CANNBot-Insight

Session-level observability for LLM coding agents: import agent session logs and analyze tokens, context, subagents, and skill usage turn by turn.

中文文档

Features

Import SQLite (opencode.db) or JSONL (native Claude Code / Codex rollout-*.jsonl / cpx proxy captures; plaintext, .jsonl.7z and legacy .jsonl.gz, cpx- prefix handled) and explore the session detail tabs:

• Tokens & Cost — 5-item per-turn token breakdown vs the model's context window, plus cost estimation
• Active / Wall time — session detail Wall Clock split (active + inactive = wall) and Active / Wall columns on the homepage: event-clustering semantics (gaps >5min count as inactive, in-flight API calls count as active)
• Context Growth — Per-subagent context growth curves with animated replay; /compact markers
• Context Governance — Per-turn LLM input composition: visible messages + stable System overhead
• Subagent Tracking — Subagent sessions and dispatch→response bridges, inline in the timeline
• Turn Quick-Jump — Type #459 + Enter in the Turns timeline to jump to that turn
• Wire Rounds (proxy captures) — Per-round full wire input/output, cross-checked against the DB reconstruction; shown with -a
• Skill Events / Coverage / Content — Per-turn skill events, available-skills vs used comparison, SKILL.md full-text rebuild
• Agent Definition Recovery — One-click full text of dispatched agents' .md definitions: session system-injection first, disk scan fallback
• Concept Tracing — Cross-turn keyword search, propagation chains and DAG; filter by source/thinking/tool
• File Read Analysis / Restore / Gather — Duplicate-read detection, per-file line-by-line rebuild, whole-directory rebuild as zip
• Session Compare — Two sessions side by side: tokens / cost / latency / tools / subagents
• Trajectory Scoring — Six-dimension scorecard (completion veto / trust / quality / efficiency / redundancy / autonomy + custom dims), baseline comparison and scoring profiles
• Audit — Workflow audit (flow chart + per-node findings + priorities) and Skill audit (real execution vs SKILL.md / agent definitions / workflow declarations)
• Key Redaction (proxy captures) — API keys never touch disk: masked (abcd…wxyz) at the single write chokepoint

The homepage Sessions list supports keyword / session-ID search; the import dialog offers one-click default paths.

Run Logging — Server-side process logs (import pipeline, lake sync, API errors) are written to ~/.cannbot-insight/logs/insight-<date>.log (daily rolling) and optionally mirrored to the startup terminal. Level, file output and screen output switches are adjustable live in Settings → Run Logging (log location shown there; no restart); env vars (CANNBOT_LOG / CANNBOT_LOG_LEVEL / CANNBOT_LOG_FILE / CANNBOT_LOG_STDOUT / CANNBOT_LOG_DIR) take precedence over the settings page.

Install & Usage

Option 0: npm install (Recommended)

npm install -g cannbot-insight
cannbot-insight            # first run: auto-build + migrate (~30s-2min), then start + open browser

• Requires Node.js >= 22.12; mainland-China networks can add --registry=https://registry.npmmirror.com
• Offline fallback: npm install -g ./cannbot-insight-<version>.tgz
• Writable state lives in ~/.cannbot-insight/; default port 21025 (auto-increments if occupied)
• CLI subcommands pass through: cannbot-insight upload --file … --list / sessions; short alias cbi
• Python 3 optional (starts the smart-agent analysis backend when present)
• Alternate brand: npm install -g context-insight — the same tool published under the Context-Insight name (commands context-insight / ctxi, data dir ~/.context-insight/); session data is fully interchangeable with cannbot-insight (shared format identity)

Flag	Effect
-a, --advanced	Show advanced tabs (wireRounds/replay)
-k, --kill	Stop all insight/smart-agent instances of this repo, then start on 21025 (falls back to the next free port if occupied by others)
-f, --fresh	Clear the build cache and rebuild
-h, --help	Print the full CLI reference — all subcommands and flags (incl. -v, --version)

Option 1: Web UI from source

./start.sh              # production mode (default): cached build + zero per-request compile
./start.sh -u           # update deps, then start
./start.sh -f           # force a fresh build
./start.sh -d           # dev mode (hot reload while editing source)

• Requires Node.js >= 22.12; ./start.sh also installs the cpx capture command (see ../cpx/README.md)
• Open http://localhost:21025

Log file locations:

• opencode: ~/.local/share/opencode/opencode.db
• Claude Code: ~/.claude/projects/ (directory auto-scan)
• Codex: ~/.codex/sessions/ (rollout files; subagent sessions merge into their parent)

The Web UI also supports exporting sessions to standalone SQLite or Markdown. The LLM provider panel embeds a known-provider Base URL quick reference.

Option 2: CLI Import + Web Analysis

For SSH remotes and other browser-less environments (one-time npm install + npx prisma migrate deploy; the import command auto-starts the backend afterwards):

# opencode: list / import specific / interactive
npx tsx src/cli/index.ts upload --file ~/.local/share/opencode/opencode.db --list
npx tsx src/cli/index.ts upload --file ~/.local/share/opencode/opencode.db --session-id <id> --yes --json

# Claude Code: directory scan / single file
npx tsx src/cli/index.ts upload --file ~/.claude/projects/ --list
npx tsx src/cli/index.ts upload --file ~/.claude/projects/<hash>/<uuid>.jsonl --session-id <uuid> --yes --json

# Codex: native rollout files (subagent sessions merge into their parent)
npx tsx src/cli/index.ts upload --file ~/.codex/sessions/ --source codex-jsonl --list
npx tsx src/cli/index.ts import --source codex-jsonl --dir ~/.codex/sessions/ --all --yes --json

Option	Description
--file <path>	Source path (.db / .jsonl / directory; type auto-detected)
--session-id <id>	Pick a specific session (skip the interactive picker)
--description <text>	Upload description
--yes / --json	Skip confirmation / JSON output (script-friendly)
--list	List available sessions only