Fix test flakiness

.github/workflows/ci.yml

@@ -19,7 +19,7 @@ jobs:
             launch.sh dashboard.sh harvest.sh \
             costs.sh progress.sh \
             lib/*.sh lib/drivers/*.sh \
-            tests/test_*.sh tests/test.sh
+            tests/*.sh
 
   unit-tests:
     strategy:

tests/_test_env.sh

@@ -0,0 +1,7 @@
+#!/bin/bash
+
+# Keep test behavior independent of host Git settings.
+export GIT_CONFIG_GLOBAL=/dev/null
+export GIT_CONFIG_SYSTEM=/dev/null
+export GIT_CONFIG_COUNT=0
+unset GIT_CONFIG GIT_CONFIG_PARAMETERS

tests/manual_interactive_e2e.sh

@@ -1,6 +1,9 @@
 #!/bin/bash
 set -euo pipefail
 
+# shellcheck source=_test_env.sh
+source "$(dirname "${BASH_SOURCE[0]}")/_test_env.sh"
+
 # Manual interactive E2E fixture generator.
 #
 # This is intentionally not named test_*.sh, so tests/test.sh --unit

tests/runtime_signal_trap.sh

@@ -2,6 +2,9 @@
 # shellcheck disable=SC2034
 set -uo pipefail
 
+# shellcheck source=_test_env.sh
+source "$(dirname "${BASH_SOURCE[0]}")/_test_env.sh"
+
 # tests/runtime_signal_trap.sh
 #
 # Drive the harness's emergency-push paths under real Docker.
@@ -180,6 +183,22 @@ wait_for_session() {
     return 1
 }
 
+wait_for_test_commit() {
+    local name="$1" tag="$2" timeout="${3:-30}"
+    local i subject
+    for i in $(seq 1 "$timeout"); do
+        subject=$(docker exec "$name" \
+            git -C /workspace log -1 --format='%s' 2>/dev/null || true)
+        if [ "$subject" = "test commit [${tag}]" ]; then
+            return 0
+        fi
+        sleep 1
+    done
+    echo "    timed out waiting for test commit [${tag}] in ${name}"
+    docker logs "$name" 2>&1 | tail -15 | sed 's/^/      /'
+    return 1
+}
+
 await_exit() {
     local name="$1" timeout="${2:-90}"
     local i status
@@ -205,8 +224,7 @@ count_landed() {
 echo
 echo "=== Item 3: SIGTERM mid-session ==="
 start_container "${NAMES[0]}" -e TC_TAG=sigterm-tag
-wait_for_session "${NAMES[0]}" || exit 1
-sleep 2
+wait_for_test_commit "${NAMES[0]}" sigterm-tag || exit 1
 echo "  sending docker stop -t 60..."
 docker stop -t 60 "${NAMES[0]}" >/dev/null
 EXIT_3=$(await_exit "${NAMES[0]}")
@@ -221,8 +239,7 @@ echo
 
 echo "=== Item 4: SIGINT mid-session ==="
 start_container "${NAMES[1]}" -e TC_TAG=sigint-tag
-wait_for_session "${NAMES[1]}" || exit 1
-sleep 2
+wait_for_test_commit "${NAMES[1]}" sigint-tag || exit 1
 echo "  sending docker kill --signal=SIGINT..."
 docker kill --signal=SIGINT "${NAMES[1]}" >/dev/null
 EXIT_4=$(await_exit "${NAMES[1]}")

tests/test.sh

@@ -1,6 +1,9 @@
 #!/bin/bash
 set -euo pipefail
 
+# shellcheck source=_test_env.sh
+source "$(dirname "${BASH_SOURCE[0]}")/_test_env.sh"
+
 # Smoke test: launch agents with a counting prompt, verify results.
 # Each agent N writes test-results/agent-N.txt with N*100..N*100+99.
 #

tests/test_activity_filter.sh

@@ -2,6 +2,9 @@
 # shellcheck disable=SC2034
 set -euo pipefail
 
+# shellcheck source=_test_env.sh
+source "$(dirname "${BASH_SOURCE[0]}")/_test_env.sh"
+
 # Unit tests for lib/activity-filter.sh.
 # No Docker or API key required.
 

tests/test_ci_workflows.sh

@@ -1,11 +1,16 @@
 #!/bin/bash
 set -euo pipefail
 
+# shellcheck source=_test_env.sh
+source "$(dirname "${BASH_SOURCE[0]}")/_test_env.sh"
+
 # Unit tests for GitHub Actions workflow shape.
 
 PASS=0
 FAIL=0
-REPO_ROOT="$(cd "$(dirname "$0")/.." && pwd)"
+TESTS_DIR="$(cd "$(dirname "$0")" && pwd)"
+REPO_ROOT="$(cd "$TESTS_DIR/.." && pwd)"
+CI_YML="$REPO_ROOT/.github/workflows/ci.yml"
 INTEGRATION_YML="$REPO_ROOT/.github/workflows/integration.yml"
 
 assert_eq() {
@@ -32,6 +37,23 @@ assert_eq "workflow has no full-matrix job" "0" \
 assert_eq "workflow does not run --all" "0" \
     "$(grep -cF './tests/test.sh --all' "$INTEGRATION_YML" || true)"
 
+echo ""
+echo "=== 2. Shared test environment ==="
+
+missing_env=""
+for test_file in "$TESTS_DIR"/*.sh; do
+    [ "$test_file" = "$TESTS_DIR/_test_env.sh" ] && continue
+    if ! grep -qF \
+            'source "$(dirname "${BASH_SOURCE[0]}")/_test_env.sh"' \
+            "$test_file"; then
+        missing_env="${missing_env} $(basename "$test_file")"
+    fi
+done
+
+assert_eq "all test scripts source the shared environment" "" "$missing_env"
+assert_eq "CI checks every test script" "1" \
+    "$(grep -cF 'tests/*.sh' "$CI_YML")"
+
 echo ""
 echo "==============================="
 echo "  ${PASS} passed, ${FAIL} failed"

tests/test_config.sh

@@ -2,6 +2,9 @@
 # shellcheck disable=SC2034
 set -euo pipefail
 
+# shellcheck source=_test_env.sh
+source "$(dirname "${BASH_SOURCE[0]}")/_test_env.sh"
+
 # Unit tests for swarm.json config parsing.
 # No Docker or API key required -- validates jq expressions only.
 

tests/test_costs.sh

@@ -1,6 +1,9 @@
 #!/bin/bash
 set -euo pipefail
 
+# shellcheck source=_test_env.sh
+source "$(dirname "${BASH_SOURCE[0]}")/_test_env.sh"
+
 # Unit tests for costs.sh aggregation and formatting logic.
 # No Docker or API key required.
 

tests/test_dashboard.sh

@@ -2,6 +2,9 @@
 # shellcheck disable=SC2034,SC2016
 set -euo pipefail
 
+# shellcheck source=_test_env.sh
+source "$(dirname "${BASH_SOURCE[0]}")/_test_env.sh"
+
 # Unit tests for dashboard.sh helper functions:
 #   format_model, truncate_str, and column layout (with optional Tag).
 # No Docker or API key required.

tests/test_drivers.sh

@@ -2,6 +2,9 @@
 # shellcheck disable=SC2034
 set -euo pipefail
 
+# shellcheck source=_test_env.sh
+source "$(dirname "${BASH_SOURCE[0]}")/_test_env.sh"
+
 # Unit tests for the driver role interface.
 # No Docker or API key required.
 

tests/test_format.sh

@@ -1,6 +1,9 @@
 #!/bin/bash
 set -euo pipefail
 
+# shellcheck source=_test_env.sh
+source "$(dirname "${BASH_SOURCE[0]}")/_test_env.sh"
+
 # Unit tests for formatting functions used by dashboard.sh and costs.sh.
 # No Docker or API key required.
 

tests/test_harness.sh

@@ -2,6 +2,9 @@
 # shellcheck disable=SC2034
 set -euo pipefail
 
+# shellcheck source=_test_env.sh
+source "$(dirname "${BASH_SOURCE[0]}")/_test_env.sh"
+
 # Unit tests for harness.sh stat extraction and logic.
 # No Docker or API key required.
 
@@ -404,11 +407,16 @@ hlog_pipe() {
 }
 
 AGENT_ID=3
-OUT=$(hlog "test message")
+OUT=$(
+    date() {
+        printf '%s\n' "12:34:56"
+    }
+    hlog "test message"
+)
 PLAIN=$(echo "$OUT" | strip_ansi)
 
 assert_contains "hlog timestamp" \
-    "$(date +%H:%M:%S)" "$PLAIN"
+    "12:34:56 harness[3] test message" "$PLAIN"
 assert_contains "hlog prefix" "harness[3]" "$PLAIN"
 assert_contains "hlog body" "test message" "$PLAIN"
 
@@ -644,6 +652,12 @@ source "$TESTS_DIR/../lib/signing.sh"
 # Exercises configure_git_signing from lib/signing.sh against
 # a sandboxed $HOME so --global writes land in a scratch dir
 # instead of the tester's real ~/.gitconfig.
+with_sandboxed_git_config() {
+    local sandbox="$1"
+    shift
+    HOME="$sandbox" GIT_CONFIG_GLOBAL="$sandbox/.gitconfig" "$@"
+}
+
 run_signing_config() {
     local create_key="$1"
     local sandbox="$TMPDIR/sign-sandbox-$$-${RANDOM}"
@@ -655,12 +669,14 @@ run_signing_config() {
         touch "$key_path"
     fi
 
-    HOME="$sandbox" configure_git_signing "$key_path" "$dst_path"
+    with_sandboxed_git_config "$sandbox" \
+        configure_git_signing "$key_path" "$dst_path"
 
     local format gpgsign
-    format=$(HOME="$sandbox" git config --global --get gpg.format \
-        2>/dev/null || echo "none")
-    gpgsign=$(HOME="$sandbox" git config --global --get commit.gpgsign)
+    format=$(with_sandboxed_git_config "$sandbox" \
+        git config --global --get gpg.format 2>/dev/null || echo "none")
+    gpgsign=$(with_sandboxed_git_config "$sandbox" \
+        git config --global --get commit.gpgsign)
     echo "${format}|${gpgsign}"
     rm -rf "$sandbox"
 }
@@ -683,7 +699,8 @@ copy_src="$copy_sandbox/src_key"
 copy_dst="$copy_sandbox/dst_key"
 echo "fake-key-bytes" > "$copy_src"
 chmod 644 "$copy_src"
-HOME="$copy_sandbox" configure_git_signing "$copy_src" "$copy_dst"
+with_sandboxed_git_config "$copy_sandbox" \
+    configure_git_signing "$copy_src" "$copy_dst"
 assert_eq "key copied to scratch location" \
     "yes" \
     "$([ -f "$copy_dst" ] && echo yes || echo no)"
@@ -693,7 +710,8 @@ assert_eq "key copy has 0600 perms" \
         || stat -f '%Lp' "$copy_dst" 2>/dev/null)"
 assert_eq "user.signingkey points at copy" \
     "$copy_dst" \
-    "$(HOME="$copy_sandbox" git config --global --get user.signingkey)"
+    "$(with_sandboxed_git_config "$copy_sandbox" \
+        git config --global --get user.signingkey)"
 assert_eq "no swarm key written under \$HOME" \
     "no" \
     "$([ -e "$copy_sandbox/.ssh/swarm-signing-key" ] && echo yes || echo no)"
@@ -709,26 +727,28 @@ fail_src="$fail_sandbox/src_key"
 fail_dst="$fail_sandbox/no/such/dir/dst_key"
 touch "$fail_src"
 fail_rc=0
-HOME="$fail_sandbox" configure_git_signing "$fail_src" "$fail_dst" \
-    >/dev/null 2>&1 || fail_rc=$?
+with_sandboxed_git_config "$fail_sandbox" \
+    configure_git_signing "$fail_src" "$fail_dst" \
+        >/dev/null 2>&1 || fail_rc=$?
 assert_eq "install failure -> non-zero return" \
     "1" \
     "$fail_rc"
 assert_eq "install failure -> user.signingkey not set" \
     "" \
-    "$(HOME="$fail_sandbox" git config --global --get user.signingkey \
-        2>/dev/null)"
+    "$(with_sandboxed_git_config "$fail_sandbox" \
+        git config --global --get user.signingkey 2>/dev/null)"
 rm -rf "$fail_sandbox"
 
 # Defaults to /etc/swarm/signing_key when called with no arg;
 # verify by pointing HOME at a sandbox where that path doesn't
 # exist and expecting the "absent" branch.
 default_path_sandbox="$TMPDIR/sign-default-$$-${RANDOM}"
 mkdir -p "$default_path_sandbox"
-HOME="$default_path_sandbox" configure_git_signing
+with_sandboxed_git_config "$default_path_sandbox" configure_git_signing
 assert_eq "default key path absent -> gpgsign false" \
     "false" \
-    "$(HOME="$default_path_sandbox" git config --global --get commit.gpgsign)"
+    "$(with_sandboxed_git_config "$default_path_sandbox" \
+        git config --global --get commit.gpgsign)"
 rm -rf "$default_path_sandbox"
 
 # ============================================================

tests/test_harvest.sh

@@ -1,6 +1,9 @@
 #!/bin/bash
 set -euo pipefail
 
+# shellcheck source=_test_env.sh
+source "$(dirname "${BASH_SOURCE[0]}")/_test_env.sh"
+
 # Unit tests for harvest.sh logic.
 # Uses local git repos (no Docker or API key required).
 

tests/test_launch.sh

@@ -2,15 +2,12 @@
 # shellcheck disable=SC2034
 set -euo pipefail
 
+# shellcheck source=_test_env.sh
+source "$(dirname "${BASH_SOURCE[0]}")/_test_env.sh"
+
 # Unit tests for launch.sh parsing logic.
 # No Docker or API key required.
 
-# Isolate from host gitconfig (signing keys, hooks, templates).
-# Several tests build scratch repos and run `git commit` /
-# `commit-tree`;
-export GIT_CONFIG_GLOBAL=/dev/null
-export GIT_CONFIG_SYSTEM=/dev/null
-
 PASS=0
 FAIL=0
 TMPDIR=$(mktemp -d)

tests/test_locale.sh

@@ -1,6 +1,9 @@
 #!/bin/bash
 set -euo pipefail
 
+# shellcheck source=_test_env.sh
+source "$(dirname "${BASH_SOURCE[0]}")/_test_env.sh"
+
 # Guard against the regression reported in the bug:
 #   dashboard.sh: line 207: printf: 52.185737173...: invalid number
 # on systems using `,` as the decimal separator.  Pin the

tests/test_manual_interactive_e2e.sh

@@ -1,6 +1,9 @@
 #!/bin/bash
 set -euo pipefail
 
+# shellcheck source=_test_env.sh
+source "$(dirname "${BASH_SOURCE[0]}")/_test_env.sh"
+
 # Unit tests for the manual interactive E2E fixture generator.
 
 PASS=0

tests/test_session_end_push.sh

@@ -1,6 +1,9 @@
 #!/bin/bash
 set -euo pipefail
 
+# shellcheck source=_test_env.sh
+source "$(dirname "${BASH_SOURCE[0]}")/_test_env.sh"
+
 # Guard against the regression reported in issue #82:
 #   session-end `git pull --rebase && git push` at
 #   lib/harness.sh fails every retry when the target repo
@@ -11,11 +14,6 @@ set -euo pipefail
 #   the hook-suppression mechanism end-to-end, and pin the
 #   park-push retry loop both structurally and behaviorally.
 
-# Isolate from host gitconfig (signing keys, hooks, templates).
-# This test exercises rebase/push mechanics, not signing.
-export GIT_CONFIG_GLOBAL=/dev/null
-export GIT_CONFIG_SYSTEM=/dev/null
-
 PASS=0
 FAIL=0
 TMPDIR=$(mktemp -d)