Devin billing importer (Cognition Labs — reconciliation only)

SpendGuard cannot enforce Devin spend — only Cognition can

Devin (Cognition Labs) runs the agent loop inside Cognition’s cloud VM. The customer’s network never carries the per-LLM-call payload; the egress proxy + SDK adapters intercept nothing. The Devin importer is reconciliation only — it pulls post-hoc usage so the dashboard CIO/CFO single-pane-of-glass view doesn’t collapse, but it does not gate, throttle, or short-circuit Devin sessions. That contract belongs to Cognition. See the framework-coverage strategy memo Archetype IV for the architectural reasoning.

What this is

The spendguard-importer-devin crate (D14) is a Rust binary + library that:

1. Pulls usage from the Devin Team API (/api/v1/teams/{id}/usage) — feature-gated behind live so the default build is HTTP-free.
2. Converts ACU (Agent Compute Unit, ≈ $2.25/ACU) to estimated USD via a vendored price table at services/importer_devin/assets/devin_acu_prices.json.
3. Emits signed spendguard.audit.import.devin_acu CloudEvents tagged with reservation_source=subscription_meter and import_source=devin_team_api — landing on the same dashboard surface as D13 subscription-meter rows.

The default merge gate is a sanitized fixture replay — the committed tests/fixtures/devin_usage.json snapshot. Live mode is operator-opt-in.

ACU → $ conversion

| Plan | usd_per_acu | Audit row shape | |------|--------------|-----------------| | team | $2.25/ACU (Cognition published rate, 2026-06) | amount_micro_usd > 0, reason_code = NULL | | enterprise | negotiated per contract | amount_micro_usd = NULL, reason_code = "devin_enterprise_negotiated_rate" |

The conversion is pure: round(acu_consumed × usd_per_acu × 1_000_000) in micro-USD, saturating at i64::MAX on overflow. The price table’s pricing_version is stamped on every emitted row at the moment of conversion — back-revisions to the rate file do not rewrite historical audit rows.

Enterprise NULL-rate caveat

Enterprise plans negotiate ACU rates under NDA. SpendGuard cannot guess the dollar amount — so the importer emits amount_micro_usd = NULL and stamps reason_code = "devin_enterprise_negotiated_rate". Dashboards must distinguish “unknown rate” from “zero spend” — the row’s ACU field is still populated, so operators see Devin consumption volume even when the dollar mapping is private.

Install

Fixture mode (default merge gate)

cargo run -p spendguard-importer-devin --bin importer_devin -- \
    --mode fixture \
    --fixture services/importer_devin/tests/fixtures/devin_usage.json \
    --tenant <your-tenant> \
    --budget <your-budget-id>

The binary prints one CloudEvent per record to stdout; the demo runner script (deploy/demo/import_devin_fixture_demo.sh) wires the INSERT path against the demo postgres.

Live mode (operator-opt-in)

export DEVIN_API_TOKEN=<Team Admin API token>
export DEVIN_API_BASE_URL=https://api.devin.ai/api/v1   # optional override

cargo run -p spendguard-importer-devin --features live --bin importer_devin -- \
    --mode live \
    --tenant <your-tenant> \
    --budget <your-budget-id>

The live feature pulls reqwest with rustls-tls only (no native-tls, no openssl-sys) and uses a 30s per-request timeout plus bounded exponential backoff (cap: 1 hour) for resilience against rate-limits.

Scope the token to the Devin Team Admin API scope only — least privilege.

Idempotency

Re-running the same window does not double-emit. The CloudEvent event.id is a deterministic UUIDv5 derived from (devin_team_id, devin_session_id, window_end); canonical_ingest dedups via the existing event_replay_dedup table.

You can re-run the demo as many times as you like:

make -C deploy/demo demo-verify-import-devin-fixture
make -C deploy/demo demo-verify-import-devin-fixture   # idempotent; 0 new rows

Sample CloudEvent

{
  "specversion": "1.0",
  "type": "spendguard.audit.import.devin_acu",
  "source": "spendguard-importer-devin",
  "id": "018f4a3a-d971-7c14-91fe-d014de71aca0",
  "time": "2026-06-08T12:00:00Z",
  "datacontenttype": "application/json",
  "subject": "tenant/demo/devin/team/TEAM_FIXTURE_001/session/SESSION_FIXTURE_001",
  "data": {
    "schema_version": "v1alpha1",
    "tenant_id": "demo",
    "budget_id": "devin-budget",
    "devin_team_id": "TEAM_FIXTURE_001",
    "devin_session_id": "SESSION_FIXTURE_001",
    "acu_consumed": 12.5,
    "usd_per_acu": 2.25,
    "amount_micro_usd": 28125000,
    "pricing_version": "devin-acu-v1-2026-06",
    "window_start": "2026-06-01T00:00:00Z",
    "window_end": "2026-06-01T01:00:00Z",
    "reservation_source": "subscription_meter",
    "import_source": "devin_team_api",
    "ingestion_mode": "fixture",
    "fixture_provenance_sha256": "aa4c172164a8a6a5d4e97c6bde4ac455e01f5f37932b8a3561ef213049144807"
  }
}

The full schema doc lives at docs/specs/coverage/D14_devin_importer/cloudevent-schema.md.

Sample price table

{
  "pricing_version": "devin-acu-v1-2026-06",
  "effective_from": "2026-06-01T00:00:00Z",
  "currency": "USD",
  "rates": [
    { "plan": "team",       "usd_per_acu": 2.25 },
    { "plan": "enterprise", "usd_per_acu": null }
  ]
}

Why we cannot gate

The Devin agent loop runs entirely inside Cognition’s cloud VM. The customer’s egress proxy never sees the LLM API call payload — Cognition’s VM dials OpenAI / Anthropic from inside their own network boundary. The Team API exposes only post-hoc ACU consumption, not a pre-call hook.

This is Archetype IV in the SpendGuard framework-coverage analysis: a fully managed cloud agent where the customer has zero network visibility. The only feasible integration is reconciliation — pull the bill after the fact, surface it on the dashboard, alert on threshold crossings. Gating is architecturally impossible without Cognition shipping a pre-call webhook.

What you see on the dashboard

• Estimated $ per session, per team — labelled with pricing_version.
• ACU consumed (always present, even for enterprise NULL-rate rows).
• reservation_source = subscription_meter filter groups Devin spend alongside Claude Code Pro / Codex on ChatGPT (advisory rows that skip the BYOK ledger).
• Plan-level pivot: team vs enterprise.

Implementation pointers

• Crate scaffold + Cargo.toml: services/importer_devin/Cargo.toml — workspace-excluded, publish = false, live feature gated.
• Pure conversion: services/importer_devin/src/import_record.rs (import_record_to_audit_row) — no I/O, no clock read.
• Price table: services/importer_devin/src/acu_price_table.rs — saturating arithmetic, rejects NaN/Infinity/negative inputs.
• CloudEvent builder: services/importer_devin/src/cloudevent_envelope.rs — golden-tested via tests/golden/cloudevent_v1alpha1_*.json.
• Fixture loader: services/importer_devin/src/fixture_loader.rs — hard-rejects non-synthetic IDs at parse time (T5).
• Live client: services/importer_devin/src/live/client.rs — rustls-only reqwest wrapper with typed 401/403/429/5xx errors.
• Migration: services/ledger/migrations/0059_import_source_widen.sql — widens the existing audit_outbox.import_source CHECK to include devin_team_api.

For the full design rationale + slice-level breakdown, see docs/specs/coverage/D14_devin_importer/design.md.