feat: add Warp provider adapter

[Amirault]

Summary

• add a new lazy-loaded warp provider that reads Oz sessions from Warp SQLite (warp.sqlite) with WARP_DB_PATH override support
• parse finalized Warp exchanges from agent_conversations + ai_queries, estimate per-exchange token attribution from conversation totals, and normalize Warp model identifiers to priced model aliases
• attribute Warp run_command blocks from blocks to the nearest preceding exchange and extract bash command usage
• wire Warp into provider registry discovery, model aliasing, and session-cache environment fingerprinting
• add provider docs (docs/providers/warp.md), README/provider index updates, and dedicated Warp tests

Validation

• npm run test (63 files, 876 tests passed)
• npm run build
• npm run test -- tests/providers/warp.test.ts tests/provider-registry.test.ts
• npm run dev -- status --provider warp --format json

Artifacts

• Conversation: https://app.warp.dev/conversation/9fd84ac4-a81a-4eac-8d6d-ad88bea520ee
• Plan: https://app.warp.dev/drive/notebook/A7kPIYnXC9OsoXoKgdDx3s

Co-Authored-By: Oz oz-agent@warp.dev

[iamtoruk]

Code Review — Performance, Security & Compatibility Audit

Ran this against a real local Warp DB (macOS, 301MB warp.sqlite, 3 conversations). Provider discovers and parses correctly. No crashes. Full test suite passes (876/876).

Must Fix

1. Remove bare auto-efficient / auto-powerful from BUILTIN_ALIASES (src/models.ts:176-177)

These two generic keys create a global collision risk — any future provider emitting a model named auto-efficient would silently resolve to gpt-5.3-codex. They're also redundant: Warp's own normalizeModel() already maps auto-efficient → warp-auto-efficient before the name reaches BUILTIN_ALIASES, so the bare keys are dead code.

Remove these two lines:

-  'auto-efficient':                'gpt-5.3-codex',
-  'auto-powerful':                 'claude-opus-4-6',

The prefixed warp-auto-efficient and warp-auto-powerful (lines 174-175) are sufficient.

2. Remove unscoped README changes

The Node badge bump (20→22) and better-sqlite3 requirements rewrite are pre-existing doc debt, not Warp-specific. Please split these into a separate commit/PR so the changelog stays accurate.

Should Consider

3. outputTokens: 0 always

All tokens are attributed to inputTokens and cost is calculated at the input rate. For models where output costs 3-5x more than input (Opus), this significantly underestimates cost. The costIsEstimated: true flag helps, but consider a heuristic 60/40 split or at least a code comment explaining the trade-off.

4. Duplicate display-name aliases

'Claude Sonnet 4.6', 'Claude Haiku 4.5', etc. appear in both BUILTIN_ALIASES (models.ts) and modelAliases (warp.ts). These will drift. Consider removing the display-name entries from warp.ts since resolveAlias in models.ts already handles them globally.

5. toolSequence for retry detection

We just landed #353 which adds toolSequence to fix 100% one-shot rates for providers that aggregate tools per-session. Warp emits per-exchange so it's not technically broken, but if a single exchange involves multiple sequential tool invocations (Edit→Bash→Edit), those get flattened into one tools[]. Consider setting toolSequence: undefined explicitly or adding a comment noting the limitation.

Verified Safe

Area	Finding
SQL injection	All 6 queries use parameterized ? placeholders. No string interpolation.
DB access	Opened readOnly: true (sqlite.ts:124). Malicious WARP_DB_PATH → failed open, caught by try/catch.
decodeSourcePath colon split	macOS paths can't contain :. Edge case with crafted DB → failed open, no harm.
blobToText / XSS	Data flows to JSON, no HTML rendering context.
Blocks query performance	Full table scan with json_extract on read-only DB (can't add index). Acceptable for typical Warp DBs (<50k blocks). ~50-200ms worst case.
Discovery N+1	Correlated subquery per conversation. Fine for a few hundred conversations.
assignCommandBlocksToExchanges	O(blocks × exchanges) with early break. Both sets tiny per-conversation.
Provider registry (index.ts)	Lazy-load pattern matches Goose/Cursor/Crush exactly.
Session cache (session-cache.ts)	WARP_DB_PATH fingerprint isolated to Warp section.

Overall well-structured contribution. Items 1 and 2 are the only blockers.