# Support Claude Code and Cursor session providers ## Traceability - Spec ID: session-analysis-claude-cursor-providers - Status: Implemented - Port source: sibling source commit `33b19bf1` - Related local contract: [Stabilize Scripts Refactoring Contracts](2026-07-21-scripts-refactor-contracts.md) ## Intent Add Claude Code and Cursor as first-class, read-only providers for the existing `session-analysis` capability. Preserve provider provenance and evidence gaps while reusing the current selection, facts, facets, insights, usage, and semantic-facet owners. Claude project transcripts are the primary Claude source. Cursor project transcripts are joined with chat metadata and optional audit events because the transcript alone does not expose timestamps, results, or usage. The implementation must not treat generated Claude Insights facets, missing Cursor timestamps, or an unavailable audit join as observed outcomes. It must also keep Qoder and Codex output backward compatible. ## Acceptance Scenarios - SCCP-AC-1: The public `scripts/session-analysis.mjs` facade accepts `--platform claude` and `--platform cursor` for `sources`, `sessions`, `facets`, `insights`, `facts`, `file-reads`, `show`, `events`, and `claude-facets`, while existing Qoder and Codex commands remain compatible. - SCCP-AC-2: A provider may expand one raw record into multiple normalized events so every nested `tool_use` and `tool_result` retains its own invocation identity and evidence reference. - SCCP-AC-3: Claude discovery reads workspace-matching JSONL transcripts under `~/.claude/projects`, verifies the embedded cwd, and treats current and legacy Claude audit logs as optional lifecycle evidence rather than a prerequisite. - SCCP-AC-4: Claude normalization preserves bounded user and assistant text, tool requests and results, model and usage fields, timestamps, cwd, permission mode, and sidechain identity without treating titles, attachments, or queue metadata as task outcomes. - SCCP-AC-5: Generated `~/.claude/usage-data/facets` never enters production facts or findings. Derived usage-data artifacts remain excluded unless a future explicit contract marks their provenance and inference boundary. - SCCP-AC-6: Cursor discovery reads workspace-scoped agent transcripts, joins matching `chats/*//meta.json` records by session id, and joins only audit records whose session, conversation, or transcript identity belongs to the selected workspace. - SCCP-AC-7: Cursor normalization maps camel-case lifecycle events, tool input and output, model and token usage, permission decisions, file edits, shell execution, MCP execution, user prompts, assistant handoffs, and subagent boundaries into the provider-neutral event contract. - SCCP-AC-8: Missing or partial timestamps, tool results, model usage, or audit joins are reported as explicit source-coverage warnings or `Unobserved` evidence. They never become zero activity, successful validation, task completion, acceptance, or a provider-quality comparison. - SCCP-AC-9: Workspace matching and provider-home routing work on Windows, macOS, and Linux without shell-only path assumptions or reliance on one observed slug spelling. - SCCP-AC-10: Active-session exclusion is provider-owned. A provider without a reliable current-session environment id falls back to the existing bounded active-window rule rather than reading another provider's environment id. - SCCP-AC-11: `usage-summary` and the non-Canvas Harness analysis path accept Claude and Cursor evidence without coercing either provider to Qoder. Qoder remains the only Canvas-output host. - SCCP-AC-12: Focused fixtures cover discovery, workspace isolation, one-to-many message expansion, transcript/meta/audit joins, stale or missing optional sources, privacy-safe facts, usage coverage, and provider-labelled Harness output. - SCCP-AC-13: Validation includes focused session-analysis and Harness tests, the full repository suite, Markdown link-graph regeneration and validation, package verification, and bounded read-only forward tests against real local Claude and Cursor provider homes without persisting raw session content. ## Non-goals - Import Claude Insights semantic facets as observed session facts. - Decode Cursor's internal content-addressed `store.db` blob format in the first implementation. - Claim complete historical coverage when audit, metadata, or transcript stores cover different time windows. - Compare provider quality from event counts or missing fields. - Add Canvas output for Claude or Cursor. - Change findings, scores, or repair state merely because another provider is available. ## Plan and Tasks 1. Add a backward-compatible one-to-many normalization hook to both the public compatibility entrypoint and the capability-owned analyzer, plus a shared provider analysis runner for the new adapters. 2. Implement Claude source discovery, transcript expansion, optional audit normalization, source coverage, and fixtures. 3. Implement Cursor transcript discovery, chat-meta and audit joins, lifecycle normalization, source coverage, and fixtures. 4. Register both providers in the public CLI, capability-owned analyzer, usage route, planning/lifecycle host normalization, active-session resolution, and non-Canvas Harness path. Refresh the deliberately frozen CLI help fixtures and command descriptions to reflect the accepted provider surface. 5. Add focused tests and bounded real-home forward probes that emit only aggregate or privacy-safe envelopes. 6. Regenerate documentation routing, run the full validation stack, and update this spec to Implemented only when the visible evidence supports it. ## Test and Review Evidence - The focused provider, Harness, shared-contract, frozen CLI, and doc-link suite passed with 183 tests: `node --test test/session-analysis-providers.test.mjs test/harness-report-run.test.mjs test/task-loop-source.test.mjs test/session-analysis.test.mjs test/session-analysis-core-facts.test.mjs test/session-usage-summary.test.mjs test/session-selection-plan.test.mjs test/scripts-refactor-contract.test.mjs test/better-harness-cli.test.mjs test/doc-link-graph.test.mjs`. - `npm test` passed after the final capability-owned analyzer, CLI snapshot, and documentation adaptations. - `env npm_config_cache=/tmp/better-harness-npm-cache npm run pack:verify` passed with 284 npm-package entries and 318 runtime-zip entries. - `node scripts/doc-link-graph/cli.mjs skills/better-harness` regenerated the routing graph, and the doc-link test confirmed it is current. - Read-only local forward probes passed for Claude and Cursor bounded `facts` and non-Canvas `harness analyze`. Neither provider had a workspace-matching local session, so both correctly returned an explicit zero-session boundary without writing report artifacts.