Usage: include Devin (devin acp) session history on the Usage page

[Dvorinka]

Devin runs in T3 Code as a first-class agent through the ACP Registry (acpRegistry_devin), and Devin subscription limits are already being wired up in #26381. But Devin session history never appears on the Usage page — the readers only cover Codex, Claude Code, Grok Build, OpenCode, Antigravity, Pi, Muse Code, and Cursor. Every token spent through a Devin thread is invisible next to the other providers.

The Devin CLI already writes everything a usage reader needs to disk, so this seems like a natural devinUsageReader alongside the existing per-provider readers.

The data already exists locally

The same binary T3 launches via devin acp persists per-session transcripts and a session index on every machine it runs on:

~/.local/share/devin/cli/transcripts/<session_id>.json — one ATIF transcript per session:

{
  "schema_version": "ATIF-v1.7",
  "session_id": "zinc-order",
  "agent": { "name": "devin", "version": "3000.11.92", "model_name": "SWE-2 High" },
  "steps": [
    {
      "step_id": 0,
      "timestamp": "...",
      "metrics": { "prompt_tokens": 34409, "completion_tokens": 357, "cached_tokens": 832 }
    }
  ],
  "final_metrics": {
    "total_prompt_tokens": 12427771,
    "total_completion_tokens": 39023,
    "total_cached_tokens": 12026929,
    "total_steps": 121
  }
}

~/.local/share/devin/cli/sessions.db — SQLite index correlating transcripts with working directory and timestamps:

CREATE TABLE sessions (
  id TEXT PRIMARY KEY,              -- matches transcript filename
  working_directory TEXT NOT NULL,
  backend_type TEXT NOT NULL,
  model TEXT NOT NULL,
  agent_mode TEXT NOT NULL,
  created_at INTEGER NOT NULL,
  last_activity_at INTEGER NOT NULL,
  title TEXT, ...
)

This covers both T3-spawned sessions and standalone devin CLI sessions — same store, same format — matching how the Claude/Codex/Pi readers count provider history regardless of who launched it.

Suggested shape

Following the existing driver pattern:

• New reader apps/server/src/provider/Drivers/devinUsageReader.ts, modeled on claudeUsageReader.ts / codexUsageReader.ts.
• Wire into the usage scan in apps/server/src/usage/UsageService.ts and the provider lists in apps/web/src/components/usage/usageProviders.ts and apps/mobile/src/features/usage/usageProviders.ts.
• Default directory ~/.local/share/devin/cli (XDG-respecting); an env override would follow OPENCODE_DATA_DIR / ANTIGRAVITY_DATA_DIR / PI_CODING_AGENT_SESSION_DIR precedent.
• Scan transcripts/*.json, join sessions.db for working_directory / timestamps so per-project breakdown (feat(usage): break usage down by project, thread, account and environment #15902) works when it lands.

Caveats worth noting

• Schema versioning: transcripts declare schema_version: ATIF-v1.7. Parsing should tolerate unknown fields and degrade gracefully across ATIF versions — the CLI updates frequently.
• Token mapping: prompt_tokens / completion_tokens / cached_tokens per step. cached_tokens maps to cache-read; there is no separate cache-write figure. final_metrics has per-session totals if step-level aggregation is preferred.
• Dedup: no request-level IDs; (session_id, step_id) is the natural key.
• Pricing: SWE-* models have no public rates, so cost shows as unpriced until the user sets one — the existing Model prices / Map to overrides already cover this.
• Platform paths: the CLI honors XDG_DATA_HOME; the reader should mirror its per-platform base-dir resolution (Devin ships Linux/macOS/Windows).
• Transcript filenames are the human-readable session IDs Devin assigns (branch-marmot, zinc-order, ...), good enough for row labels.

Related work

• fix(usage): show Devin in subscription Limits #26381 — fix(usage): show Devin in subscription Limits (complementary, not overlapping)
• feat(usage): track Mistral Vibe session usage on the usage page #23380 — feat(usage): track Mistral Vibe session usage (precedent: third-party agent usage reader)
• feat(usage): GitHub Copilot sessions show in usage history #27059 — feat(usage): GitHub Copilot sessions show in usage history
• feat(provider-kiro): Usage shows Kiro credits #27023 — feat(provider-kiro): Usage shows Kiro credits
• feat(usage): add command-based usage limit sources #24031 — feat(usage): add command-based usage limit sources

Reproduction

Any machine with the Devin CLI used through T3 or standalone already has the data:

ls ~/.local/share/devin/cli/transcripts/   # one <session>.json per session
sqlite3 ~/.local/share/devin/cli/sessions.db \
  'SELECT id, model, working_directory FROM sessions;'

Verified on Devin CLI 3000.11.92–3000.11.234, Linux.

[piotrnowakowski]

I implemented this in #29631 and would like a maintainer to sign off on the direction and scope.

A few things I found that differ from the proposal above:

• The storage format depends on the Devin version. Devin 3000.11.234, the version the ACP Registry currently installs, writes no transcripts/ folder. Token metrics live in sessions.db, in message_nodes.chat_message → metadata.metrics (input_tokens, cache_read_tokens, cache_creation_tokens, output_tokens), with generation_model and request_id alongside. The PR reads that. The same binary also includes a migration to a files-based session store (DEVIN_SESSION_STORE=files|sqlite), so a later reader may need to cover that layout too.
• Dedupe. One inference is stored on several message nodes with identical metrics. In my data, 80 assistant nodes were 40 requests. request_id is the key.
• Cache writes are reported. cache_creation_tokens is a separate field, so cache-write cost can be priced.
• Pricing. Model ids carry effort and fast-mode suffixes (claude-opus-5-5-high-fast, gpt-6-astra-medium). Stripping them gives a known rate model, with -fast priced at the fast tier. Devin bills in ACUs, so this is an API-rate estimate, like the other providers.
• Path. $DEVIN_SESSIONS_DIR (the CLI's own override), else $XDG_DATA_HOME/devin/cli, else ~/.local/share/devin/cli. I could only verify Linux.

Limits stay out of scope, since #26381 covers them.