@langcost/adapter-warp

LangCost adapter for Warp — reads Oz agent session data directly from Warp's local SQLite database and ingests it into the langcost analysis engine.

Quick start

# Install langcost + the adapter
npm install -g langcost @langcost/adapter-warp

# Scan your Warp sessions
langcost scan --source warp

# Open the dashboard
langcost dashboard

How it works

Warp stores AI session data in a local SQLite database at:

~/Library/Group Containers/2BBY89MBSN.dev.warp/Library/Application Support/dev.warp.Warp-Stable/warp.sqlite

The adapter reads three tables:

| Table | Maps to | Description | |---|---|---| | agent_conversations | Trace | One per Oz agent session — token totals, model, status | | ai_queries | Span (llm) | One per exchange — model, timestamp, prompt, status | | blocks | Span (tool) | One per run_command call — timing, exit code, output |

The database is opened read-only and Warp uses WAL mode, so scanning is safe while Warp is running.

CLI options

# Scan the last 30 days (default)
langcost scan --source warp

# Scan all history
langcost scan --source warp --since all

# Point to a custom warp.sqlite path
langcost scan --source warp --path /path/to/warp.sqlite

# Choose Warp credit-rate assumptions used for credit->USD attribution
langcost scan --source warp --warp-plan build
# accepted: build | business | add-on-low | add-on-high | byok

# Force re-analysis of all sessions
langcost scan --source warp --force

Waste detection

| Rule | Supported | Notes | |---|---|---| | tool_failures | ✅ | Detected via blocks.exit_code ≠ 0 | | agent_loops | ✅ | Repeated run_command patterns across exchanges | | model_overuse | ✅ | Per-exchange model ID available | | retry_patterns | ✅ | Consecutive failed/cancelled exchanges | | high_output | ⚠️ approximate | Token counts are estimated (see limitations) | | low_cache | ❌ | Cache hit/miss data not exposed per exchange |

Known limitations

Token counts are estimated, not exact. Warp does not store per-exchange token counts in ai_queries. The adapter uses a two-signal estimation:

1. Estimates input size from the prompt JSON length using estimateTokenCount
2. Normalizes all estimates against the conversation's actual total from conversation_data.token_usage

Estimates sum to the exact conversation total and are proportionally accurate, but the input/output split is unknown — all tokens are treated as input for cost calculation. Aggregate cost per session is accurate; per-span cost is approximate.

Warp credit pricing is estimated from plan-level rates. Warp does not expose a deterministic per-model credit burn formula, so creditCostUsd is derived as: credits_spent × effectiveCreditRateUsd

effectiveCreditRateUsd is selected via --warp-plan:

• build: $20 / 1500 credits = $0.0133 per credit
• business: $50 / 1500 credits = $0.0333 per credit
• add-on-low: $0.0153 per credit (high-volume add-on tier)
• add-on-high: $0.0250 per credit (base add-on tier)
• byok: $0 (operational mode for API-direct sessions)

apiCostUsd is the equivalent direct API comparator from model pricing, and costMarkupPct compares paid-path cost against that comparator.

BYOK attribution is split from Warp credits. byok_tokens always contribute to direct API cost and never to creditCostUsd; warp_tokens contribute to Warp credit cost. For mixed sessions, totalCostUsd follows the paid path (creditCostUsd + byokApiCostUsd) while preserving apiCostUsd for arbitrage comparison.

Tool spans cover run_command only. Other tools (read_files, grep, apply_file_diff, etc.) are not persisted in the blocks table and cannot be reconstructed.

No assistant response content. Warp does not store the AI's response text. Only user messages (from the prompt) are available.

macOS only. Warp currently ships for macOS; the adapter auto-detects both Warp Stable and Warp Preview.

Schema stability. warp.sqlite is a private implementation detail of Warp and may change without notice. The adapter validates required tables at startup and fails with a clear error if the schema is incompatible.

Auto-detection

The adapter checks for Warp Stable first, then Warp Preview:

1. ~/Library/Group Containers/2BBY89MBSN.dev.warp/Library/Application Support/dev.warp.Warp-Stable/warp.sqlite
2. ~/Library/Group Containers/2BBY89MBSN.dev.warp/Library/Application Support/dev.warp.Warp-Preview/warp.sqlite

Use --path to override.