Part of the DEVIN ecosystem
Track: Understand · Nature: product
For: Developers, AI engineers, Data Scientists
Interface: CLI
Path: Data Scientists · step 2/4 — after devin-graph, before devin-history

devin-search

Unofficial community project. Not affiliated with, endorsed by, or sponsored by Cognition AI. "Devin" is a trademark of Cognition AI.

Linux · Personal Windows · Corporate Windows

Part of the awesome-devin ecosystem: the curated hub for the devin-* tools.

Full-text search across all your agent sessions — find that command, that error message, that file path or that decision from months ago in under a second. Devin (sessions.db, acp-messages/*.db), GitHub Copilot CLI, OpenAI Codex CLI and VS Code chat sessions all land in one index.

The problem

Devin keeps your entire session history in local SQLite stores (sessions.db, User/acp-messages/*.db) — but offers no way to search across them. You remember Devin fixed a flaky test or ran a specific kubectl command three weeks ago, and the only way back is scrolling through sessions one by one. Generic grep over the raw databases mostly hits JSON noise and knows nothing about who said what.

Prior art

• Full-text search over chat/agent history is well-established: everything here leans on SQLite's built-in FTS5 engine with BM25 ranking — the same approach used by ripgrep-style tools, mail clients and sqlite-utils.
• tokmesh and UniSessions document/parse the CLI sessions.db; the schema itself is tracked by devin-internals-spec, which this project uses for versioned, read-only access.
• devin-history (sibling repo) exports sessions to Markdown/JSON; devin-search complements it with instant ranked lookup instead of a static dump.

What makes it Devin-native

Results are role-tagged and session-linked, not raw grep hits. The indexer understands Devin's real message structure through devin-internals-spec's typed parsers: user prompts vs assistant replies vs tool calls (including prompt_history shell commands and GUI acp-messages sessions), each stamped with its working directory and a ref back to the exact source row. And because the store schema has had 17 migrations already, indexing fails loudly on an unknown schema version instead of silently misreading it.

Install

Requires Python ≥ 3.10 and pipx or uv: uv tool install devin-search or pip install devin-search. Per-OS setup lives in the platform guides: Linux · Personal Windows · Corporate Windows.

Usage

# build/update the index (auto-detects Devin's stores; safe to re-run —
# only new rows are indexed each time)
devin-search index

# search everything
devin-search query "kubectl delete pod"

# filters: role, project, date, limit — --json on every command
devin-search query "TypeError" --role assistant --project myrepo
devin-search query "migration" --since 2026-09-01 --limit 5 --json

# link hits to devin-history export notes
devin-search query "kubectl" --history-dir ~/notes/devin-history

# deterministic index: sessions.db only, never the auto-detected dirs
# (--no-acp, --no-copilot, --no-codex and --no-vscode-chat each disable
# one source; --copilot-dir/--codex-dir/--vscode-chat-dir pin it to one
# root instead of auto-detecting)
devin-search index --no-acp --no-copilot --no-codex --no-vscode-chat

# zero-hit stats from the local query log (evidence for/against semantic
# search — SE-1's trigger; log is opt-out via query --no-log)
devin-search misses

Hits print as WHEN · ROLE · PROJECT · SESSION · SNIPPET with the match wrapped in «»; each hit carries a ref (e.g. node:1234, devin-chat:file.db:7, codex-cli:…/rollout-x.jsonl:41, vscode-chat:…/vs-1.jsonl:0:resp) pointing back to the exact source row.

With --history-dir <dir> pointing at a devin-history export directory, each hit whose session has an exported note gets a note: <path> line under its row (and a history_note field in --json). Notes are matched by the <YYYY-MM-DD>_<session-id>.md filename devin-history writes (.json exports work as a fallback). A missing note produces no link and no error — the flag is purely additive.

Works with Devin alone (Devin-only mode)

devin-search builds and queries a fully local index over Devin's session stores. Nothing is sent anywhere; the index lives on your disk next to the data it covers.

Platform support

Tested on Windows and Linux (windows-latest + ubuntu-latest in CI). The Devin stores are auto-detected from the OS-specific locations listed in the platform guides — Linux · Personal Windows · Corporate Windows — and --sessions-db/--acp-dir override them.

The non-Devin surfaces auto-detect per the session-surface spec (devin-internals-spec → docs/SESSION-SURFACES.md §6.3–§6.5): copilot-cli merges $COPILOT_DIR, $COPILOT_HOME and ~/.copilot (session-state/ dir + flat files, legacy history-session-state/); codex-cli reads $CODEX_HOME else ~/.codex (newest state_<N>.sqlite plus sessions//archived_sessions/ rollouts); vscode-chat merges the Code and Code - Insiders User dirs plus ~/.vscode-server/data/User (workspaceStorage/*/chatSessions + globalStorage/*ChatSessions). Use --copilot-dir/--codex-dir/--vscode-chat-dir to pin one root or --no-copilot/--no-codex/--no-vscode-chat to disable a surface.

Limitations

• Schema-gated. Only sessions.db schema v15–v17 is accepted; newer versions fail loudly (update devin-internals-spec first).
• Opaque payloads. chat_message, tool_call_*_json, acp payload, copilot event data, codex rollout payloads and vscode-chat requests[] shapes are undocumented/unstable — text extraction is best-effort and tolerant, not contractual. Malformed JSONL lines are skipped, never fatal.
• Keyword search only (M1). BM25 over tokens — no synonyms or embeddings; semantic search is an opt-in M2 candidate.
• Deletion lag. New rows are picked up incrementally, but rows deleted mid-table upstream can stay in the index until --rebuild (tail-pruned sources are detected and re-indexed automatically).
• Read-only by design — the tool never writes to Devin's stores; the only file it creates is search.db.

Development

pip install -e ".[dev]"
pytest

Fixtures are generated at test time by devin_internals.fixtures (real v17 DDL, synthetic rows) plus per-surface writers in tests/conftest.py for the copilot/codex/vscode-chat stores — no binary fixtures are committed.

Adapters (MCP server, Devin skill, plugin)

devin_search.mcp_server exposes one read-only tool via the devin-search-mcp entry point (pip install 'devin-search[mcp]'): search_query (same JSON as devin-search query --json).

adapters/ is a self-contained Devin plugin root (adapters/.devin-plugin/plugin.json + adapters/skills/devin-search/SKILL.md), installed with devin plugins install Icaro0310/devin-explore#packages/search/adapters — it ships in the repo, not inside the wheel. The manifest launches the server through uvx --from 'devin-search[mcp]' devin-search-mcp, which resolves once the matching PyPI release ships; until then run the source-installed devin-search-mcp directly or point a local manifest copy at the checkout (uvx --from './packages/search[mcp]' devin-search-mcp).

MCP host configs

Same stdio server in other hosts — server name devin-search:

Zed — context_servers in ~/.config/zed/settings.json:

{ "context_servers": { "devin-search": { "command": { "path": "devin-search-mcp", "args": [] } } } }

JetBrains AI Assistant — Settings → Tools → AI Assistant → Model Context Protocol → Add:

{ "mcpServers": { "devin-search": { "command": "devin-search-mcp" } } }

Antigravity — ~/.gemini/config/mcp_config.json:

{ "mcpServers": { "devin-search": { "command": "devin-search-mcp" } } }

When to use this

• You remember Devin ran a command, hit an error, touched a file, or made a decision, and scrolling sessions one by one is too slow.
• You want ranked results tagged by role (user / assistant / tool call) with a ref back to the exact source row.
• Your session history must stay on disk — the index is fully local, no telemetry, no network calls.
• You want incremental indexing: re-running devin-search index only picks up new rows.

When NOT to use this

• You need semantic or synonym-aware search — M1 is keyword BM25 only; embeddings are an opt-in M2 candidate.
• You need session observability (activity, context size, token peaks) — use devin-metrics; or relationship queries — use devin-graph.
• Your sessions.db schema is outside v15–v17 — indexing refuses loudly rather than misreading it.

FAQ

What is devin-search? A local full-text search engine over your Devin session history. It indexes Devin's SQLite stores with FTS5/BM25 and answers queries like devin-search query "kubectl delete pod" in under a second, with results tagged by role and linked back to the source row.

Does devin-search send my session data anywhere? No. Everything runs locally: it reads Devin's stores read-only and writes a single search.db index next to your data. There are no network calls and no telemetry.

Does it write to or modify Devin's databases? No. Devin's stores are opened read-only by design; the only file devin-search creates is its own search.db index.

How is this different from devin-history? devin-history exports sessions to static Markdown/JSON files. devin-search complements it with instant ranked lookup across all sessions, without exporting anything.

License

MIT — see LICENSE.