에이전트 전반에 걸쳐 AI 코딩 세션을 지속적으로 만들기

Vendor-neutral handoff for AI coding tasks — between Claude, GPT, Gemini, DeepSeek, local LLMs. Reads from Claude Code, Codex, Cursor, Aider. Local-first. Claude stalls. Codex refuses. Cursor goes off-track. You spend 30 minutes re-explaining context you already solved. - What decisions were made? - What approaches already failed? - Which file you were editing? 👉 None of that survives a session. Git shows what changed. It does NOT tell you: - why it was written - what was rejected - which agent produced it - what context led here That reasoning disappears. Drift is git blame for AI decisions - Copy-paste chat history - Re-explain everything - Lose decisions - AI repeats mistakes - Structured markdown brief - Decisions + rejected paths included - Resume instantly - No re-explaining The problem: Your AI coding agent stalled — refused, rate-limited, or just got dumb. Now you need to transfer 30 minutes of context to another agent. Re-pasting a chat history doesn't work; the new agent doesn't know which decisions are settled, which approaches you already rejected, or which file you were halfway through. drift handoff packages your in-progress task into a markdown brief any LLM can absorb cold: $ drift handoff --branch feature/oauth --to claude-code ⚡ scanning .prompts/events.db ⚡ extracting file snippets and rejected approaches ⚡ compacting brief via claude-opus-4-7 ✅ written to .prompts/handoffs/2026-04-25-1530-feature-oauth-to-claude-code.md The brief lists what you've decided, what you tried and rejected, what's open, and where to resume. Paste it into the next agent and they pick up mid-task without you re-explaining. drift is built on top of an attribution engine that watches the local session logs of your AI coding agents (Claude Code, Codex, Cursor, Aider...), LLM-compacts each completed session, stores the result in .prompts/ inside your git repo, and binds each session to its matching commit via git notes . The handoff feature is the cross-agent task-transfer wedge; the attribution engine powers drift blame / drift log underneath. After installation, drift log still shows multi-agent attribution per commit: commit abc1234 — Add OAuth login 💭 [claude-code] 7 events accepted, 0 rejected 💭 [codex] 3 events accepted, 1 rejected ✋ [human] 2 manual edits …and drift blame still resolves any line back to its full timeline. See docs/VISION.md for the broader thesis. AI coding stopped being a single-agent workflow. A real session today involves: - Switching agents mid-task: Claude rate-limits, Codex stalls, the model goes off the rails — refusing tasks, hallucinating tool calls, or producing low-quality output. You move to another agent and burn 10 minutes re-explaining what you'd already decided and rejected. - Forgetting your own work: a week later, git blame tells you which line you wrote, but not which prompt produced it, which approach was settled, or why. - Onboarding teammates with zero context: they see the code, but the reasoning lived in someone else's chat history on someone else's laptop. drift turns that disposable AI trail into durable project memory: - Capture, locally: drift capture (anddrift watch for live mode) reads the session JSONL your agents already write under~/.claude/projects/ and~/.codex/sessions/ . Nothing leaves your machine except an optional Anthropic compaction call you can turn off. - Compact, into markdown: each session becomes a small markdown summary in .prompts/sessions/ — decisions kept, approaches rejected, files touched. Cheap to read, cheap to grep, survives any vendor migration because it's just text in your repo. - Bind, to commits: drift bind /drift auto-bind attaches each session to the commit it produced viagit notes (refs/notes/drift ). The link travels with the repo; it does not pollute commit history. - Hand off, when you switch agents: drift handoff --branch --to produces a brief the next agent can absorb cold — what's done, what's open, what was already rejected, and where to resume. - Reverse-lookup, when you forget: drift blame [--line N] returns the full timeline behind a line of code: which session, which prompt, which agent, plus the human edits that landed on top. - Forward-lookup, when you remember the session but not the diff: drift trace lists everyCodeEvent that session produced. - Audit, across a release: drift log isgit log with a per-agent summary under each commit — useful when you need to answer "how much of this release was AI vs. human" without trusting LOC ratios. Net effect: multi-agent AI coding becomes something you can hand off, review, and reconstruct months later — instead of a chat history that disappears the next time you close a tab. Homebrew (macOS arm64/x86_64, Linux arm64/x86_64): brew install ShellFans-Kirin/drift/drift crates.io (Rust 1.85+ toolchain required): cargo install drift-ai Pre-built binaries (GitHub Releases): curl -sSfL https://github.com/ShellFans-Kirin/drift_ai/releases/latest/download/drift-v0.4.2-$(uname -m)-unknown-linux-gnu.tar.gz \ | tar xz -C /tmp && sudo mv /tmp/drift /usr/local/bin/drift drift --version From source: git clone https://github.com/ShellFans-Kirin/drift_ai.git cd drift_ai cargo install --path crates/drift-cli drift does not scrub session content. Anything you typed into your Claude Code / Codex session — including secrets you may have pasted — will be mirrored into .prompts/ and, by default, committed to your repo. Two knobs today: - Set [attribution].db_in_git = false in.prompts/config.toml to keepevents.db local only. - Review .prompts/sessions/ beforegit add . By default, db_in_git = true (teams share blame via the repo). For solo repos or sensitive content, flip it to false — see Configuration. A future release will add a regex-based redaction pass. For full coverage today, pair drift with gitleaks or trufflehog as a pre-commit hook. If you routinely paste secrets into AI sessions, wait for the redaction pass before enabling drift on shared repos. The first time you run drift capture , you'll see a one-shot notice restating the above; press Enter to acknowledge. Set DRIFT_SKIP_FIRST_RUN=1 to suppress in CI. See docs/SECURITY.md for the full threat model and roadmap. Six commands, zero config: cd your-git-repo drift init # scaffold .prompts/ drift capture # pull sessions from ~/.claude + ~/.codex drift handoff --branch feature/oauth --to claude # task transfer (since v0.2) drift blame src/foo.rs # reverse lookup: who wrote what drift trace # forward lookup: session → events drift install-hook # auto-run after each commit drift handoff is the v0.2 headline feature: package your in-progress task into a brief the next agent can absorb cold. See §Handoff for the full flow. Verified from /tmp with zero prior state: rm -rf /tmp/drift-smoke && mkdir -p /tmp/drift-smoke && cd /tmp/drift-smoke git init -q && git config user.email "x@y" && git config user.name x drift init && drift capture && drift list drift handoff reads your local events.db (built up by drift capture or drift watch ), filters down to the sessions in the scope you ask for, and produces a markdown brief structured for handoff: - What I'm working on — 3-5 sentences of intent (LLM-compacted). - Progress so far — done / in-progress / not-started bullets. - Files in scope — modified ranges with ±5 lines of context. - Key decisions — with session+turn citations. - Rejected approaches — pre-extracted from session tool errors. - Open questions / blockers. - Next steps. - How to continue — the prompt to paste into the target agent. # scope by branch (recommended): all sessions whose commits land on this # branch since it diverged from main drift handoff --branch feature/oauth --to claude-code # scope by time drift handoff --since 2026-04-25T08:00:00Z --to codex # single-session debug drift handoff --session abc12345-xxx --print # pipe to clipboard or another tool drift handoff --branch feature/oauth --print | pbcopy The default model is claude-opus-4-7 — the brief is what the next agent reads verbatim, so narrative quality matters more than for the per-session compaction in v0.1. Each handoff costs ≈ $0.10–$0.30 USD at Opus rates. To trade narrative for ~30× cost reduction, drop to Haiku in .prompts/config.toml : [handoff] model = "claude-haiku-4-5" # default is "claude-opus-4-7" drift watch is an event-driven daemon backed by the platform's native file-system notifications (FSEvents on macOS, inotify on Linux, ReadDirectoryChangesW on Windows). It coalesces rapid writes against the same session file inside a 200ms window, so a long running Claude Code or Codex session produces one capture pass per idle moment, not one per tool call. State is persisted to ~/.config/drift/watch-state.toml so a restart resumes rather than rescans. Ctrl-C finishes the current capture and exits cleanly. drift watch # drift watch · event-driven; Ctrl-C to stop # watching /home/you/.claude/projects # watching /home/you/.codex/sessions # first run; capturing every session seen # drift capture · provider=anthropic # Captured 10 session(s), wrote 192 event(s) to .prompts/events.db # ... # drift watch · interrupt received; exiting after last capture Every Anthropic compaction call is logged to compaction_calls in events.db with input / output / cache token counts and a computed USD cost (built-in pricing table per model; cross-check against https://www.anthropic.com/pricing before trusting as invoice-ready). drift cost # drift cost — compaction billing # total calls : 10 # input tokens : 120958 # output tokens : 6582 # cache creation : 0 # cache read : 0 # total cost (USD) : $0.1539 drift cost --by model # ── grouped by model (descending cost) # key calls input_tok output_tok cost (USD) # claude-haiku-4-5 10 120958 6582 $0.1539 drift cost --by session # ── grouped by session (descending cost) # key calls input_tok output_tok cost (USD) # 4b1e2ba0-621c-4977-af3f-2a9df5ac45ec 2 51696 2448 $0.0564 # ad01ae46-156f-403b-b263-dd04a232873a 1 33662 2390 $0.0456 # ... Filter with --since , --until , --model . Switching from Opus to Haiku on the same 10-session corpus takes compaction from $2.91 → $0.15 — a ~19× reduction at the cost of slightly terser summaries. Same 4-turn fixture session, different provider for the handoff brief: | Provider | Model | Cost USD | Latency | |---|---|---|---| | Anthropic | claude-haiku-4-5 | $0.0028 | 2281 ms | | OpenAI | gpt-4o-mini | $0.0003 | 3201 ms | | Gemini | gemini-2.5-flash | $0.0003 | 1505 ms | | DeepSeek (compat) | deepseek-chat | $0.0005 | 1906 ms | | Ollama | local model | $0 | varies | DeepSeek vs default Anthropic Opus on a real handoff: roughly 30× cheaper at similar narrative quality. Switch with one line in .prompts/config.toml — see Configuration. Source: docs/V030-V040-SMOKE.md + docs/V042-BUGFIX-VERIFY.md . drift blame is the reverse lookup. Given a line of code, it returns the full timeline of who touched it (multi-agent + human edits), with each entry linked back to the originating session and prompt. See docs/VISION.md for the three core scenarios: reverse (drift blame ), forward (drift trace ), and audit (drift log ). Drift AI ships its own stdio MCP server (drift mcp ). Any MCP-compatible client can call the five read-only tools — drift_blame , drift_trace , drift_rejected , drift_log , drift_show_event — to query the attribution store without spawning a subshell. Claude Code (one command): claude mcp add drift -- drift mcp Codex: codex mcp add drift -- drift mcp Tools are read-only by design — anything that mutates state (capture / bind / sync ) stays CLI-only. | Command | Purpose | |---|---| drift init | scaffold .prompts/ + project config | drift capture | one-shot: discover sessions, compact, attribute | drift watch | background daemon, debounced re-capture | drift handoff [--branch B --to A --print --output P] | v0.2 — cross-agent task brief | drift list [--agent A] | list captured sessions | drift show | render a compacted session | drift blame [--line N] [--range A-B] | reverse lookup | drift trace | forward lookup | drift diff | single event's unified diff | drift rejected [--since DATE] | list rejected AI suggestions | drift log [-- ] | git log + per-agent session summaries | drift bind | attach a session to a commit note | drift auto-bind | timestamp-pair every session to its closest commit | drift install-hook | install a non-blocking post-commit hook | drift sync push|pull | push/pull refs/notes/drift | drift config get|set|list | global + project TOML merge | drift mcp | run the stdio MCP server | Global: ~/.config/drift/config.toml Project (overrides): /.prompts/config.toml [attribution] db_in_git = true # default — teams share blame via the repo [connectors] claude_code = true codex = true aider = false # feature-gated stub [compaction] provider = "anthropic" # default; switch to "mock" for offline / testing model = "claude-haiku-4-5" # or claude-sonnet-4-6 / claude-opus-4-7 [handoff] model = "claude-opus-4-7" # narrative-quality is the value; switch to # "claude-haiku-4-5" for ~30x cost reduction ANTHROPIC_API_KEY : required for the live API compaction path. When it's unset, drift_ai transparently falls back to MockProvider and every summary is labelled [MOCK] so you never mistake a fallback run for a real one — nothing else in the pipeline changes. How drift compares to Cursor / Copilot history, Cody, and git blame itself: docs/COMPARISON.md . - Human-edit detection is SHA-ladder only — we do not claim authorship, the human slug means "no AI session produced this". See VISION.md. Bash python -c "open(...).write(...)" is best-effort; anything the shell lexer misses is caught by the SHA ladder and attributed tohuman .- Codex reasoning items are encrypted; we count them, we do not surface them. - Cost totals use a hardcoded pricing table — cross-check against https://www.anthropic.com/pricing before treating as invoice-ready. - Context-window truncation is deterministic head+tail elision (Strategy 1); hierarchical summarization (Strategy 2) is stubbed behind a feature flag for a future release. drift is an independent open-source project by @ShellFans-Kirin (shellfans.dev). It is not affiliated with Anthropic, OpenAI, or any other vendor whose agents it integrates with — drift is built on top of their session logs, not by them. Originally built for myself when I kept losing context between Codex stalls and Claude rate-limits. The v0.2 drift handoff feature is the part I personally use most. Apache 2.0 — see LICENSE. See CONTRIBUTING.md — it walks through adding a new connector using the Aider stub as the worked example.