Open source Pre-release · Apache-2.0

Context your agents can check.

When an agent session ends, what it learned usually ends with it. Alethic keeps unfinished work in your repository and compiles it into a short, cited briefing for whoever continues.

View on GitHub

git clone https://github.com/shubhambaid/alethic.git

alethic resume --budget 2500

Invalidate sessions after password reset

Task: task-invalidate-sessions-after-password-reset
Handed from: codex to claude-code
Code since: unchanged

Verified behavior and checks run

node --test failed · exit 1 at 87e0b47 rcpt-node-test-20260914t094000z ⚠ unverified

Failed approaches

Delete session rows on reset. Failed because refresh() reads the refresh cache first, so cached sessions keep working. cp-invalidate-sessions-…-094000z ⚠ unverified

Next safe action

Bump users.tokenVersion on reset and compare it in refresh(). cp-invalidate-sessions-…-094000z

## Verified behavior and checks run
- `node --test` failed (exit 1) at 87e0b47 (code unchanged since),
  2026-09-14T09:40:00Z; (receipt rcpt-node-test-20260914t094000z) ⚠ unverified

## Failed approaches
- Delete session rows on reset. Failed because: refresh() reads the
  refresh cache first, so cached sessions keep working.
  [cp-invalidate-sessions-after-password-reset-20260914t094000z] ⚠ unverified

## Next safe action
- Bump users.tokenVersion on reset and compare it in refresh().
  [cp-invalidate-sessions-after-password-reset-20260914t094000z]

Principles

Small files. Honest labels. No service.

Plain files in Git

One YAML record per file under .alethic/. Reviewed in pull requests, merged like code. No account or database.

Anchored to code

Records fingerprint the files they describe. When that code changes, they're flagged may be stale, even after squash merges.

Honest trust

Agent claims are marked unverified. A human confirmation is an attribution tied to the exact text, never shown as authenticated.

Agent-neutral

A CLI any agent can run, short blocks for AGENTS.md, CLAUDE.md, and GEMINI.md, plus an MCP server.

How it works

Three commands carry work across sessions.

Codex tries an approach and it fails. Claude Code resumes from the briefing and fixes it. Gemini runs the suite and closes the task. None of them sees the others' chats.

1 Set up the repositoryCreate .alethic/ and tell agents to start from a briefing.

Shell

alethic init
alethic render agents-md --write

2 Leave a checkpointBefore stopping, record what ran, what failed, and the next safe step.

Shell

alethic receipt run -- npm test
alethic checkpoint create \
  --failed "Delete session rows::the refresh cache still serves them" \
  --next "Compare a token version on refresh"

3 Resume anywhereA fresh session, any agent, gets a cited briefing within a budget.

Shell

alethic resume --budget 2500
alethic validate

Alethic isn't on npm yet. Clone the repository, run npm install && npm run build && npm link, and try the three-agent demo with examples/demo/run-demo.sh.

Dashboard

Every session, record, and handoff. One view.

A read-only page on your machine. Here a code edit contradicts an earlier decision, so the decision is flagged and its evidence is compared file by file.

From alētheia, Greek for unconcealment

Nothing learned should stay hidden in a closed session.

Get started on GitHub Read the specification