bombot-forge

Use when driving a browser: BrowserSkill (`bsk`) is the default for read-only checks, QA, screenshots and dbgQuery reads in the browser you're already logged in to; the cdp lane (Chrome for Testing + `cdp.py`, port 9333) is for writes, anything where a native dialog matters, and `lens`/`netlog`/`stub`/`diff`. Covers which engine to pick (with measured numbers), the bsk safety recipes verified on a Mac (own session + pinned tab, dialog guard, NetSuite identity gate, bounded observe), and the cdp launch/commands/coordinator rules. For NetSuite session recovery (black screen, SSO loop, permission gates) pair with `netsuite-qa-browser`.

Use when turning an existing NetSuite bundle (account-owned, "Convert to SDF Project" available on its Bundle Details page) into a version-controlled SDF repo — extracting the conversion zip, cleaning up legacy auto-generated scriptids, wiring the repo so a future `file:upload`/deploy still lands on the bundle's real live paths, and scaffolding it as a Teibto standards repo.

Use when you need to read or verify live NetSuite account state — chart of accounts, GL postings, custom record/field values, saved searches, or script deployment status — without opening the NetSuite UI, especially to confirm real state before or after a change instead of assuming.

Use when writing to a live NetSuite record from a logged-in browser tab (updating a field, fixing a bad value, a scoped submitFields or load/save), instead of allow-listing raw `cdp.py eval`. Covers the scoped `ns_write.py` helper (structured args only, submit + save modes, account guard, dry-run default), its two engines (bsk: sandbox freely, production only with an explicit `--allow-bsk-prod`; cdp: anywhere), the exact `permissions.allow` line it needs, and a validator that checks each machine.

Use when deploying NetSuite SDF customizations (scripts or objects) to a live or production account, running a per-round prod deploy, or importing files/objects from a target account to compare before deploying. Covers the import-compare-confirm flow and the temp defaultAuthId swap.

Use when setting up browser automation on a machine for the first time or changing it: choose between bsk (BrowserSkill, everyday Chrome) and cdp (Chrome for Testing) with the trade-offs, install the chosen one, optionally connect Claude in Chrome, then offer a Dev Bridge trial and record whether Claude may click the NetSuite Login button for you. Interactive — asks before each step; saves the choices to a per-machine prefs file.

Use when setting up or using a coding agent CLI (Cline, or OpenCode if it is what the machine has) as the worker for ALL code/text edits, with Claude only briefing it and reviewing the result — for example to spend the company's DeepSeek tokens instead of Claude's. Covers detecting which CLI exists, pointing it at the TEIBTO DeepSeek endpoint, the privacy settings to check, a smoke test, the worktree-isolated run helper `agent_run.py`, and the review-then-apply protocol. Not for browser, production, deploy, or secret-handling work — those stay with Claude.

Use when setting up or syncing a machine's global `~/.claude/CLAUDE.md` to match BomBot's canonical NetSuite-dev instructions — on a fresh machine, or to reconcile drift between machines. Compares the bundled reference snapshot against the machine's existing global CLAUDE.md and proposes a section-by-section merge (never a blind overwrite). The reference is redacted; real customer/email/path values are filled per-machine.

Use when setting up, checking, or repairing the `local-llm` worker on a machine — the free, private subagent that delegates bulk text work to Ollama on the tailnet GPU box (`bombot-gaming`) through the `ollama` MCP server. Installs the MCP script and agent file from canonical copies, registers the MCP server, and runs a PASS/FAIL checker. Also covers re-snapshotting when the agent file or MCP script is edited on a machine.

Use when setting up, testing, or troubleshooting the `teibto-worker` subagent on a machine — the worker that delegates bulk low-judgement text work to a cheap external model (deepseek via the TEIBTO OpenAI-compatible endpoint). Covers saving `TEIBTO_API_KEY` once per machine without it touching a repo or shell history, the smoke test, switching model/endpoint by env var, and the python.org-macOS empty-CA-store gotcha.

Use when delegating a coding task to teibto-worker (DeepSeek on the TEIBTO endpoint) — the user types /teibto-code <task>, or asks teibto-worker to write or draft code. Decides whether delegating is worth it, runs the data check, attaches the repo's worker brief if it has one, delegates, then verifies every result line by line before the calling Claude applies it.

Use when the user types /teibto-worker <task>, or asks to send a task to teibto-worker (DeepSeek on the TEIBTO cloud endpoint). Hands the task to the `teibto-worker` subagent and relays its answer plus the verbatim token/cost line. For private data the default is the local `local-llm` (Ollama) instead.

Use when turning an informal or ambiguous requirement (chat, comment, BRD, handoff) into a decision document for stakeholders, when the current (as-is) behaviour must be proven against real code and live data before proposing changes, or when producing an analysis brief another person or AI will act on.

Use when transcribing a video or audio file to text (SRT with timestamps, plain TXT, and a cleaned Markdown) via the homelab `bombot` MCP server — a Tailscale bridge to a Windows GPU box (`bombot-gaming`) running ffmpeg + faster-whisper. Covers the prereq check, scp-to-inbox → verify-size → `transcribe_video` → scp-results-back flow, the sync-call wait warning, and the SSH job-object gotcha.