Alejandro Gutiérrez 3fab89c61a fix(cli): a resumed session keeps its registration and its key (bug1, bug3) ...

bug1 — the registration belonged to the launch WRAPPER (its pid, a token
in its tmpdir). When claude outlived it (restart, re-exec, resume) the
reaper dropped the entry, the token vanished, and every send from that
session went out silently under the MEMBER key; replies to it fanned out
to all of the member's sessions.

- services/session/reattach.ts: ensureSessionRegistered re-registers the
  session with its persisted keypair, a fresh attestation and claude's
  pid (nearest `claude` ancestor). Called by the MCP at startup (moves
  the anchor off the wrapper), by `send` when the session is unknown
  (auto-heal), and by the new `claudemesh session reattach`.
- The session token lives next to the keypair
  (~/.claudemesh/sessions/<mesh>/<stem>.token) and survives the wrapper;
  token readers fall back to it via CLAUDEMESH_SESSION_ID (+ the new
  CLAUDEMESH_MESH_SLUG, or the mesh holding the keypair file).
- `send` inside a session that can't be re-attached refuses unless
  --as-member; the daemon answers 401 session_unknown to a send whose
  session token it doesn't know.
- The registry's absolute 24 h TTL is gone (liveness = pid + start time).

bug3 — sessions reappeared under a new pubkey:
- launch keys the keypair, the registration and the token on ONE id (it
  used to register a random UUID next to a keypair keyed elsewhere).
- non-UUID session ids persist under a hashed stem instead of a
  throwaway keypair per call.
- the daemon's boot reload is load-only — it never mints a key a live
  session didn't have.
- a same-token, same-key re-register only moves the pid anchor, so the
  MCP's startup re-attach doesn't close/reopen the session WS.

Tests: session-reattach.test.ts (2/5 fail on the old registry),
keypair-store.test.ts updated to the new non-UUID rule.
Spec: .artifacts/specs/2026-09-26-session-identity-and-routing.md §2, §4

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

2026-09-26 21:29:03 +01:00

3.7 KiB

Raw Permalink Blame History

View Raw