SessionFS starts with portable AI coding sessions: capture the work, resume it in another tool, hand it to a teammate or agent, then distil it into shared memory and audit.
The foundation is the portable .sfs session: captured from local tools, resumable across runtimes, shareable with teammates, and compiled into knowledge.
Background daemon captures every session from 9 local tools — Claude Code, Codex, Gemini, Copilot, Cursor, Amp, Cline, Roo Code, Kilo Code — into the portable .sfs format. Zero developer behaviour change.
Start in Claude Code, resume in Codex. Auto-launch + full transcript via --append-system-prompt-file. Four-tool resume; 9-tool capture.
Sessions distil into claims, decisions, patterns, and bugs. Auto-compiled wiki with backlinks. Agents read AND write — the next session inherits everything the last one learned.
Shared architecture, conventions, and key decisions — loaded into every AI session automatically. Explain once, never repeat.
Rules and personas follow every tool, every session, every teammate.
One canonical rules document, compiled to CLAUDE.md, codex.md, .cursorrules, copilot-instructions.md, and GEMINI.md. Resume syncs the target tool's rules file before launch.
Portable AI roles per project (atlas, sentinel, scribe, …). ASCII name regex, soft-delete preserves history, server-side guard refuses delete while non-terminal tickets reference. Pro+ tier.
Every captured manifest records which persona + ticket + rules version was active. Sync propagates the tags into sessions.persona_name + sessions.ticket_id, so every audit query can trace any line of code back to the agent identity that produced it.
Handoffs move captured sessions between people and agents. Tickets carry the work, personas claim it, and AgentRuns enforce CI policy on the way out.
Server-enforced FSM (suggested → open → in_progress → blocked → review → done). Acceptance criteria, dependencies, comments. Active-ticket bundle tags every captured session with the assigned persona. Team+ tier.
Auditable execution record for one persona run, with CI-friendly policy enforcement (severity × fail_on → exit_code). GitHub Actions + GitLab CI workflows shipped, hardened across 10 rounds of cross-agent review.
Cloud-agent-ready via integration docs: AWS Bedrock action group, Google Vertex AI function-calling schema, and a closed-OPERATIONS dispatch table for custom API clients. Same personas + tickets, hosted-LLM-side.
Hand off a session to a teammate with full context. Session data copies to recipient — full conversation, workspace state, git snapshot, status stepper.
Built in. The foundation under everything else — verify, sanitise, audit.
Cross-references every AI claim against actual tool outputs. When the AI says "test passes" but exit code is 1 — that's a critical contradiction, flagged with full evidence + CWE mapping.
Server-side scan of every archive file. 14 PHI patterns + 19 secret patterns. BLOCK / REDACT / WARN modes. Org-level policy via settings JSON; dashboard surfaces findings per-session.
Every captured session carries instruction provenance, persona, ticket, and the AgentRun that signed off. The dashboard's Agent Runs tab is a queryable audit log; severity + policy verdict + findings JSON are all stored per run.
Start with local memory, then add shared knowledge, tickets, and governance as your team adopts agents.