Two ways to use it: a local CLI for individual engineers reading their own Claude Code & Codex usage, or a self-hosted Platform with multi-tenant gateway, evaluator, and cost budgeting for engineering teams.
AI 開發績效評核工具 — 個人 CLI 讀取本地資料；團隊 Platform 提供 multi-tenant gateway、評核器與成本控管。
Pick whichever fits — they share no runtime state.
v0.1.x · for individual engineers
~/.claude/ and ~/.codex/ on your laptopnpm install -g @hanfour.huang/aidev0.2 → v0.5 · self-hosted for teams
/v1/messages + /v1/chat/completions with shared upstream account pool, per-user API keys, usage + cost dashboardsghcr.io/hanfour/aide-{api,web,gateway}Everything an individual engineer needs to evaluate their own AI-assisted development. (Platform-mode capabilities are listed in the Platform section below.)
Reads Claude Code session metadata, facets, SQLite cost data, JSONL conversations, and Codex thread data from local storage.
Detects iterative refinement, multi-task coordination, and active corrections to evaluate decision-making quality.
Identifies security awareness, performance discussions, bug catching, and other risk-control behaviors.
Generate reports in colored terminal, JSON (machine-parseable stdout), or Markdown format.
Bring your own evaluation criteria, keywords, thresholds, and superior rules via custom JSON standard files.
All data is read locally and read-only. No data is sent to any external service. Your code stays yours.
Built-in monthly and quarterly commands with --previous flag for calendar-aligned KPI reviews.
Auto-generated executive summary with headline, observations, and recommendations for KPI reviews.
Set defaults for locale, theme, output format, period days, and data directories via config command.
Both modes share an "extract → analyze → score → report" pipeline; Platform adds storage + multi-tenancy + LLM augmentation.
~/.claude/ & ~/.codex//v1/messages; bodies AES-256-GCM-encrypted into request_bodiesrequest_body_facets (per-session classification)Self-hosted multi-tenant evaluator. Released incrementally — every version is a feature flag flip away from the previous one.
Next.js UI + Fastify API on Postgres + Redis. OAuth sign-in, organization-scoped RBAC, member invites, immutable audit log.
Anthropic-native + OpenAI-compatible HTTP proxy that pools admin-donated upstream accounts and meters per-user usage.
sk-ant-... + OAuth bundles; scheduler picks per requestak_...) authenticated against the poolusage_logs with per-call token / cost attributionOpt-in content capture + dual-layer evaluation (rule-based + optional LLM Deep Analysis), with admin-customizable scoring rubrics.
Per-org monthly USD cap with two enforcement modes. Spend tracked in a dedicated ledger + visualised in a cost dashboard.
llm_monthly_budget_usd + llm_budget_overage_behavior (degrade or halt)llm_usage_events ledger (one row per LLM call)/dashboard/organizations/<id>/evaluator/costsOpt-in second LLM pass that classifies each captured session into structured JSON, persisted for rubric signals + report drill-downs.
sessionType, outcome, claudeHelpfulness, frictionCount, bugsCaughtCount, codexErrorsCountrequest_body_facets table; prompt_version cache prevents duplicate LLM callsfacet_*) wired end-to-endProduction-ready observability shipped with the release. The pipeline can be deployed and supported without secondary tooling.
docs/runbooks/release-blocker issue on canary failureAll data is read locally. Nothing leaves your machine. (Platform mode persists captured request bodies in Postgres with AES-256-GCM encryption — see the Platform section above.)
| Source | Path | Data |
|---|---|---|
| Claude Code Sessions | ~/.claude/usage-data/session-meta/*.json | Tokens, duration, tools, languages, git commits |
| Claude Code Facets | ~/.claude/usage-data/facets/*.json | Goals, outcomes, friction, helpfulness |
| Claude Code Costs | ~/.claude/__store.db | Per-message cost (USD), model, duration |
| Claude Code JSONL | ~/.claude/projects/*/*.jsonl | Full conversations for keyword scanning |
| Codex Threads | ~/.codex/state_5.sqlite | Tokens, model, title, git info |
| Codex History | ~/.codex/history.jsonl | Full user prompts by thread |
| Codex Logs | ~/.codex/logs_2.sqlite | Tool calls and error events |
Pick the path that matches your use case. Requires Node.js ≥ 18 for the CLI, or Docker + Postgres + Redis for Platform mode.
# Install globally from npm npm install -g @hanfour.huang/aide # Verify installation aide --help # Update to latest npm install -g @hanfour.huang/aide@latest
Multi-arch images on ghcr.io/hanfour/aide-{api,web,gateway}. aide-web is amd64-only since v0.5.0; aide-api and aide-gateway publish both amd64 and arm64.
# Clone, configure, start core stack (api + web + postgres + redis) git clone https://github.com/hanfour/aide.git && cd aide/docker cp .env.example .env # fill in OAuth + bootstrap email + secrets docker compose up -d # Opt-in: add gateway service docker compose --profile gateway up -d # Opt-in: enable evaluator subsystem (v0.4.0+) echo 'ENABLE_EVALUATOR=true' >> .env # Opt-in: enable LLM facet extraction (v0.5.0) echo 'ENABLE_FACET_EXTRACTION=true' >> .env
Full setup walk-through: docs/SELF_HOSTING.md · Gateway operator guide: docs/GATEWAY.md · v0.5.0 upgrade: docs/UPGRADE-v0.5.0.md
Generate reports with a single command. (Platform mode is operated via the web UI; see the Platform section above for screenshots / docs.)
# Quick summary (last 7 days) aide summary # Full report (last 30 days) aide report # Monthly KPI report as Markdown aide monthly --format markdown --output report.md # HTML report for manager review aide report --format html --output report.html # Previous quarter, JSON for automation aide quarterly --previous --format json | jq '.sections[].score' # Custom date range with engineer metadata aide report --since 2026-03-01 --until 2026-03-31 \ --engineer "Jane Doe" --department "R&D" # Use a custom evaluation standard aide report --standard my-standard.json
# View settings aide config # Set defaults aide config set locale zh-TW aide config set theme no-color aide config set defaultFormat markdown aide config set defaultPeriodDays 14 # Reset to defaults aide config reset
Built-in OneAD R&D AI-Application Evaluation Standard. Fully customizable. CLI ships the v1.0.0 baseline; Platform mode v0.5.0 ships v1.1.0 with additive facet supports.