md, and prompts/fragments/session_defaults.md. Return a JSON object with {name, description, synonyms} only — the script will generate attributes."
• Context: the row or source content for this one finding only. Do not include other findings, prior conversation, or drafts.

Fan out in parallel groups of 5. Collect returned JSON snippets.

Step 4 — Create

Load prompts/fragments/create_invocation.md. Assemble a batch config (session source / contributors / tags + the collected snippets) and run scripts/finding_authoring/create_model.py --batch via heredoc.

Parse the filepath|name|oifm_id stdout into a {rowID → oifm_id} mapping for later CSV writeback.

Step 5 — Mechanical lint

Load prompts/fragments/mechanical_lint.md. Run review_model.py over the created files.

Step 6 — Quality review pass (fan-out, sub-agents, parallel groups of 5)

For each created file, spawn one sub-agent via the Agent tool with:

• Instructions: "Quality-review the finding model at <path>. Read prompts/fragments/core_concept.md and prompts/fragments/quality_checklist.md. Return issues, suggested fixes (concrete edits), extraction candidates (use exact attribute names), and warnings."
• Context: one file path. No neighbors, no batch context.

Fan out in parallel groups of 5.

Step 7 — Apply fixes and write review files

Apply the sub-agents' suggested fixes with Edit. Then, inline in the main skill (no sub-agent), load prompts/fragments/review_file_generation.md and write reviews/review_<label>_<n>.md files, splitting into groups of ~8-10 entries.

Step 8 — TUI handoff

Load prompts/fragments/tui_handoff.md. Tell the user how to launch the TUI. Wait for them to come back. The TUI step is mandatory for this skill.

Step 9 — Read back responses and apply feedback

For each entry: "ok" or blank → no action; specific feedback → apply the change with Edit; question → answer and iterate. If substantive changes were made, re-generate the affected review files and re-surface to the user.

Step 10 — CSV writeback (if source was a CSV)

Load prompts/fragments/csv_writeback.md. Inspect the CSV, decide columns with the user, extend the header if adding a new OIFM column, then run scripts/finding_authoring/update_csv.py with the mapping from step 4.

Stop conditions

• All created findings have passed quality review, TUI sign-off, and (for CSV-sourced batches) the OIFM IDs are written back.

Report: count created, count already-existing (referenced by OIFM IDs), count synonyms added, count remaining unprocessed. Do not commit without explicit user permission.