code-recall 0.2.2

Semantic search across local coding-agent sessions. Find past conversations by intent, not just keywords.

code-recall

Find the coding-agent session you vaguely remember.

Search, inspect, chat with, and resume your local Claude Code and Codex history from one fast terminal UI.

pip install 'code-recall[all]'
code-recall "stripe webhook signature"

Why

Coding agents leave behind useful local transcripts, but built-in resume pickers are optimized for recent sessions. After a few weeks, you remember the shape of the work, not the exact title:

• "the session where we fixed Stripe webhook signatures"
• "the dashboard query that got slow after adding invoices"
• "the Docker healthcheck that broke deploys"
• "the branch where the OAuth callback tests were added"

code-recall turns those transcripts into a searchable memory layer. It indexes Claude Code and Codex sessions, ranks results with keyword search, semantic search, reranking, and graph signals, then lets you jump back into the right project with the right provider command.

What It Does

• Search Claude Code and Codex sessions together.
• See exactly why a session matched your query.
• Filter and search by provider, project, branch, file, or command.
• Resume in the original project directory with claude --resume ... or codex resume ....
• Chat with a selected transcript from the TUI.
• Keep the index fresh with quick incremental indexing and Claude Code session-end hooks.
• Stay local-first: your SQLite index and transcript reads stay on your machine.

Screenshots

All screenshots in this README are generated from synthetic demo data using scripts/generate_demo_assets.py. They do not come from a real work index.

Search across providers	Understand why it matched	Chat with a transcript

For a short GIF or product demo, a good sanitized sequence is:

1. Search stripe webhook signature.
2. Highlight the top result and show the provider badge, project, branch, and why: line.
3. Press 2 for the Why tab and show matched evidence.
4. Press p to cycle Claude Code and Codex provider scopes.
5. Press 4 to show related sessions that touched the same files.
6. Press 5 and ask: What changed and what should I verify before resuming?

Install

pip

# Everything: TUI, semantic search, reranking, sqlite-vec
pip install 'code-recall[all]'

# Smaller installs
pip install code-recall                  # keyword search only
pip install 'code-recall[semantic]'      # + embeddings and reranking
pip install 'code-recall[tui]'           # + interactive terminal UI

uv

uv tool install code-recall --with textual --with fastembed --with sqlite-vec

From source

git clone https://github.com/lupuletic/code-recall
cd code-recall
uv venv && source .venv/bin/activate
uv pip install -e ".[all]"

Usage

# Interactive TUI - starts with recent sessions; type to search
code-recall

# Direct search
code-recall "stripe webhook signature"
code-recall "slow account dashboard query"
code-recall "docker healthcheck deploy"

# Search graph metadata directly
code-recall "file:src/webhooks/stripe.ts"
code-recall "cmd:pnpm test"
code-recall "branch:fix/stripe-webhook-signature"

# Combine graph prefixes with regular text
code-recall "file:schema.prisma migration"
code-recall "branch:main release workflow"

# AI investigation over indexed evidence + read-only transcript access
code-recall ask "where did we fix webhook signature verification?"
code-recall investigate "summarize sessions related to checkout retries"

# Index controls
code-recall index                  # indexes Claude Code + Codex by default
code-recall index --no-codex       # Claude Code only for this run

# Output formats
code-recall "query" --no-tui       # plain text
code-recall "query" --json         # JSON for scripting

# Updates
code-recall update
code-recall update --yes

TUI Controls

Key	Action
Type or /	Focus search and search as you type
Tab / Shift+Tab	Move between search, results, and detail panes
Down / Up	Navigate results
Enter / Right	Focus the detail pane for the selected result
1-5	Switch detail tabs: Overview, Why, Activity, Related, AI
p	Cycle provider scope: all, Claude Code, Codex
r	Resume selected session
c	Copy selected resume command
o	Open the selected project folder
Ctrl+A	Ask AI to investigate the current query/results
Ctrl+K	Open the command palette
Ctrl+O / f	Open settings
?	Show help
Esc	Close detail/help or quit

With an empty search box, the TUI shows your most recent indexed sessions. Results include provider badges, project/branch context, activity, score quality, file/command counts, and a short why: line. The detail pane has Overview, Why, Activity, Related, and AI tabs. On narrow terminals, the detail pane opens as a focused view.

Search

The default search mode is hybrid.

code-recall config search_mode hybrid   # keyword + semantic + reranking
code-recall config search_mode keyword  # FTS only, fastest, no semantic deps
code-recall config search_mode llm      # + assistant reranking, slower
code-recall config limit 20

The ranking pipeline combines:

Stage	What it does
FTS5 keyword search	Fast BM25 search across summaries, prompts, messages, and project paths.
Structured prefixes	Exact graph lookups for file:, cmd:, and branch: queries.
Semantic search	Local embeddings over conversation chunks when semantic extras are installed.
Knowledge graph	Connects sessions through shared files, commands, branches, and project context.
Cross-encoder rerank	Improves top-result precision for hybrid mode.
Optional LLM rerank	Uses an installed Claude Code or Codex CLI when search_mode = llm.

Claude Code subagent content is bubbled up to parent sessions during indexing, so work done by delegated agents remains searchable from the main session.

Providers

code-recall currently indexes:

Provider	Source	Resume command
Claude Code	~/.claude/projects/*/*.jsonl	claude --resume <id>
Codex	~/.codex/sessions/**/*.jsonl and local Codex state	codex resume <id>

AI features use the matching assistant when possible: Claude Code sessions prefer claude -p, Codex sessions prefer codex exec. If the matching CLI is not installed, code-recall falls back to the other supported CLI when available and shows the assistant provider in the AI tab.

Index Freshness

code-recall keeps the index fresh in two ways:

• Every search/TUI startup runs a quick incremental index over Claude Code and Codex sources before searching.
• On first run or code-recall install-hooks, code-recall installs a Claude Code SessionEnd hook that runs code-recall index --quiet when Claude Code sessions end.

Codex sessions are indexed on the next incremental index/search run. There is no Codex session-end hook installed today because Codex does not expose the same SessionEnd hook surface in this repo's integration path.

Privacy

code-recall is designed for local developer recall:

• The index is a local SQLite database under your app data directory.
• Keyword and graph search do not require network calls.
• Semantic search uses local embeddings.
• AI investigation and transcript chat are explicit actions, not automatic background calls.
• README screenshots are synthetic and reproducible from scripts/generate_demo_assets.py.

Architecture

graph TD
    subgraph "Data Sources"
        CLAUDE_JSONL["Claude Code JSONL"]
        CODEX_JSONL["Codex sessions"]
        CODEX_DB["Codex local state"]
    end

    subgraph "Indexer"
        PARSE["Parse provider sessions"]
        CHUNK["Chunk conversations"]
        ENRICH["Enrich parent sessions"]
        FTS_IDX["FTS5 keyword index"]
        VEC_IDX["Local embedding index"]
        KG_IDX["Knowledge graph"]
    end

    subgraph "Search"
        QUERY["User query"]
        PREFIX["file: cmd: branch:"]
        PIPELINE["Hybrid ranking"]
        RESULTS["Ranked sessions"]
    end

    CLAUDE_JSONL --> PARSE
    CODEX_JSONL --> PARSE
    CODEX_DB --> PARSE
    PARSE --> CHUNK --> VEC_IDX
    PARSE --> ENRICH --> FTS_IDX
    PARSE --> KG_IDX
    QUERY --> PREFIX --> PIPELINE
    FTS_IDX --> PIPELINE
    VEC_IDX --> PIPELINE
    KG_IDX --> PIPELINE
    PIPELINE --> RESULTS

Search Quality

The project includes an eval runner so you can measure recall against your own local history:

code-recall eval generate ~/.code-recall/eval-cases.json 30
code-recall eval ~/.code-recall/eval-cases.json
uv run python scripts/eval_search.py ~/.code-recall/eval-cases.json --json

Example categories worth testing:

Category	Example query
Exact project names	payments api, customer portal
Vague task recall	the webhook signature bug
Technology queries	prisma dashboard query, docker healthcheck
Last-message context	what was left before release
Graph queries	file:src/webhooks/stripe.ts, cmd:pnpm test

First Run

On first run, code-recall:

1. Builds a keyword index of Claude Code and Codex sessions.
2. Shows results immediately.
3. Generates embeddings in the background when semantic extras are installed.
4. Builds graph metadata for files, commands, branches, and session chains.
5. Installs the Claude Code session-end hook if you choose to enable hooks.

Development

git clone https://github.com/lupuletic/code-recall
cd code-recall
uv venv && source .venv/bin/activate
uv pip install -e ".[all,dev]"

uv run pytest -q
uv run python scripts/check_version.py
uv run python -m compileall -q src tests

# Regenerate sanitized README screenshots
uv run python scripts/generate_demo_assets.py

# Optional: benchmark against your own populated local index
code-recall eval generate ~/.code-recall/eval-cases.json 30
CODE_RECALL_RUN_QUALITY_TESTS=1 \
CODE_RECALL_QUALITY_CASES=~/.code-recall/eval-cases.json \
uv run pytest tests/test_search_quality.py -v

Release

Releases are tag driven. CI verifies the version declarations match before publishing. The release workflow builds distributions, publishes to PyPI through Trusted Publishing, and creates or updates the matching GitHub release.

Trusted Publisher fields:

Field	Value
PyPI project name	code-recall
Owner	lupuletic
Repository name	code-recall
Workflow name	release.yml
Environment name	pypi

uv run python scripts/check_version.py
uv run pytest -q
uv run python -m compileall -q src tests
uv build --no-sources

git tag v0.2.2
git push origin main --tags

License

MIT