[Bug]: Pi extension breaks with in-process subagents

[noctuid]

Which fff frontend?

Other / multiple

has logs

No response

Description

This will happen for any subagents plugin that uses createAgentSession instead of completely separate pi processes. (My personal subagents extension works this way, and there are many significant benefits to doing things this way, e.g. subagents can still use UI tools like permissions or ask/question like in other harnesses).

fff-node can't open the same db environment twice in one process, so on subagent start you will get an error message because FileFinder.create() runs against the same frecency/history db again.

Potential fixes:

1. fff-node: allow multiple FileFinder instances in the same process to share one db environment
2. pi-fff: share a single FileFinder per process (e.g. a globalThis/Symbol.for singleton) and/or fall back to an isolated DB path when the environment is already open; this is how pi-pretty's fff implementation works fine with in-process subagents

This is related to #700, but the fix there does not handle createAgentSession

[gustav-fff]

[triage-bot] REPRODUCED: Reproduced on origin/main. Branch: triage-bot/issue-760.

Root cause is heed 0.22, not fff. OPENED_ENV is a process-global registry (env_open_options.rs:405-422): opening the same canonical LMDB path twice in one process returns Error::EnvAlreadyOpened. pi-fff's main finder always opens the fixed ~/.pi/agent/fff/{frecency,history} (packages/pi-fff/src/index.ts:313-321), so a second FileFinder.create() in the same process (in-process createAgentSession subagent) fails. Aux finders already dodge this by opening no DB (#700).

Reproduced at the Rust layer (crates/fff-core/src/dbs/frecency.rs):

let a = FrecencyTracker::open(dir); // ok
let b = FrecencyTracker::open(dir); // Err
// SECOND OPEN ERROR: Failed to open frecency database env: environment
// already open in this program; close it to be able to open it again with
// different options

Going with reporter option 2 (pi-fff singleton/fallback) over option 1: making fff-node share one env across instances is a core-DB change with broad blast radius. Instead, when the main finder fails with EnvAlreadyOpened, fall back to a transient no-DB finder — same precedent as aux finders (#700).

Working on fix on triage-bot/issue-760.
Honk-Honk 🪿