test(mobile): allow a locally built CLI in the E2E remote-cli harness

[iscekic]

What

apps/mobile/e2e/remote-cli.sh gains an opt-in KILO_CLI_BIN mode: when set to a locally built kilo binary, prepare_env skips the npm install of @kilocode/cli, prepends the binary's own directory to the PATH exported into .cli-env (the built CLI keeps sibling assets next to the executable), and fails loudly if the path is not an executable file.

Under the same opt-in, the generated .cli-env now also carries the flags the session-tail E2E depends on:

• KILO_NO_DAEMON=1 — the TUI must not attach to a machine daemon, so the embedded worker owns the ingest queue (otherwise the verification is a false pass/fail);
• KILO_DEBUG_SESSION_INGEST=true — the ingest sync / ingest flush ok lines are the primary evidence channel;
• KILO_REMOTE=1 — without it the session never mirrors live to the device.

They must be emitted: prepare_env rewrites .cli-env on every prepare/start, silently discarding anything hand-appended (hit for real during bug reproduction). With KILO_CLI_BIN unset, behaviour and the generated env file are byte-for-byte identical to before.

Why this PR exists at all

The fix under verification is CLI-side: Kilo-Org/kilocode#12545 (flush the session ingest tail on shutdown). A CLI-side fix cannot be exercised on device while the harness can only ever install a published CLI — this harness change is the smallest thing that makes that verification reachable, and the only reason this cloud PR exists. The batch instruction said one PR from this worktree; the two-PR deviation is deliberate and stated in both PR bodies.

Checks

pnpm format && pnpm typecheck && pnpm lint && pnpm check:unused in apps/mobile, plus git diff --check — all green. The script was additionally validated by dry-running the emitted-env branch with a stub KILO_CLI_BIN (default path unchanged). Live verification against the stack runs with the companion CLI PR's on-device E2E; results will be posted as a follow-up comment.

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Executive Summary

The opt-in KILO_CLI_BIN path in remote-cli.sh is correctly gated, validates the binary before use, and leaves default (non-opt-in) behavior byte-for-byte unchanged, matching the PR description.

Files Reviewed (1 files)

• apps/mobile/e2e/remote-cli.sh

---

_{Reviewed by claude-sonnet-5 · Input: 18 · Output: 7.3K · Cached: 355.4K}

_{Review guidance: REVIEW.md from base branch main}

[iscekic]

(bot) Standin review — round 1: No findings. (review of record; head reviewed: 8dd6e45580315dc6933a816f3c7fd30f213d9942, full diff vs main)

The standin reviewer (reviewer-bot substitute per the workflow override) statically traced the full diff against plan §5, set -euo pipefail semantics, quoting, PATH precedence, the heredoc/append mechanics, and default-path equivalence, and returned no actionable findings. The loop is converged on the exact latest head.

Residual testing risks recorded (no action): the new branches' first live exercise is the orchestrator-run on-device E2E by design; a symlinked KILO_CLI_BIN would resolve the wrong directory — the E2E handoff passes the real built-binary path and records its version stamp as the guard; the bad-path validation ordering (after the token mint) is cosmetic and pre-existing; the header says "absolute path" while relative paths also work (more permissive, no failure mode).

The on-device E2E for the companion fix (Kilo-Org/kilocode#12545) exercises this harness end to end; results will be posted here and there as (bot) comments.

[iscekic]

(bot) Harness exercised end to end by the companion on-device E2E: PASS. The acceptance run for Kilo-Org/kilocode#12545 used this branch's KILO_CLI_BIN mode exactly as designed: prepare_env skipped the npm install, put the locally built binary's directory first on PATH (kilo --version resolved to 0.0.0-fix-cli-ingest-flush-on-shutdown-202607262141), and the emitted KILO_NO_DAEMON=1 / KILO_DEBUG_SESSION_INGEST=true / KILO_REMOTE=1 held across every prepare/start rewrite of .cli-env — the TUI footer showed ◆ Remote and the version stamp on every attempt, and the ingest sync / ingest flush ok lines were the primary evidence channel.

The full E2E matrix (11/11: /exit ×3, SIGINT ×3, SIGHUP ×3, delayed-/exit control ×2, plus the coalescing check) is posted on Kilo-Org/kilocode#12545. The default developer flow (no KILO_CLI_BIN) was left byte-for-byte identical and was not disturbed by the run.