[feature]: Track MiniMax Code (desktop coding agent) usage from ~/.minimax/v2/sessions JSONL

[Theoz001]

What problem does this solve?

MiniMax Code (MiniMax's desktop coding agent, Electron app; CN bundle id com.minimax.agent.cn, verified on v3.0.73 / macOS) is not among the 39 tracked tools, so its token usage is invisible in the dashboard — even though the app writes complete per-message usage records to local disk.

This is complementary to #429, which covered MiniMax cloud quota polling and was closed pending a safe credential source. Local session parsing needs no credentials and no network calls, so it doesn't hit that blocker.

Proposed solution

Add MiniMax Code as a passive reader (same family as pi / OmO / Kimi Code — nothing gets installed into the tool).

Where the data lives (verified on a live install, macOS):

~/.minimax/v2/sessions/YYYY/MM/DD/<HH-MM-SS-mmm>-session_<base64>/messages.jsonl

One JSON object per line (redacted sample):

{
  "message_id": "…",
  "turn_id": "…",
  "message": {
    "role": "assistant",
    "model": "k3-256k",
    "provider": "…",
    "timestamp": 1789743567882,
    "usage": {
      "input": 132719,
      "output": 16692,
      "cacheRead": 439298,
      "cacheWrite": 0,
      "totalTokens": 588709,
      "cost": { "input": 0, "output": 0, "cacheRead": 0, "cacheWrite": 0, "total": 0 }
    }
  }
}

Notes from inspecting a real install (290 session files spanning 2026-05 → today):

• Only message.role === "assistant" entries carry usage. The usage shape is the pi-family input / output / cacheRead / cacheWrite convention with a ms-epoch message.timestamp — exactly what parsePiLikeIncremental already consumes.
• The two format differences vs pi: there is no type: "message" wrapper, and the dedup id is message_id (top level) instead of id. Everything else matches, so a thin entry adapter over the existing pi-style engine should be sufficient.
• Migrated legacy sessions replay with model: "historical-transcript" and all-zero usage — naturally skipped by the zero-token guard.
• MiniMax Code is a router: the upstream model is recorded per message (MiniMax-M2.7, kimi-for-coding, glm-5.3, k3, …), so no global fallback model is needed (same situation as OmO).
• There is no session-header line carrying cwd, so project attribution can be left out initially.
• Suggested path override: TOKENTRACKER_MINIMAX_HOME following the TOKENTRACKER_OMO_HOME pattern.

Also present but apparently stale: ~/.minimax/sqlite.db has an ideal token_usage table (model, ts, input_tokens, output_tokens, reasoning_tokens, cache_read_tokens, cache_write_tokens, cost_usd), but on this install it only holds 2 rows from install day back in June — live data is in the v2 JSONL above. The v2 ~/.minimax/v2/sqlite/runtime-state.sqlite has no token tables at all. Session directories also contain subagent/task transcripts (*.readable.v1.jsonl) in case subagent attribution ever matters.

Privacy-wise this fits the house rule: the parser only needs message_id, role, model, timestamp, and usage — never any content.

Alternatives considered

• Parsing the Electron ~/Library/Application Support/MiniMax storage — Chromium LevelDB/IndexedDB, binary and brittle; unnecessary since the JSONL is complete.
• Waiting for [Feature Request] Add MiniMax (Minimax) provider for quota/limit tracking #429's quota integration — orthogonal: that path needs credentials, this one doesn't.

Category

New AI tool integration

Anything else?

• Verified on macOS; the Windows/Linux data directory (%USERPROFILE%\.minimax) is presumed but unverified.
• Happy to turn this into a PR following CONTRIBUTING.md's 5-step integration checklist if welcome — it looks like a thin adapter over the pi-style engine plus a recursive messages.jsonl walker.

[linear]

001-328