Cursor Agent provider misses CLI sessions stored in `~/.cursor/chats/**/store.db`

Summary

CodeBurn misses recent Cursor Agent CLI sessions because the cursor-agent provider discovers only transcript exports under ~/.cursor/projects/**/agent-transcripts/. Recent sessions are stored locally in per-session SQLite databases at:

~/.cursor/chats/<hash>/<session-id>/store.db

The separate cursor provider does not cover these databases either; it reads Cursor IDE state from the platform-specific globalStorage/state.vscdb.

The agent-tools/*.txt files observed alongside recent sessions are terminal-output captures, not conversation transcripts or usage records.

Current CodeBurn behavior

src/providers/cursor-agent.ts probes only:

~/.cursor/projects/
~/.cursor/ai-tracking/ai-code-tracking.db

Its discovery logic enters only directories named agent-transcripts and parses JSONL or legacy TXT transcripts.

src/providers/cursor.ts reads only Cursor IDE state.vscdb.

Neither provider discovers ~/.cursor/chats/**/store.db, so store-only sessions—and sessions whose transcript export is absent, delayed, or incomplete—are silently omitted.

Observed store.db format

Each session store contains:

Decoded root metadata includes:

Resolving the graph from latestRootBlobId yields structured user, assistant, and tool records, including conversation text, reasoning blocks, tool calls, tool results, request/model fields, and internal timestamps. Some observed stores also contain token- or context-related fields.

blobEncryptionKey is sensitive local material and must never be included in logs, diagnostics, errors, cache keys, exports, snapshots, or committed fixtures.

The attribution DB is not a fallback

~/.cursor/ai-tracking/ai-code-tracking.db tracks AI-authored code attribution and related metadata. It may help enrich a session with model or project information, but it does not contain the complete current conversation stream or reliable request-level usage records.

It cannot substitute for reading store.db.

Expected behavior

CodeBurn should discover and reconstruct Cursor Agent sessions from ~/.cursor/chats/**/store.db, including sessions that have no matching transcript export.

For the same session UUID, source precedence should be:

1. store.db
2. JSONL transcript
3. Legacy TXT transcript

A malformed or unsupported store may fall back to a matching transcript, but a successfully decoded store and its transcript must not both be counted.

Suggested implementation

• Add ~/.cursor/chats to cursor-agent discovery and probeRoots().
• Discover chats/*/*/store.db without following arbitrary symlink trees.
• Open each SQLite database read-only and validate the expected meta and blobs schema.
• Hex-decode meta['0'], validate agentId, and begin reconstruction at latestRootBlobId.
• Reconstruct the reachable blob graph while tolerating unknown protobuf fields and message block types.
• Use validated internal request/message timestamps when present.
• Use session metadata only as an explicitly session-level timestamp fallback; do not report database file mtime as an exact request timestamp.
• Prefer explicit request-level input, output, and cache counters only after validating their semantics.
• Do not treat a running context-window gauge as exact billed token usage.
• Estimate only missing token components and preserve exact/estimated provenance in the emitted calls.
• Deduplicate matching stores and transcripts using the session UUID plus stable request ID or turn index.
• Include each database’s WAL state in cache invalidation so active sessions refresh correctly.
• Retain the existing JSONL/TXT parser as a compatibility fallback.

Acceptance criteria

• A store-only Cursor Agent session appears under --provider cursor-agent.
• User and assistant messages are reconstructed in order.
• Structured tool calls and tool results are retained.
• Valid internal timestamps determine date attribution.
• Explicit, semantically validated token counts take precedence over estimation.
• Estimated values remain visibly identified as estimated.
• A store and matching transcript produce one set of calls.
• An unsupported or corrupt store can fall back to its transcript without double-counting.
• blobEncryptionKey cannot reach logs, errors, caches, exports, snapshots, or fixtures.
• Active WAL changes invalidate cached results.

Tests should cover:

• An empty chats directory.
• A valid minimal store.
• A multi-turn blob graph.
• Malformed hex metadata.
• Missing or invalid root/blob references.
• Unknown protobuf fields and message variants.
• Seconds versus milliseconds timestamp validation.
• Exact, partial, gauge-only, and absent token data.
• Store/transcript deduplication.
• Transcript fallback after store decoding fails.
• WAL-aware cache invalidation.
• Secret-redaction behavior.