Agent/daemon: hold the unlocked vault, serve just-in-time child-scoped secret injection

[arimxyer]

Summary

Add an optional background agent/daemon that holds the unlocked vault in memory and serves just-in-time, child-scoped secret injection to exec / export / run. This is the foundational capability under the ergonomics work in #115 — the piece that actually attacks secret lifetime / blast-radius, which none of the #115 flags fix on their own.

Split out of #115 deliberately: #115 is the ergonomic layer (how mappings are declared and delivered); this is the substrate (where the unlocked key lives and how long a materialized secret survives). They compose; this one is the dependency.

Problem

Two gaps in today's model:

1. Lifetime. exec is well-scoped — the secret dies with the child. But the export-into-a-shell pattern proposed in Env-injection ergonomics: export/direnv, run --env-file/inject, transforms, mapping manifest #115 lives for the whole shell session (hours), /proc/<pid>/environ-readable the entire time. Shipping export alone trades a file leak for a long-lived in-memory leak.
2. Unlock friction. Every exec / get opens and decrypts the vault, so just-in-time fetching either means repeated master-password/keychain unlocks, or pushing users toward keeping things unlocked some other ad-hoc way.

A daemon resolves both: unlock once, then every fetch is fast, just-in-time, and scoped to the consuming process.

Proposal

pass-cli agent — a background process that:

• Holds the derived key / unlocked vault in memory only (never on disk).
• Serves exec / export / run (and get) over a local socket so they fetch on demand instead of opening the vault themselves.
• When the daemon isn't running, all commands fall back to today's direct-open behavior — the daemon is an optimization/safety layer, not a hard dependency.

Composed with #115: single declared source (manifest) → daemon serves JIT → materialization scoped to the child = one auditable source, zero persistent plaintext, minimal lifetime.

Design surface (the parts that need decisions)

• IPC & auth. Unix domain socket (Windows named pipe), perms 0600, plus peer-credential checks (SO_PEERCRED / getpeereid) so only same-user processes can talk to it. Define the request/response protocol (request a mapping set, return values, never log them).
• Lifecycle / auto-lock. Explicit pass-cli agent (and agent stop / lock), idle timeout and max-TTL re-lock, behavior across multiple shells, and what happens on logout/suspend.
• Unlock & key handling. Unlock via master password or the existing OS keychain integration; cache the derived key in locked memory (mlock, disable core dumps) to keep it out of swap/crash dumps.
• Honest security model. The daemon's own memory is /proc/<pid>/mem-readable by the same user — the same ceiling as ssh-agent / gpg-agent. Document it; don't oversell isolation. This is strictly better than long-lived shell env or plaintext files, not a sandbox.
• Integration. How it interacts with keychain (could supply the unlock) and sync; whether exec/export/run auto-detect and prefer the socket.
• Platform service. Optional systemd --user unit / launchd agent / Windows service, or pure run-on-demand.

Prior art

ssh-agent, gpg-agent, op daemon / op biometric session, aws-vault's server mode. The lifetime/scope model is well-trodden; this issue is about adopting it for pass-cli's vault.

Non-goals

Does not remove the consumer-side plaintext-at-boundary ceiling (an app reading process.env still gets a plaintext value at the moment of use — see #115's "hard ceiling"). This issue minimizes how long and how broadly that plaintext exists, not whether it exists at the boundary.

Relationship

• Substrate for the ergonomics layer in Env-injection ergonomics: export/direnv, run --env-file/inject, transforms, mapping manifest #115 (export / run --env-file / manifest become thin clients of the agent).
• Implement independently; Env-injection ergonomics: export/direnv, run --env-file/inject, transforms, mapping manifest #115's flags work without it (direct-open fallback) and get safer with it.

Filed alongside #115 from a real session wiring OpenTelemetry/Langfuse export into an eve agent; a separate Claude Code session may pick up the implementation.