opencode-wiretap-viewer

Reads the request bodies captured by the opencode-wiretap plugin and gives you somewhere decent to read them. One process, one port, UI and API together — no build step, no config file.

bunx opencode-wiretap-viewer      # or: npx opencode-wiretap-viewer

Sessions on the left, that session's requests in the middle, the decoded payload on the right. Subagent sessions nest under the conversation that spawned them.

You need the plugin too

This package is only the reader. Without captures on disk it starts fine and shows you nothing. Add the plugin to your OpenCode config first:

// ~/.config/opencode/opencode.jsonc
{ "plugin": ["opencode-wiretap"] }

Restart OpenCode, send a message, then start the viewer.

What it shows you

A system prompt is rarely one thing, so the viewer splits it back into the pieces that built it — your AGENTS.md, the <env> block, MCP instructions, the skill catalog — each independently collapsible. Tool definitions get their own tab with full descriptions, which is normally where your token budget went. When you don't trust the pretty version, the original { timestamp, url, body } envelope is one click away, byte for byte as it was serialized.

Options

  -p, --port <n>       port to listen on            (env PORT / API_PORT, default 3001)
  -l, --log-dir <dir>  wiretap capture directory    (env LOG_DIR)
      --db <file>      OpenCode SQLite database     (env OPENCODE_DB)
      --models <file>  models.dev price table       (env OPENCODE_MODELS)
  -h, --help           show this message

CLI flags beat environment variables, which beat the defaults.

| Variable | Default | Purpose | | -------------------- | -------------------------------------- | ----------------------------- | | LOG_DIR | ~/.config/opencode/logs/wiretap | where the plugin writes | | PORT | 3001 | listen port | | OPENCODE_DB | ~/.local/share/opencode/opencode.db | read-only, for session titles | | OPENCODE_MODELS | ~/.cache/opencode/models.json | read-only, for cost estimates | | WIRETAP_COST_CACHE | ~/.cache/opencode-wiretap/costs.json | memoised per-file costs |

Point --log-dir at a copied capture directory to inspect someone else's session, or --port at something else when 3001 is taken.

Runtime

Node 20+ or Bun. Session titles come from OpenCode's own SQLite database, and the backend is chosen at runtime: bun:sqlite under Bun, node:sqlite under Node 22.13+, and no titles at all if neither is available. Everything else works regardless — titles are the only thing you lose, and sessions still list by ID.

The database is opened read-only. The viewer never writes to it, and never writes to your captures either.

Cost

Requests are priced from OpenCode's own models.dev snapshot at ~/.cache/opencode/models.json — the same table OpenCode bills against, read off disk with no network call. Missing table or unrecognised model means no cost is shown; it is never reported as $0.

Session totals need every capture parsed, which on a large corpus is minutes of disk the first time. So they are produced by a background sweep that memoises each file against its (mtime, size) in ~/.cache/opencode-wiretap/costs.json and re-runs as a cheap stat pass thereafter. /api/sessions answers instantly from that cache and reports null for anything not yet swept; the UI polls /api/cost/status and fills totals in as they land. The cache is disposable — delete it and it rebuilds.

The figures are estimates reconstructed from published rates. They ignore negotiated pricing and credits, so expect them to disagree with an invoice at the margin.

API

The UI is a client of these; they're stable enough to script against.

| Route | Returns | | -------------------------- | ----------------------------------------------------------------- | | GET /api/sessions | every session, newest first, with titles, parent IDs, cost totals | | GET /api/sessions/:id | that session's requests — model, message count, size, cost | | GET /api/sessions/:id/:f | one captured request, streamed verbatim | | GET /api/config | resolved paths, and whether the DB and price table were found | | GET /api/cost/status | background cost sweep progress |

Session IDs and filenames are pattern-checked and resolved inside the log directory, so .. gets you a 400 rather than a file. Anything else under /api 404s as JSON; every other path serves the SPA.

Links

• Repository and full documentation
• Issues

MIT