@code-insights/cli 3.3.1 → 3.4.0

package/CHANGELOG.md

@@ -2,6 +2,47 @@
 
 All notable changes to `@code-insights/cli` will be documented in this file.
 
+## [3.4.0] - 2026-03-02
+
+### Fixed
+
+- **Codex CLI parser rewrite** — Complete rewrite of the Codex CLI provider to handle the current JSONL format (v0.104.0+). Previously produced 0 assistant messages and 0 tool calls for all sessions. Now correctly parses `function_call`, `function_call_output`, `custom_tool_call`, `agent_message`, and `task_complete` events. Also adds support for the legacy single-JSON format (pre-2025 `.json` files).
+- **Cursor provider data quality** — Three fixes: (1) project path inference from code block URIs when workspace.json is unavailable, (2) Lexical JSON rich text extraction with proper paragraph separators, (3) VSCode URI object unwrapping for file paths in tool calls.
+- **Copilot VS Code metadata** — `models_used` and `primary_model` were always NULL (307 sessions affected). Provider now collects model IDs from session and per-request fields into a `SessionUsage` object.
+- **Copilot CLI timestamps** — `started_at` and `ended_at` were identical (collapsed to file mtime). Event timestamps live at the event root level, not inside `event.data` — parser now extracts from the correct location.
+- **Copilot CLI tool call IDs** — Were synthetic (`copilot-tool-N`) instead of using the original `toolCallId` from events. Also extracts tool calls from `assistant.message` `toolRequests` array.
+- **Copilot CLI model extraction** — Model field was always empty. Now extracted from `tool.execution_complete` events.
+
+### Added
+
+- **Provider-agnostic session character detection** — `detectSessionCharacter` now recognizes tool names from all providers (Claude Code, Copilot VS Code, Copilot CLI, Codex CLI, Cursor) instead of only Claude Code tool names (`Edit`, `Write`, `Read`, `Grep`, `Glob`). Uses `EDIT_TOOLS` and `READ_TOOLS` sets with provider-agnostic file path extraction.
+- **Dashboard agent message rendering** — Agent team coordination messages (`<task-notification>`, `<teammate-message>`) previously rendered as "You" bubbles. Now displayed as distinct notification cards with amber borders (task notifications) and colored borders (teammate messages).
+- **`usageSource: 'session'` type** — New usage source value for providers that have model info but no token data.
+
+## [3.3.2] - 2026-03-02
+
+### Added
+
+- **Richer analysis prompts (v3.0.0)** — Decomposed insight schemas: decisions now include situation, choice, reasoning, alternatives (with rejection reason), trade-offs, and revisit conditions. Learnings include symptom, root cause, transferable takeaway, and applicability. Summaries include outcome status.
+- **Session traits in prompt quality** — Detects higher-level behavioral patterns: context drift, objective bloat, late context, no planning, and good structure. Each trait includes severity, evidence, and suggestions.
+- **LLM-based session character classification** — Sessions are classified into one of 7 types (deep_focus, bug_hunt, feature_build, exploration, refactor, learning, quick_task) by the LLM during analysis, replacing the heuristic-only approach.
+- **PR link extraction** — GitHub PR links referenced in session messages are automatically detected and displayed as clickable badges on the session detail page.
+- **Few-shot examples in analysis prompt** — Two curated examples (a decision and a learning) set the quality bar for LLM output.
+- **Chain-of-thought pre-analysis** — Prompt quality analysis now uses a 6-step mental walkthrough before scoring.
+
+### Changed
+
+- **Analysis version bumped to 3.0.0** — New decomposed schemas are not backward-compatible with v2 insight format. Re-analyze sessions to generate v3 insights.
+- **Tool result cap raised from 200 to 500 chars** — Better context for error messages in analysis input.
+- **Source tool badge shown for all sessions** — Previously hidden for claude-code sessions.
+
+### Fixed
+
+- **Dashboard compact layout** — Higher information density across dashboard pages.
+- **7-day range filter** — Default range filter applied to dashboard and analytics.
+- **LLM JSON truncation** — 3-layer fix: max_tokens 8192, jsonrepair fallback, conciseness constraints.
+- **PostHog exception tracking** — Stack trace frames included in error reports.
+
 ## [3.3.1] - 2026-03-02
 
 ### Added

package/README.md

@@ -56,7 +56,7 @@ Opens the built-in React dashboard at `http://localhost:7890`. The dashboard pro
 
 - **Session Browser** — search, filter, and view full session details
 - **Analytics** — usage patterns, cost trends, activity charts
-- **LLM Insights** — AI-generated summaries, decisions, learnings, and techniques
+- **LLM Insights** — AI-generated summaries, decisions (with reasoning and trade-offs), learnings (with root cause analysis), prompt quality scoring with session traits
 - **Settings** — configure your LLM provider for analysis
 
 ### Options