fix(devkitd): parse ps rows with padded columns on Linux

[codeaholicguy]

Summary

Fixes #371 — ai-devkit agent list reported "No running agents detected." on Linux despite running agents.

Root cause. devkit-core's sweep parser split each ps -axo pid=,ppid=,tty=,command= row with splitn(4, char::is_whitespace), which splits on each whitespace character. Linux procps right-aligns/pads columns with runs of spaces (" 61524 60968 pts/1 claude"), producing empty fields that fail i64::parse — so every row was dropped and the daemon's agent.list was permanently empty. The CLI then treated that empty-but-non-null answer as authoritative and never ran its local adapters. The existing PS_FIXTURE test never caught this because its padding sits between tty and command (where the remainder absorbs it), not between the numeric columns.

Changes

• rust/crates/devkit-core/src/discover.rs: new split_ps_row splits pid/ppid/tty on whitespace runs and keeps the remainder as command verbatim. sweep() now tracing::warn!s when it parses 0 processes from non-empty ps output (the issue's suggested tripwire).
• packages/agent-manager/src/AgentManager.ts (defense in depth): only a non-empty daemon answer is authoritative. An empty answer falls through to local adapters; if local detection finds agents, the local result wins and a one-line stderr warning is emitted (a daemon whose sweep yields nothing while agents exist is broken, per fix(devkitd): ps parser drops all rows on Linux (padded columns), so agent list shows no agents #371). Daemon-absent/throw behavior is unchanged.
• packages/daemon-client/src/__tests__/daemon-sweep.test.ts (new): end-to-end coverage at the daemon boundary — spawns the real devkitd with a stub ps on PATH printing padded procps rows (the issue's own repro technique) and asserts agent.list reports them.
• .github/workflows/ci.yml: the rust job now runs cargo test --workspace (it previously ran only fmt + clippy — no Rust tests ran in CI at all), and the npm test job now builds devkitd so the daemon-client integration tests actually execute instead of skipping via describe.runIf(bin != null). Why CI never caught this: no Rust cargo test step, no devkitd binary in the JS test job, and no agent list e2e existed.

Issue test matrix — every row must yield procs=1

stub ps line	before	after
61524 60968 pts/1 claude	1	1
61524 60968 pts/1 claude	1	1
61524 60968 pts/1 claude	0	1
61524 60968 pts/1 claude	0	1
61524 60968 pts/1 claude	0	1

Validation

• cargo fmt --all -- --check, cargo clippy --workspace --all-targets -- -D warnings, cargo test --workspace — all green (94 tests).
• npm run lint — clean; npm test — 8 projects green incl. new daemon-sweep integration test (verified failing-then-passing TDD on both Rust and CLI sides); npm run test:e2e — 42 passed.

Test plan

• Issue matrix unit test: all five rows parse to procs=1
• Command remainder preserves inner spacing (claude --resume abc123)
• CLI falls back to local adapters on empty-daemon/local-nonempty disagreement (+warning); silent when both empty
• Real devkitd + stub ps e2e: agent.list returns the padded-row agents
• CI: rust job runs cargo test; npm job builds devkitd so daemon tests don't skip

Closes #371

Generated with Devin