code-coordinator 0.5.130__tar.gz → 0.5.132__tar.gz

{code_coordinator-0.5.130 → code_coordinator-0.5.132}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: code-coordinator
-Version: 0.5.130
+Version: 0.5.132
 Summary: Coordinate Claude Code workers across multiple machines and repos.
 Author-email: John Donaghy <johnfdonaghy@gmail.com>
 License-Expression: LicenseRef-FSL-1.1-MIT

{code_coordinator-0.5.130 → code_coordinator-0.5.132}/code_coordinator.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: code-coordinator
-Version: 0.5.130
+Version: 0.5.132
 Summary: Coordinate Claude Code workers across multiple machines and repos.
 Author-email: John Donaghy <johnfdonaghy@gmail.com>
 License-Expression: LicenseRef-FSL-1.1-MIT

{code_coordinator-0.5.130 → code_coordinator-0.5.132}/code_coordinator.egg-info/SOURCES.txt

@@ -539,6 +539,7 @@ tests/test_conflict_fix.py
 tests/test_conflict_fix_semantic.py
 tests/test_coord_test.py
 tests/test_coord_test_runner_baseline.py
+tests/test_coord_test_runner_populated_home.py
 tests/test_coord_test_runner_routing.py
 tests/test_coord_test_runner_toolchain.py
 tests/test_coord_web_sse_restart_live.py

{code_coordinator-0.5.130 → code_coordinator-0.5.132}/code_coordinator.egg-info/scm_file_list.json

@@ -581,6 +581,7 @@
     "tests/test_conflict_fix_semantic.py",
     "tests/test_coord_test.py",
     "tests/test_coord_test_runner_baseline.py",
+    "tests/test_coord_test_runner_populated_home.py",
     "tests/test_coord_test_runner_routing.py",
     "tests/test_coord_test_runner_toolchain.py",
     "tests/test_coord_web_sse_restart_live.py",

{code_coordinator-0.5.130 → code_coordinator-0.5.132}/code_coordinator.egg-info/scm_version.json

@@ -1,7 +1,7 @@
 {
-  "tag": "0.5.130",
+  "tag": "0.5.132",
   "distance": 0,
-  "node": "g14a2d9301195335040144bf7d82384aeddb54b2b",
+  "node": "g77edefa9b01b23c5ab4770f72c62d1469f11346e",
   "dirty": false,
   "branch": "HEAD",
   "node_date": "2026-08-17"

{code_coordinator-0.5.130 → code_coordinator-0.5.132}/coord/drive_queue.py

@@ -309,6 +309,26 @@ RESUME_PROBE_TIMEOUT_SECONDS = 5.0
 # an unsatisfiable pre-req.
 DEFAULT_MAX_ATTEMPTS = 2
 
+# #2363: the WIDER ceiling for one specific drive-death signature — an
+# acceptance-author or plain work session that exited DONE/ADVISORY claiming
+# success while its branch carried zero commits (see
+# `_is_empty_branch_death_reason`). Every recorded instance of this shape in
+# `~/.coord/queue-block-log.jsonl` self-healed 0% of the time inside
+# `DEFAULT_MAX_ATTEMPTS` (2) — `claude-coordinator#2283`, `#2348`;
+# `coord-portal#74` (×2); `space-invaders#1`, `#3` — every one needed a
+# by-hand `operator_removed` after exactly 2 attempts, and the by-hand
+# recovery is `coord acceptance author <repo> <tracking> --issue <n>` (or a
+# bare `coord drive`) — the exact command the queue already tried and could
+# just try again itself. Unlike a CI signal (`is_ci_infra_reason` et al in
+# `coord.merge_queue`), there is no free, passive re-check here — the only
+# way to learn whether a retry would succeed is to actually spend one — so
+# the fix is a WIDER attempt budget for this signature alone, not a global
+# increase: every other death reason keeps `DEFAULT_MAX_ATTEMPTS` unchanged.
+# Still finite: once THIS budget is also exhausted, the entry blocks with
+# the same diagnosis-and-recovery instructions it has today (#1526/#2273
+# discipline — no silent infinite retry).
+EMPTY_BRANCH_MAX_ATTEMPTS = 6
+
 # ── retry spacing (#2273) ────────────────────────────────────────────────────
 #
 # 2026-08-15: quadraui#508 and coord-portal#83 each spent their ENTIRE
@@ -1679,6 +1699,48 @@ def _dispatch_produced_nothing(entry: QueueEntry, facts: IssueFacts) -> bool:
     return facts.last_dispatched_at < entry.launched_at
 
 
+def _is_empty_branch_death_reason(reason: str | None) -> bool:
+    """#2363: does *reason* name the "claimed success, wrote nothing"
+    signature — an acceptance-author or plain work session that exited
+    DONE/ADVISORY while its branch carried zero commits?
+
+    Text-matched against the drive's own `drive_exited` reason, the same
+    convention `is_ci_infra_reason`/`is_ci_flaky_reason`/
+    `is_ci_unreadable_reason` in `coord.merge_queue` use to classify a CI
+    status string — not marker-based like `is_permanent_block_reason`,
+    because neither `_die` call site this recognizes embeds a dedicated
+    marker; today nothing reads their reason text for anything but display.
+
+    Two independent shapes, both required to include "no commits" so a
+    death that merely MENTIONS one of the other words in passing does not
+    false-positive:
+
+    * the acceptance-author JIT-slice death (`coord/drive.py:850-865`
+      ADVISORY, `:887-918` DONE) — "acceptance author ... exited
+      (ADVISORY|DONE) ... no commits".
+    * the plain work-row death (`coord/drive.py`'s `_decide_advisory`) —
+      "work ... exited ADVISORY ... no commits ... nothing was pushed".
+
+    Deliberately does NOT match the sibling "work ... finished with no
+    branch — nothing was pushed" reason (`coord/drive.py`'s `_decide`, the
+    'done'-status-with-no-branch arm): that shape never reaches
+    `branch_has_commits` at all, is not one of the two shapes the issue's
+    acceptance criteria name, and has no recorded evidence of its own in
+    `queue-block-log.jsonl` — folding it in unasked would be exactly the
+    un-scoped widening #2230's own docstring warns a naive rule invites.
+    """
+    if not reason:
+        return False
+    lowered = reason.lower()
+    if "no commits" not in lowered:
+        return False
+    if "acceptance author" in lowered and (
+        "exited advisory" in lowered or "exited done" in lowered
+    ):
+        return True
+    return "exited advisory" in lowered and "nothing was pushed" in lowered
+
+
 def _retry_backoff_reason(
     entry: QueueEntry,
     facts: IssueFacts,
@@ -1840,6 +1902,15 @@ def _reconcile_running(
     absent, refused or not — only ever changes the WORDING below; whether the
     entry gets another attempt is otherwise unaffected by #1845/#1844 (still
     ``retry`` until ``max_attempts``, still ``exhausted`` → ``blocked`` after).
+    The ONE exception is #2363: when ``own_reason`` matches
+    :func:`_is_empty_branch_death_reason` — an acceptance-author or work
+    session that exited DONE/ADVISORY claiming success with a zero-commit
+    branch — the ceiling widens to :data:`EMPTY_BRANCH_MAX_ATTEMPTS` instead
+    of the passed-in ``max_attempts``, because that signature's own recorded
+    history (see the constant's comment) shows a 0% self-heal rate inside
+    the default budget. Still finite, still ``exhausted`` → ``blocked`` with
+    the same diagnosis-and-recovery instructions once ITS wider ceiling is
+    also hit.
     """
     facts = board.facts(entry.key)
 
@@ -2128,18 +2199,39 @@ def _reconcile_running(
         else ""
     )
 
+    # #2363: the "claimed success, wrote nothing" signature gets a WIDER
+    # ceiling than every other death reason — see `EMPTY_BRANCH_MAX_ATTEMPTS`
+    # and `_is_empty_branch_death_reason`. `max()` against the passed-in
+    # `max_attempts` rather than an outright override so a caller-supplied
+    # `--max-attempts` larger than the #2363 default is never narrowed.
+    # Additive only: any other reason (including `dispatch_only` above) still
+    # uses the plain `max_attempts` ceiling, unchanged.
+    empty_branch = _is_empty_branch_death_reason(own_reason)
+    effective_max_attempts = (
+        max(max_attempts, EMPTY_BRANCH_MAX_ATTEMPTS) if empty_branch else max_attempts
+    )
+    empty_branch_note = (
+        f" — empty-branch DONE/ADVISORY exit (#2363): widened retry budget, "
+        f"{effective_max_attempts} attempts before blocking instead of the "
+        f"usual {max_attempts}"
+        if empty_branch
+        else ""
+    )
+
     attempts = entry.attempts + 1
-    if attempts < max_attempts:
+    if attempts < effective_max_attempts:
         if own_reason:
             reason = (
-                f"{own_reason} (attempt {attempts}/{max_attempts}) — "
-                f"requeued at position {entry.position}{dispatch_note}"
+                f"{own_reason} (attempt {attempts}/{effective_max_attempts}) — "
+                f"requeued at position {entry.position}"
+                f"{dispatch_note}{empty_branch_note}"
             )
         else:
             reason = (
                 f"drive session died without landing the work"
-                f"{launched} (attempt {attempts}/{max_attempts}) — requeued at "
-                f"position {entry.position}{dispatch_note}"
+                f"{launched} (attempt {attempts}/{effective_max_attempts}) — "
+                f"requeued at position {entry.position}"
+                f"{dispatch_note}{empty_branch_note}"
             )
         return (
             Reconcile(
@@ -2168,14 +2260,14 @@ def _reconcile_running(
 
     if own_reason:
         reason = (
-            f"{own_reason} ({attempts}/{max_attempts} attempts) — giving up"
-            f"{dispatch_note}"
+            f"{own_reason} ({attempts}/{effective_max_attempts} attempts) — "
+            f"giving up{dispatch_note}{empty_branch_note}"
         )
     else:
         reason = (
             f"drive session died without landing the work"
-            f"{launched} {attempts}/{max_attempts} times — giving up"
-            f"{dispatch_note}"
+            f"{launched} {attempts}/{effective_max_attempts} times — giving up"
+            f"{dispatch_note}{empty_branch_note}"
         )
     return (
         Reconcile(entry.key, "exhausted", reason, occupies=False),

{code_coordinator-0.5.130 → code_coordinator-0.5.132}/scripts/coord-test-runner.sh

@@ -91,6 +91,24 @@
 #     see the long note above `ensure_baseline_worktree` for why the fallback
 #     arm is excluded rather than "not done yet".
 #
+#  7. THE POPULATED-$HOME ARM (#2269).  A branch can be green here and red in
+#     CI's `populated-home` job, and until #2269 whether that was caught
+#     depended on WHICH MACHINE the Test stage landed on.  That job's
+#     environment is SYNTHESIZED by scripts/run_tests_in_populated_home.sh (a
+#     thin-client ~/.coord with no coordinator.yml, no `sqlite3` on PATH, a
+#     $TMPDIR under an ancestor pytest config); this runner otherwise inherits
+#     whatever environment the assigned machine happens to have.  dellserver —
+#     the daemon host — is structurally incapable of reproducing knob 1, so
+#     #2174 passed the Test gate, spent a full adversarial review, and was
+#     caught only when `coord fix` noticed CI disagreed.  So once the ordinary
+#     python arm is green, the test files THE DIFF TOUCHED are re-run through
+#     that same harness.  Diff-scoped because running the suite twice is not
+#     affordable (#2169).  This arm can only ever FAIL, never downgrade: the
+#     harness's own `exit 2` ("a knob failed to take effect") is a
+#     FAIL-with-warning, never a skip.  Python arm only, and it names itself
+#     (`python-populated-home`) in every verdict line it emits, so a red result
+#     is attributable rather than a mystery.
+#
 # Exit codes: 0 pass (or skip — nothing to test), 1 genuine failure or refusal
 # (cannot determine what to test), 2 usage, 3 INFRASTRUCTURE — the suite could
 # not run at all (a required toolchain is missing); no verdict may be inferred
@@ -215,9 +233,77 @@ else
     log "routing: repo=$REPO_NAME — no path rule and no --fallback-command"
 fi
 
+# ── populated-$HOME arm scoping (#2269) ─────────────────────────────────────
+#
+# WHICH FILES. Only the python test files THIS DIFF touched, and only ones that
+# still exist in the worktree (a diff that DELETES a test file must not hand
+# pytest a path that is not there). `tests/acceptance/**` is excluded for the
+# same reason the ordinary arm passes `--ignore=tests/acceptance`: that sealed
+# suite is gated separately through the #2164 `--ci` wrapper, and a
+# red-by-design slice must not fail every concurrent branch.
+#
+# WHY DIFF-SCOPED AND NOT THE WHOLE SUITE. The full suite already exceeds
+# Claude Code's 600s Bash ceiling (#2169); running it a second time under the
+# harness is not on the table. Diff-scoping is what makes this arm affordable —
+# it costs roughly one test FILE's runtime, paid only by branches that touch
+# python tests at all. The always-on cover for the rest of the suite is
+# tests/test_ambient_home_isolation.py (which runs everywhere, on every branch);
+# this arm exists for the case that file cannot cover, a NEWLY WRITTEN test of
+# this class (#2174 was exactly that).
+#
+# Computed HERE, beside the routing decision, rather than inside `run_python`,
+# because `--print-routing` must be able to name the arm before a worker pushes.
+POPULATED_TESTS=""
+POPULATED_COUNT=0
+POPULATED_ARM=0
+POPULATED_SKIP_REASON=""
+
+is_python_test_file() {
+    case "$1" in
+        tests/acceptance/*) return 1 ;;
+        tests/*)            ;;
+        *)                  return 1 ;;
+    esac
+    case "${1##*/}" in
+        test_*.py) return 0 ;;
+        *)         return 1 ;;
+    esac
+}
+
+if [[ "$ROUTE_MODE" == "coordinator" && "$RUN_PY" -eq 1 ]]; then
+    if [[ "$DIFF_FAILED" -eq 1 ]]; then
+        POPULATED_SKIP_REASON="the diff against $BASE_REF could not be computed, so the arm has no scope to run over"
+    else
+        while IFS= read -r f; do
+            [[ -z "$f" ]] && continue
+            is_python_test_file "$f" || continue
+            # Deleted-by-the-diff files are still listed by `git diff --name-only`.
+            [[ -f "$WT/$f" ]] || continue
+            POPULATED_TESTS="${POPULATED_TESTS:+$POPULATED_TESTS }$f"
+            POPULATED_COUNT=$((POPULATED_COUNT + 1))
+        done <<<"$CHANGED"
+        [[ "$POPULATED_COUNT" -gt 0 ]] && POPULATED_ARM=1
+        [[ "$POPULATED_ARM" -eq 0 ]] && \
+            POPULATED_SKIP_REASON="the diff touches no python test files under tests/ (outside the sealed tests/acceptance/)"
+    fi
+elif [[ "$ROUTE_MODE" == "coordinator" ]]; then
+    POPULATED_SKIP_REASON="the python arm is not routed for this diff"
+else
+    POPULATED_SKIP_REASON="the populated-\$HOME arm is this repo's python arm only"
+fi
+
+if [[ "$ROUTE_MODE" == "coordinator" ]]; then
+    if [[ "$POPULATED_ARM" -eq 1 ]]; then
+        log "routing: populated-\$HOME arm (#2269) will re-run $POPULATED_COUNT diff-scoped test file(s): $POPULATED_TESTS"
+    else
+        log "routing: populated-\$HOME arm (#2269) SKIPPED — $POPULATED_SKIP_REASON"
+    fi
+fi
+
 if [[ "$PRINT_ROUTING" -eq 1 ]]; then
     case "$ROUTE_MODE" in
-        coordinator) printf 'ROUTING mode=coordinator pytest=%s cargo=%s\n' "$RUN_PY" "$RUN_RS" ;;
+        coordinator) printf 'ROUTING mode=coordinator pytest=%s cargo=%s populated-home=%s\n' \
+                        "$RUN_PY" "$RUN_RS" "$POPULATED_ARM" ;;
         fallback)    printf 'ROUTING mode=fallback run=%s\n' "$RUN_FALLBACK" ;;
         unknown)     printf 'ROUTING mode=unknown\n' ;;
     esac
@@ -245,6 +331,8 @@ EXIT_INFRA=3
 # baseline is red, so this is not a verdict on the branch either (#2170).
 BASELINE_RED_SUITES=()
 EXIT_BASELINE_RED=4
+# See the note at the dispatch site below.
+PY_FAIL_SUITE="python"
 
 # ── baseline comparison (#2170) ──────────────────────────────────────────────
 #
@@ -570,6 +658,98 @@ run_python_acceptance_ci() {
     return 1
 }
 
+# ── the populated-$HOME arm (#2269) ─────────────────────────────────────────
+#
+# Re-run the diff's own python test files in the environment of a REAL FLEET
+# MACHINE instead of this machine's, using the SAME harness CI's
+# `populated-home` job uses (scripts/run_tests_in_populated_home.sh, from the
+# BRANCH's tree — so a branch that changes the harness is tested against its own
+# version, exactly like CI). See item 7 in the header for why this is not
+# optional: without it, catching this failure class is a property of which
+# machine the Test stage was routed to.
+#
+# THIS ARM ONLY EVER FAILS. It is the mirror image of #2170's baseline check,
+# which only ever downgrades and is therefore deliberately hard to trigger. This
+# one is the STRICT direction, so every ambiguity resolves the other way:
+#
+#   * The harness's own `exit 2` — "a knob failed to take effect", i.e. the
+#     environment it claims to synthesize was not synthesized — is a FAIL with a
+#     warning, NOT a skip. A guard that cannot guard reporting green is the exact
+#     failure mode #2170 exists to prevent, and a Test gate that silently
+#     tolerates it is back to routing luck.
+#   * Any other non-zero exit is a FAIL naming this arm.
+#
+# There is no flake filter here on purpose: these same files went green in the
+# ordinary arm moments ago, so a failure here is by construction environmental,
+# and re-running would only launder it.
+#
+# The two SKIPs it can report (no python test files in the diff; the harness
+# absent from this branch's tree) are announced in the verdict stream and by
+# `--print-routing`, never silent.
+run_python_populated_home() {
+    local venv="$1"
+    local rel="scripts/run_tests_in_populated_home.sh"
+    local script="$WT/$rel"
+
+    if [[ "$POPULATED_ARM" -eq 0 ]]; then
+        say "SKIP(python-populated-home): not run — $POPULATED_SKIP_REASON (#2269)"
+        return 0
+    fi
+    if [[ ! -f "$script" ]]; then
+        # A branch (or a checkout) predating the harness. Skipping is the only
+        # option that does not redden every such branch, but it is stated out
+        # loud rather than inferred from silence.
+        warn "populated-home: $rel is not present in this worktree — the arm cannot run"
+        say "SKIP(python-populated-home): not run — the harness $rel is absent from this branch's tree (#2269)"
+        return 0
+    fi
+
+    local out="$WT/.pytest-populated-home.out"
+    local started=$SECONDS
+    local rc=0
+    log "running: $rel over $POPULATED_COUNT diff-scoped test file(s): $POPULATED_TESTS"
+    # shellcheck disable=SC2086  # the file list is intentionally word-split
+    (cd "$WT" && bash "$script" "$venv/bin/python" -m pytest -q --tb=short $POPULATED_TESTS) \
+        >"$out" 2>&1 || rc=$?
+    local elapsed=$((SECONDS - started))
+
+    if [[ "$rc" -eq 0 ]]; then
+        say "PASS(python-populated-home): $POPULATED_COUNT diff-scoped test file(s) green in a synthesized fleet \$HOME (+${elapsed}s)"
+        return 0
+    fi
+
+    # The harness's own guard failures. `run_tests_in_populated_home.sh`
+    # reserves exit 2 for exactly this case, so check `rc` directly first —
+    # it can't drift out of sync with itself. The message grep is a second,
+    # belt-and-braces signal for a harness that fails this way but exits with
+    # some other code (or an older/forked copy of the script); if the exact
+    # wording ever changes, `rc -eq 2` alone still catches the exit-2 case.
+    # Distinguished from a test failure because the two need different fixes
+    # — but BOTH are a FAIL, so mis-classifying one as the other can only
+    # ever mis-word a red verdict, never hide one.
+    if [[ "$rc" -eq 2 ]] || grep -qE "the guard itself is broken|the guard is too broad|is not thin-client shaped" "$out"; then
+        say "FAIL(python-populated-home): the harness $rel exited $rc because A KNOB FAILED TO TAKE EFFECT — the environment it claims to synthesize (thin-client ~/.coord, no sqlite3, \$TMPDIR under an ancestor pytest config) was not the environment the tests ran in, so a green here would have meant nothing. This is reported as a FAILURE, not a skip, on purpose (#2269)."
+        tail -n 20 "$out" | sed 's/^/      /'
+        return 1
+    fi
+
+    say "FAIL(python-populated-home): $POPULATED_COUNT diff-scoped test file(s) PASS in this machine's ambient environment but FAIL in a synthesized fleet \$HOME (thin-client ~/.coord with no coordinator.yml, no sqlite3 on PATH, \$TMPDIR under an ancestor pytest config) — this is CI's populated-home job's failure class, caught here instead of at the merge gate (#2269). Reproduce with: $rel python -m pytest $POPULATED_TESTS"
+    printf '%s\n' "$POPULATED_TESTS" | tr ' ' '\n' | sed 's/^/      /'
+    tail -n 40 "$out" | sed 's/^/      /'
+    return 1
+}
+
+# The arms that run AFTER the ordinary pytest suite has gone green (or been
+# judged a tolerated flake), in cost order. Short-circuiting on the first
+# failure is deliberate: the verdict is already attributable to a named arm, and
+# there is nothing to learn from paying for the next one.
+run_python_post_arms() {
+    local venv="$1"
+    run_python_acceptance_ci "$venv" || return 1
+    run_python_populated_home "$venv" || { PY_FAIL_SUITE="python-populated-home"; return 1; }
+    return 0
+}
+
 run_python() {
     local venv="$WT/.venv"
     if [[ ! -x "$venv/bin/python" ]]; then
@@ -618,7 +798,7 @@ run_python() {
     # `set -u` on older bash.
     if (cd "$WT" && "$venv/bin/python" -m pytest -q --tb=short --ignore=tests/acceptance ${par[@]+"${par[@]}"}) >"$out" 2>&1; then
         log "ordinary suite: $(grep -oE '[0-9]+ passed[^)]*' "$out" | tail -1)"
-        run_python_acceptance_ci "$venv"
+        run_python_post_arms "$venv"
         return $?
     fi
 
@@ -645,7 +825,7 @@ run_python() {
         say "FLAKE(python): $count test(s) failed in the full run but PASS in isolation"
         printf '%s\n' "$failed" | sed 's/^/      /'
         FLAKES+=("python:$count")
-        run_python_acceptance_ci "$venv"
+        run_python_post_arms "$venv"
         return $?
     fi
 
@@ -854,8 +1034,13 @@ mark_failed() {
     FAILED_SUITES+=("$suite")
 }
 
+# Which name a failing python arm books itself under. "python" for every arm
+# that existed before #2269; the populated-$HOME arm overwrites it with its own
+# name so `RESULT: FAIL (python-populated-home)` says which environment went
+# red, not merely that pytest did (#2269's "a red result is attributable"
+# requirement). It is only ever read after `run_python` returns non-zero.
 if [[ "$RUN_PY" -eq 1 ]]; then
-    run_python || mark_failed python
+    run_python || mark_failed "$PY_FAIL_SUITE"
 fi
 if [[ "$RUN_RS" -eq 1 ]]; then
     run_rust || mark_failed rust