tests/fm-secondmate-harness.test.sh does not neutralise ambient CLAUDECODE, so the whole detect_own block fails inside a Claude Code session

[ironerumi]

What happens

tests/fm-secondmate-harness.test.sh fails for anyone running the suite from inside a Claude Code session:

not ok - unmarked shared signed-wrapper ancestry resolved 'claude', expected pi

Nothing is wrong with the harness resolution. The test reads an ambient environment variable it never neutralises.

Cause

The assertion at tests/fm-secondmate-harness.test.sh:162-163 builds a fake ps ancestry and expects bin/fm-harness.sh to resolve pi:

  got=$(PATH="$fakebin:$BASE_PATH" PI_CODING_AGENT=true "$ROOT/bin/fm-harness.sh")
  [ "$got" = pi ] || fail "unmarked shared signed-wrapper ancestry resolved '$got', expected pi"

But bin/fm-harness.sh:38 short-circuits on the Claude marker before any ancestry walk:

  [ "${CLAUDECODE:-}" = "1" ] && { echo claude; return; }

A Claude Code session exports CLAUDECODE=1, so it outranks the test's PI_CODING_AGENT=true pin and the fake ancestry is never consulted. The whole detect_own block of this file is affected, not just the one assertion that trips first.

Reproduce

At 99533c5 on any host, from inside a Claude Code session (or with CLAUDECODE=1 exported):

CLAUDECODE=1 bin/fm-test-run.sh tests/fm-secondmate-harness.test.sh

Invocation	Runs	Result
CLAUDECODE=1 bin/fm-test-run.sh tests/fm-secondmate-harness.test.sh	2	exit=1, resolved 'claude', expected pi
env -u CLAUDECODE bin/fm-test-run.sh tests/fm-secondmate-harness.test.sh	1	exit=0 (duration_ms=114185 gate_skip=false)

Expected: the suite is hermetic and passes regardless of which harness the contributor happens to be running it from.
Actual: it fails whenever the surrounding session exports CLAUDECODE=1.

Test-only counterfactual

Single variable, nothing under bin/ touched: env -u CLAUDECODE turns the file green. That isolates ambient environment as the sole cause — it is not host sensitivity and not a regression in the pi-signed adapter.

Proposed minimal fix

Scrub the harness markers at the top of the file, the same way tests/fm-session-start.test.sh already does. That file's prior art is at :408-420:

# Markers today: CLAUDECODE (claude), PI_CODING_AGENT plus FM_PI_HARNESS
# (Pi family), GROK_AGENT (grok).
    env -u CLAUDECODE -u PI_CODING_AGENT -u FM_PI_HARNESS -u GROK_AGENT \

So either an unset CLAUDECODE PI_CODING_AGENT FM_PI_HARNESS GROK_AGENT near the top of tests/fm-secondmate-harness.test.sh (every assertion in the file already pins its harness explicitly, so nothing depends on the inherited values), or env -u on each detect_own invocation.

This is the same class PR #432 ("test: isolate session-start suite from ambient harness markers", merged as 3e3dff6) already fixed for the session-start suite; tests/fm-secondmate-harness.test.sh was missed.

Existing work

Open, unmerged PR #1200 ("fix(bin): keep firstmate parsing under stock macOS Bash 3.2") adds exactly this unset line to this file as one piece of a broad sweep. Filing separately so the defect is tracked independently of whether that sweep lands; if #1200 merges as-is, this closes with it. No existing issue reports it — searched CLAUDECODE, fm-secondmate-harness, ambient environment test, test fails macOS.

Environment

• firstmate 99533c5d7d3702050e6084429dddff6ea4fe1aa0
• macOS 15.7.7, Darwin 24.6.0, arm64
• Bash 5.3.9 (Homebrew), stock /bin/bash 3.2.57
• Reproduced from inside Claude Code (ambient CLAUDECODE=1); platform-independent