feat(providers): add forge provider support

[sleicht]

Summary

• Add Forge as a supported provider by reading conversations from ~/.forge/.forge.db.
• Parse assistant usage into prompt, completion, cached token counts, model names, tool calls, and shell commands.
• Register Forge in the provider list and document its local data source.

Fixes #400

Testing

• I have tested this locally against real data (not just unit tests)
• npm test passes
• npm run build succeeds

For new providers only:

• I installed the tool and generated real sessions by using it
• npm run dev -- today shows correct costs and session counts for this provider
• npm run dev -- models --provider forge shows correct model names and pricing
• Screenshot or terminal output attached below proving it works with real data

Real Forge DB smoke test:
- Data source: ~/.forge/.forge.db
- Discovered 167 Forge sources
- Parsed 68 calls from the first 5 sources
- Provider output uses provider: forge
- Parsed model, input tokens, output tokens, cached tokens, tools, and session id without printing message content

Verification commands:
- CI=1 npm test -- tests/providers/forge.test.ts tests/provider-registry.test.ts tests/cli-status-menubar.test.ts
  - 23 tests passed
- CI=1 npm test -- --testTimeout=15000
  - 68 files passed
  - 993 tests passed
- npx tsc --noEmit
  - passed
- npx tsup
  - passed
- NODE_USE_ENV_PROXY=1 npm run build
  - passed

[iamtoruk]

Review

Security — clean. All SQL is parameterized (WHERE conversation_id = ?), DB opened read-only, no credentials/auth exposed, splitSourcePath handles : in paths correctly via lastIndexOf, user messages truncated to 500 chars.

Correctness — solid. Token math matches the codex.ts pattern: inputTokens = Math.max(0, promptTokens - cachedInputTokens) normalizes OpenAI-style inclusive prompt tokens to Anthropic semantics. calculateCost parameter order verified against the signature. CAST(workspace_id AS TEXT) for BigInt is tested and documented. Timestamp parsing handles SQLite format, ISO, and epoch fallback.

Tests — thorough. 8 tests covering core parsing, missing DB, zero-token skip, multi-message user prompt tracking, null title fallback, large integer workspace_id, invalid JSON, and registry integration. All 994 tests pass, tsc clean, build clean.

No breaking changes — purely additive.

---

One fix needed before merge:

discoverFromDb (line 238) selects context, created_at, updated_at but only uses conversation_id, title, and workspace_id. The context column holds the full conversation JSON (potentially megabytes per row). With 167+ conversations this pulls all message data into memory just to build a list of IDs.

-- current (wasteful)
SELECT conversation_id, title, CAST(workspace_id AS TEXT) AS workspace_id, context, created_at, updated_at
FROM conversations WHERE context IS NOT NULL

-- fix
SELECT conversation_id, title, CAST(workspace_id AS TEXT) AS workspace_id
FROM conversations WHERE context IS NOT NULL

You can narrow the ConversationRow type or just use a separate type for discovery rows. The parse query (line 155) still needs all columns — only discovery is affected.

Once that's addressed this is good to go.

[sleicht]

Thank you for the review. I addressed the conversationrow issue.

[iamtoruk]

Discovery query fix verified — only selects the 3 columns needed, separate DiscoveryRow type, source-level test guarding against regression. 995/995 tests pass, tsc clean, build clean.

[sleicht]

Thank you.