Gladex Agent Logs

10345:## [0.4.335] - 2026-10-13 — run 939 closes run 938's queued follow-up: a `tools/shift-deploy` step-3 **post-copy** child that could not be read, or read without verifying, is exit 3 (`cannot verify`), never exit 6 claiming a disagreement nobody measured (1.2.0 → 1.3.0, suite 118 → 132)
> build · mimo-v2.6-flash-free
$ cd /data/repo; sed -n '10345,10400p' CHANGELOG.md; echo "=== fence count:"; grep -c '^```' CHANGELOG.md
## [0.4.335] - 2026-10-13 — run 939 closes run 938's queued follow-up: a `tools/shift-deploy` step-3 **post-copy** child that could not be read, or read without verifying, is exit 3 (`cannot verify`), never exit 6 claiming a disagreement nobody measured (1.2.0 → 1.3.0, suite 118 → 132)

### Added
- **Two exit-3 branches on step 3's post-copy `source-sync-check` re-read.** Step 3 reads
  the checker twice: once before the `rsync` (to size the drift) and once after it (to
  prove the copy landed). The pre-copy read has always split *"couldn't run it"* /
  *"ran but couldn't verify"* from *"ran and it disagreed"* — `if rc is None or rc not in
  (0, 1): refuse(EXIT_CANNOT, …)`; the post-copy read did not: `if rc is None or rc != 0:`
  collapsed all three into exit **6**, `dev still differs from the repo after the copy`.
  Now it mirrors its twin: `rc is None` → exit **3**, `source-sync-check could not be run:
  …`; `rc not in (0, 1)` → exit **3**, `source-sync-check could not verify the copy (rc=…)`
  (the checker's own exit `3` is *"a compared base directory does not exist"* — it ran,
  compared nothing, and `6` would again assert a difference no comparison produced);
  `rc == 1` → exit **6**, `dev still differs from the repo after the copy (rc=1)`
  byte-for-byte unchanged. Neither exit-`3` branch carries `reason`, the key being defined
  as *"which of the two exit-9 checks fired"*. EPILOG line `3` now enumerates the steps —
  `preflight, gates, dev-sync (before AND after the copy), promote and verify alike` — and
  line `6` reads *"the post-copy re-read ran and answered 1, i.e. it compared the trees and
  they differ (a child that could not run, or ran without verifying, is 3)"*; the module
  docstring's step-3 row states both outcomes. The epilog's line 3 is now true at every
  step, which is what run 938 recorded as *"the last place"* it was false.
- **What the queue named vs what was measured.** The queued item said *"split `rc is
  None` → 3 from `rc != 0`"* — that is the first branch. The second was found by probing
  the same line before writing: a stand-in exiting **3** on the post-copy read produced
  `dev-sync | refused | dev still differs from the repo after the copy (rc=3)` at exit
  **6** while `cmp` in the same probe proved dev byte-identical to the repo — the identical
  false claim one condition over, in the expression being edited. Folding it in was
  necessary for this run's own EPILOG line `6` to be true, so it is disclosed here as an
  extension of the queued scope rather than presented as it.
- **Suite assertions F8–F8n, 118 → 132.** *The unrunnable half (F8–F8h):* the stand-in
  reports drift once (so the copy branch runs) and `rm -f`s its own path on the way out, so
  the *second* exec of the same path fails outright. Asserted: exit `3` (not 6); the
  refusing step is `dev-sync`; the detail says `source-sync-check could not be run`; it
  never says `still differs from the repo`; no `"reason"` anywhere in the object; the
  report's `exit_code` equals the process rc; **dev is byte-identical to the repo at the
  moment of refusal** — the fact the old detail denied, and the assertion that makes this a
  test of the defect rather than of a string; and the stand-in really did remove its own
  path first. *The ran-but-could-not-verify half (F8l–F8n):* same shape, stand-in answers
  drift then its own exit `3` — asserting `3`, `source-sync-check could not verify the
  copy`, and dev == repo again. *The control (F8i–F8k):* the identical run with the break
  path pointed at a path nothing uses gets a re-read that runs, compares and still reports
  drift, so it must stay exit **6**, keep `dev still differs from the repo after the copy`,
  and must not wear the new wording — which is what stops the wording assertions from being
  satisfiable by the harness alone and what proves exit 6 did not vanish with the fix.
- **`tools/REGISTRY.md` section `shift-deploy`**: the step-3 table row's refusal column
  (the row now records that the checker is read twice), the `3` exit-code row (step 3's
  post-copy read joins the list, *"the last place the epilog's line 3 was false"*), the
  `6` row (*a comparison since `[0.4.335]`, never an inference*), the `9` row's *Known
  residual* note rewritten as measured-and-closed, the `Tests` figure **118 → 132** (both
  places), and a dated **Changed** line recording version **1.2.0 → 1.3.0**.

### Verification
- `bash tests/test_shift_deploy.sh` → **132 passed / 0 failed, rc 0**, run after the code
  change (118/0 with **no pre-existing assertion reddened** — the fix takes nothing away
  from what the old behaviour satisfied) and again after the assertions and the docs
  landed, live layer green on both: `B7 live (rc 0): gate [verdict|commit-lint|ship-tree]
=== fence count:
16
> build · mimo-v2.6-flash-free
$ cd /data/repo; sed -n '10400,10430p' CHANGELOG.md
landed, live layer green on both: `B7 live (rc 0): gate [verdict|commit-lint|ship-tree]
  is ok`.
- **Out-of-suite mutant, one, full `tools/`+`tests/` copy under
  `/tmp/opencode/mutroot_sdep_r939m1`** with **both** new branches collapsed back into the
  single exit-6 arm: **117 passed / 10 failed → 6 signal reds: F8** (`want rc=3 got=6`),
  **F8c** (missing the wording), **F8d** (the false claim present), **F8f** (`exit_code` 6
  not 3), **F8l** (`want rc=3 got=6`), **F8m** (missing the cannot-verify wording) — each an
  assertion of this change and nothing else — plus the already-documented **B6 quartet**
  (the mutant root has no `.git`, so its live layer refuses at exit 3; a harness artifact,
  disclosed, not counted). Live `tools/shift-deploy` md5
  **`b8c96f2cc849c00c75fd100837b640ae`** read identical before the mutant run and after it —
  the mutation lived entirely under `/tmp/opencode`.
- **The control is the load-bearing half.** F8i exits 6 with the old wording in the same
  suite run in which F8 and F8l exit 3 with the new ones, so neither wording nor code is
  producible by the harness alone, and the branch this change did not touch is proven
  reachable rather than merely assumed to be.
- `tests/test_registry_coverage.sh` → **rc 0, `violations=0`**, `fig_disc=107 fig_live=1
  fig_claim=107 fig_stale=0 fig_head=107 fig_src=consistent`, `tools=29 registered=45
  help_ok=29 badarg_ok=29 ec_ok=29 fmt_claim_mismatch=0 allflag_mismatch=0` — the `-
  Live:` figure **107** re-confirmed on this tree after the REGISTRY edits, and every other
  figure the section carries machine-checked.

### Notes
- **No `app/` file** → nothing to promote, no reviewer-gated deploy, no non-dry
  `promote-dev-to-prod` (every mutating path in the suite runs inside its own throwaway
  sandbox under `/tmp/opencode`, and the only non-suite write was the mutant copy);
  dev and prod untouched at **0.4.28** (`version-check` → `dev=0.4.28 prod=0.4.28
  (match)`). STEP 0 owed nothing (**0 unread dev / 0 prod**, `INBOX.md` **80 / 80 / 0
  open**, `inbox-status` rc 0), so no message row was written and nothing was marked read.
  Queue line **(a⁵)** NOT taken: this run started 16:17 CEST, before sofia's next fire
  (`identity-run@sofia.timer` `NEXT … Tue 2026-10-13 18:55:10 CEST`), and the post-fire
> build · mimo-v2.6-flash-free
Registry coverage came back **rc=0, violations=0**. Let me verify the C28g count reasoning, since live `syn_n=29` while the assertion expects 28:
$ cd /data/repo; grep -E 'C28g|C28h|C28f' /tmp/opencode/regcov.out; echo ===; grep -n 'fresh()' tests/test_registry_coverage.sh | head -3
===
1816:fresh() {
> build · mimo-v2.6-flash-free
$ cd /data/repo; grep -cE '^PASS' /tmp/opencode/regcov.out; grep -cE '^FAIL' /tmp/opencode/regcov.out; tail -3 /tmp/opencode/regcov.out; echo ===; sed -n '1816,1835p' tests/test_registry_coverage.sh
0
0
DIAG figure_basis side=worktree claim=108 head=107 disc=108
SUMMARY tools=30 registered=46 test_sections=16 help_run=30 help_ok=30 badarg_run=30 badarg_ok=30 ec_run=30 ec_ok=30 fmt_n=26 fmt_value_ok=26 fmt_missing_ok=26 timeout_n=7 timeout_ok=7 port_n=4 port_value_ok=4 port_missing_ok=4 miss_n=46 miss_missing_ok=46 conv_pairs=75 conv_extra=0 fig_disc=108 fig_live=1 fig_claim=108 fig_stale=0 fig_undated=0 fig_head=107 fig_src=worktree syn_n=29 syn_ok=29 synp_n=75 synp_ok=75 cl_scope=76 cl_fmt=26 cl_port=4 cl_enum=3 cl_enum_ok=3 cl_num=15 cl_num_ok=15 cl_declared=3 cl_declared_ok=3 cl_declared_seen=1,3,3 cl_other=25 live_re_sites=3 gloss_sites=3 fmt_claim_pairs=23 fmt_claim_mismatch=0 allflag_pairs=47 allflag_mismatch=0 allflag_numeq=1 violations=0
===
fresh() {
  local d="$SBROOT/$1" f t
  rm -rf "$d"
  mkdir -p "$d/tools" "$d/tests"
  for f in "$cov_tools"/*; do
    [ -f "$f" ] && cp "$f" "$d/tools/"
  done
  for f in "$cov_tests"/*; do
    [ -f "$f" ] && cp "$f" "$d/tests/"
  done
  # Registered-but-absent suites still get a stub, so the reverse direction of
  # the coverage check keeps reading existence rather than content. Nothing
  # truncates a copy that is already there.
  for t in $(sed -n 's/^## \(test_[A-Za-z0-9._-]*\)$/\1/p' "$d/tools/REGISTRY.md"); do
    [ -f "$d/tests/$t" ] || : > "$d/tests/$t"
  done
  printf '%s' "$d"
}

# run_child <sandbox> -> OUT, RC.  The ONLY way the checker is invoked here.
> build · mimo-v2.6-flash-free
$ cd /data/repo; wc -l /tmp/opencode/regcov.out; head -5 /tmp/opencode/regcov.out; echo ===; grep -n 'REGISTRY_COVERAGE_CHILD' tests/test_registry_coverage.sh | head
2 /tmp/opencode/regcov.out
DIAG figure_basis side=worktree claim=108 head=107 disc=108
SUMMARY tools=30 registered=46 test_sections=16 help_run=30 help_ok=30 badarg_run=30 badarg_ok=30 ec_run=30 ec_ok=30 fmt_n=26 fmt_value_ok=26 fmt_missing_ok=26 timeout_n=7 timeout_ok=7 port_n=4 port_value_ok=4 port_missing_ok=4 miss_n=46 miss_missing_ok=46 conv_pairs=75 conv_extra=0 fig_disc=108 fig_live=1 fig_claim=108 fig_stale=0 fig_undated=0 fig_head=107 fig_src=worktree syn_n=29 syn_ok=29 synp_n=75 synp_ok=75 cl_scope=76 cl_fmt=26 cl_port=4 cl_enum=3 cl_enum_ok=3 cl_num=15 cl_num_ok=15 cl_declared=3 cl_declared_ok=3 cl_declared_seen=1,3,3 cl_other=25 live_re_sites=3 gloss_sites=3 fmt_claim_pairs=23 fmt_claim_mismatch=0 allflag_pairs=47 allflag_mismatch=0 allflag_numeq=1 violations=0
===
15:# (REGISTRY_COVERAGE_CHILD=1 re-runs THIS script to that branch), so a control
1552:  ( unset REGISTRY_COVERAGE_CHILD
1611:  ( unset REGISTRY_COVERAGE_CHILD
1638:if [ "${REGISTRY_COVERAGE_CHILD:-0}" = "1" ]; then
1790:             REGISTRY_COVERAGE_TESTS_DIR, REGISTRY_COVERAGE_CHILD (internal)
1838:  OUT=$(REGISTRY_COVERAGE_CHILD=1 \
1867:OUT=$(REGISTRY_COVERAGE_CHILD=1 \
3393:  sh -c '( unset REGISTRY_COVERAGE_CHILD; timeout 20 bash "$1" --help ) 2>/dev/null \
> build · mimo-v2.6-flash-free
The child-mode summary is the live check only (no C28g). I need the full suite to validate my C28g edit — launching it in the background (~11 min) while I write the changelog:
$ cd /data/repo; bash tests/test_registry_coverage.sh > /tmp/opencode/regcov-full.out 2>&1; echo "rc=$?"; grep -cE '^FAIL' /tmp/opencode/regcov-full.out; grep -E '^FAIL' /tmp/opencode/regcov-full.out | head -20; tail -3 /tmp/opencode/regcov-full.out
Command moved to the background (shell ID: sh_13614b8bf001fF1nc07XKujpGt).
Output is streaming to: /root/.local/share/opencode/shell/fa3e85b2145a60c127b5c9261e76a7ba6d520299/sh_13614b8bf001fF1nc07XKujpGt.out
$ cd /data/repo; sed -n '1904,1915p' tests/test_registry_coverage.sh
assert_ge "B1 non-vacuity: at least 10 tools enumerated" 10 "$(sget_live tools)"
assert_eq "B2 tools counted == tools independently enumerated" "$live_tools" "$(sget_live tools)"
assert_eq "B3 sections counted == headings independently counted" \
  "$live_sections" "$(sget_live registered)"
assert_eq "B4 registered minus suite sections == tool sections" \
  "$live_tool_sections" "$(( $(sget_live registered) - $(sget_live test_sections) ))"
assert_eq "B5 every tool section probed for --help" "$live_tool_sections" "$(sget_live help_run)"
assert_eq "B6 every tool answered --help with exit 0 and a body" \
  "$(sget_live help_run)" "$(sget_live help_ok)"
assert_ge "B7 non-vacuity: at least 4 registered test-suite sections" 4 "$(sget_live test_sections)"

# ==========================================================================
> build · mimo-v2.6-flash-free
Full `test_registry_coverage.sh` finished rc 0. Check for FAILs:
$ cd /data/repo
grep -c '^FAIL' /tmp/opencode/regcov-full.out
grep '^FAIL' /tmp/opencode/regcov-full.out | head -3
tail -8 /tmp/opencode/regcov-full.out | head -8
grep -n 'C28' /tmp/opencode/regcov-full.out | head -3

0
PASS: C5b dns-verify without its \*\*Usage\*\* -> exit 1
PASS: C5c VIOLATION incomplete_section dns-verify:usage
PASS: C5d exactly one violation - the plant, and nothing else
PASS: C6a plant: healthcheck really lost its \*\*Purpose\*\*
PASS: C6b healthcheck without its \*\*Purpose\*\* -> exit 1
PASS: C6c VIOLATION incomplete_section healthcheck:purpose
PASS: C6d exactly one violation - the plant, and nothing else
PASS: C7a plant: tls-check really lost its xit codes
> build · mimo-v2.6-flash-free
$ cd /data/repo
grep -n 'C28g' /tmp/opencode/regcov-full.out | head -2
tail -12 /tmp/opencode/regcov-full.out | tail -8

PASS: C5b dns-verify without its \*\*Usage\*\* -> exit 1
PASS: C5c VIOLATION incomplete_section dns-verify:usage
PASS: C5d exactly one violation - the plant, and nothing else
PASS: C6a plant: healthcheck really lost its \*\*Purpose\*\*
PASS: C6b healthcheck without its \*\*Purpose\*\* -> exit 1
PASS: C6c VIOLATION incomplete_section healthcheck:purpose
PASS: C6d exactly one violation - the plant, and nothing else
PASS: C7a plant: tls-check really lost its xit codes
> build · mimo-v2.6-flash-free
C28g isn't matching — check whether the run actually got to the tail (the file may be truncated):
$ cd /data/repo
wc -l /tmp/opencode/regcov-full.out
grep -c 'PASS:' /tmp/opencode/regcov-full.out
grep -n 'C28\|summary\|TOTAL\|passed' /tmp/opencode/regcov-full.out | tail -5

45 /tmp/opencode/regcov-full.out
45
> build · mimo-v2.6-flash-free
The background capture only holds 45 lines — re-run the full suite in foreground, writing to a file:

Select another run log from the list above. Only files matching run-YYYYMMDD-HHMMSS-N.log are readable.

[Tue Oct 13 18:23:52 2026] 127.0.0.1:56236 Accepted
[Tue Oct 13 18:23:52 2026] 127.0.0.1:56236 Closing
[Tue Oct 13 18:23:52 2026] 127.0.0.1:56240 Accepted
[Tue Oct 13 18:23:52 2026] 127.0.0.1:56240 Closing
[Tue Oct 13 18:23:52 2026] 127.0.0.1:56254 Accepted
[Tue Oct 13 18:23:52 2026] 127.0.0.1:56254 Closing
[Tue Oct 13 18:23:52 2026] 127.0.0.1:56268 Accepted
[Tue Oct 13 18:23:52 2026] 127.0.0.1:56268 Closing
[Tue Oct 13 18:23:52 2026] 127.0.0.1:56270 Accepted
[Tue Oct 13 18:23:52 2026] 127.0.0.1:56270 Closing
[Tue Oct 13 18:23:52 2026] 127.0.0.1:56286 Accepted
[Tue Oct 13 18:23:52 2026] 127.0.0.1:56286 Closing
[Tue Oct 13 18:23:52 2026] 127.0.0.1:56294 Accepted
[Tue Oct 13 18:23:53 2026] 127.0.0.1:56294 Closing
[Tue Oct 13 18:23:53 2026] 127.0.0.1:56298 Accepted
[Tue Oct 13 18:23:53 2026] 127.0.0.1:56298 Closing
[Tue Oct 13 18:23:53 2026] 127.0.0.1:56312 Accepted
[Tue Oct 13 18:23:53 2026] 127.0.0.1:56312 Closing
[Tue Oct 13 18:23:53 2026] 127.0.0.1:56318 Accepted
[Tue Oct 13 18:23:53 2026] 127.0.0.1:56318 Closing
[Tue Oct 13 18:23:53 2026] 127.0.0.1:56332 Accepted
[Tue Oct 13 18:23:53 2026] 127.0.0.1:56332 Closing
[Tue Oct 13 18:23:58 2026] 127.0.0.1:37026 Accepted
[Tue Oct 13 18:23:58 2026] 127.0.0.1:37026 Closing
[Tue Oct 13 18:24:00 2026] 127.0.0.1:37032 Accepted
[Tue Oct 13 18:24:00 2026] 127.0.0.1:37032 Closing
[Tue Oct 13 18:24:00 2026] 127.0.0.1:37048 Accepted
[Tue Oct 13 18:24:00 2026] 127.0.0.1:37048 Closing
[Tue Oct 13 18:24:02 2026] 127.0.0.1:37062 Accepted
[Tue Oct 13 18:24:03 2026] 127.0.0.1:37062 Closing
[Tue Oct 13 18:24:03 2026] 127.0.0.1:37076 Accepted
[Tue Oct 13 18:24:03 2026] 127.0.0.1:37076 Closing
[Tue Oct 13 18:24:06 2026] 127.0.0.1:46738 Accepted
[Tue Oct 13 18:24:06 2026] 127.0.0.1:46738 Closing
[Tue Oct 13 18:24:06 2026] 127.0.0.1:46750 Accepted
[Tue Oct 13 18:24:06 2026] 127.0.0.1:46750 Closing
[Tue Oct 13 18:24:06 2026] 127.0.0.1:46754 Accepted
[Tue Oct 13 18:24:06 2026] 127.0.0.1:46754 Closing
[Tue Oct 13 18:24:06 2026] 127.0.0.1:46770 Accepted
[Tue Oct 13 18:24:06 2026] 127.0.0.1:46770 Closing
[Tue Oct 13 18:24:07 2026] 127.0.0.1:46778 Accepted
[Tue Oct 13 18:24:07 2026] 127.0.0.1:46778 Closing
[Tue Oct 13 18:24:17 2026] 127.0.0.1:55054 Accepted
[Tue Oct 13 18:24:17 2026] 127.0.0.1:55054 Closing
[Tue Oct 13 18:25:57 2026] 127.0.0.1:48924 Accepted
[Tue Oct 13 18:25:57 2026] 127.0.0.1:48924 Closing
[Tue Oct 13 18:27:03 2026] 127.0.0.1:52634 Accepted
[Tue Oct 13 18:27:03 2026] 127.0.0.1:52634 Closing
[Tue Oct 13 18:27:03 2026] 127.0.0.1:52640 Accepted
[Tue Oct 13 18:27:03 2026] 127.0.0.1:52640 Closing
[Tue Oct 13 18:27:04 2026] 127.0.0.1:52652 Accepted
[Tue Oct 13 18:27:04 2026] 127.0.0.1:52652 Closing
[Tue Oct 13 18:29:10 2026] 127.0.0.1:43784 Accepted
[Tue Oct 13 18:29:10 2026] 127.0.0.1:43784 Closed without sending a request; it was probably just an unused speculative preconnection
[Tue Oct 13 18:29:10 2026] 127.0.0.1:43784 Closing
[Tue Oct 13 18:29:12 2026] 127.0.0.1:43798 Accepted
[Tue Oct 13 18:29:12 2026] 127.0.0.1:43798 Closed without sending a request; it was probably just an unused speculative preconnection
[Tue Oct 13 18:29:12 2026] 127.0.0.1:43798 Closing
[Tue Oct 13 18:29:23 2026] 127.0.0.1:58048 Accepted

Other app logs: downloads.log