Gladex Agent Logs

## [0.4.257] - 2026-10-08 — queue item (ax): **the access-time cut gets a home — `tools/cache-show` (new tool) measures the Go build cache's cold half, the growth that undoes the trim and the day `/` reaches the threshold, and prints the prune command instead of leaving the next run to re-derive it** — one `find` pass over 119,106 files / 26,891,002,127 B in **0.54 s** reports **10,461,521,958 B cold** (not read in 48 h, 46,335 files), **7,065,323,144 B/day** written and **2.7 days** to 90 % of `/` at today's growth; `tests/test_cache_show.sh` (new suite) → **105 / 0** with all three mutations diverging; `tests/test_registry_coverage.sh` → **464 / 0** after three deliberate scope refreshes (Live suite figure **92 → 93**, C28g **23 → 24**, G38 **22 → 24**)

### Decisions
- **Candidate (iii) of the (ax) queue, not (i) or (ii).** The queue named three: (i) put the 48 h cut on a schedule — a system timer is the operator's to execute, never mine; (ii) move `GLADEX_GO_GOCACHE` off `/` onto `/data` — a three-file change (`tools/system-status`, `tools/repo-lint`, `tests/test_system_status_go_tests.sh`) whose first build from a cold cache on a new filesystem risks trading a measured **28 s** `go test` for an unmeasurable one under `GO_TIMEOUT` 120, and which does not remove the regrowth (the cache refills at ~7 GB/day wherever it lives); (iii) give the cut a home beside `disk-show`, which measures and never deletes. (iii) is the only one that is both mine to take and verifiable inside one run, and it is the one that makes (i) and (ii) cheap later: the projection below is the number both of them are arguing about.
- **Read-only, and the prune command is text.** `disk-show` set the shape: the tool measures, prints `find <cache-dir> -type f -atime +2 -print -delete` as a string and never runs it, so reclaiming stays a deliberate act by whoever reads the report. Three static guards pin it — `-delete` never appears as a quoted argv element, the only subprocess in the file is the read-only `find` walk, and there is no delete/write API in it at all.
- **Cold is decided by ACCESS time, growth by MODIFICATION time.** The root mount is `rw,relatime`, so atime bumps on the first access after 24 h and "not accessed in 48 h" means "nobody read it in two days" — the property a trim needs and the one mtime cannot give (run 785's own reasoning, now written down instead of remembered). The growth rate comes from bytes *written* in the last 24 h, which needs no second sample and therefore **no state file** — a tool that must be told yesterday's number by its own yesterday is a tool that cannot answer on its first run.
- **`48 h` and `-atime +2` are one number in two spellings**, held together by a single constant: the window that decides warm/cold is the window the printed command encodes, so the report and the command it prints cannot drift apart.

### Measurement (live box, 2026-10-08, taken with the tool itself)
- `./tools/cache-show --format json` → exit **0**: `mount / total 105,089,261,568 · used 75,212,972,032 · avail 24,490,803,200 · use_pct 75.4 · breach false`; `cache /tmp/gocache 26,891,002,127 B / 119,106 files`; `atime 48h → recent 16,429,480,169 B / 72,771 files, stale 10,461,521,958 B / 46,335 files`; `mtime → written 24 h 7,065,323,144 B / 31,148 files, written 7 d 26,891,002,127 B / 119,106 files`; `growth_bytes_per_day 7,065,323,144`; `days_to_threshold 2.7`; `prune_command find /tmp/gocache -type f -atime +2 -print -delete`.
- **The projection is this run's headline**: at 7,065,323,144 B/day the watched mount reaches 90 % in **2.7 days**, and the cold half is **10,461,521,958 B** — i.e. one hand-run cut buys ~1.5 days while the tool says so in advance. The **10,461,521,958 B** cold figure is 6.4 % smaller than the pre-cut world of run 785 (`31.5 GB` atime-≥48 h) because this box has been running `go test` every six minutes since; the number moves, which is exactly why it is measured per run rather than written down.
- **The walk is cheap enough to run every time**: `find /tmp/gocache -type f -printf '%A@ %T@ %s\n'` → **119,106 lines in 0.54 s**, against the **30 s** ceiling in the tool, so a cache ten times larger still answers inside the budget.

### Verification
- **`bash tests/test_cache_show.sh` → 105 passed / 0 failed (rc 0)** — arg contract (10 probes: unknown flag, `--format xml`, `--format`/`--threshold`/`--cache-dir`/`--mount` without a value, `--threshold abc|0|101`, each exit **2** with stdout empty); the exact documented JSON key set; breach **95 % → 3**, boundary **90 % at `--threshold 90` → 3** (`>=`, and 91 → 0); precedence **4 > 3 > 0**; the cold/warm split against a synthetic cache whose **`reused` file (written 3 days ago, read today) is warm by atime and cold by mtime** — the one shape that separates the clocks; projection arithmetic asserted exactly (**(900−100)/400 = 2.0**, **(500−100)/400 = 1.0**, `0.0` when already breached, `null` when growth is zero); hook-file errors → exit **1** with nothing on stdout; the live box content-agnostic (exit 0 or 3, `/` watched, no fixture residue); three static guards; and section **Z** (the (128) ledger family: 6 sandboxes, directory-only readers, `cleanup()` removing everything).
- **3 mutations, all precondition-asserted and all diverging**: **M1** `return 3 → 0` in `decide()` (real 3 / mutant 0, still 0 on clean and 4 on cannot-verify — surgical); **M2** `>= → >` on the threshold comparison (real 3 / mutant 0 on the boundary, still 3 on a real 95 %); **M3** the cold split moved from **atime to mtime** (real reads **650 B** warm, mutant **400 B** — it calls the 3-day-old reused file cold, i.e. hands the next trim a working set — while the breach world stays 3/0 as before).
- **`tests/test_registry_coverage.sh` → 463 passed / 1 failed, then the one red refreshed and re-run** — the single failure was **G38**, which pins `cl_other` as a *scope claim about this tree* (22 → **24**: `cache-show`'s `--cache-dir DIR` and `--mount PATH` are pairs no metavar class can judge invalid without asking the tool). Refreshed with its reasoning, alongside the two refreshes this suite's own rules demand of any run that adds a tool: **C28g** (option-list scope **23 → 24**, the pin's own history note extended to its fifth landing) and REGISTRY's **`- Live:` suite figure 92 → 93** (section F). Re-run after the refresh: **464 / 0**, `violations=0`, `fig_claim=93 fig_disc=93 fig_stale=0`.
- **`./tools/regression-run --list` → 93 suite(s) discovered**; `./tools/system-status` → `queue-source [OK]`, `investor-duty [OK] owed=0 unread=0 unreplied=0 open=0`, `git-tree` clean at the time of writing (re-read at close); `./tools/version-check` → `OK: dev=0.4.28 prod=0.4.28 (match)` — **0.4.28 in both environments, and 0.4.29 is never quoted as a production version** (investor's 2026-10-05 correction).

### Notes
- **STEP 0 was paid first and owed nothing**: `SELECT count(*) FROM messages WHERE direction='investor_to_agent' AND read=0` → **0** on `/opt/startup/dev/data/messages.db` and **0** on `/opt/startup/prod/data/messages.db`; `./tools/inbox-status` → **rc 0**, *"OK - nothing owed (0 unread, 0 open entries all replied)"* (last reply dev **144** / prod **109**); `INBOX.md` **80 entries, 80 handled, 0 open** — no row written, nothing marked read, `INBOX.md` untouched, re-read with the same three results at close. **(93)/(97)(a)** still stand on the investor's hiring decision (`hiring/queue/marco-steiner.json` queued, untouched) and the `SOA:gladex.de [WARN] … mname=placeholder (NEEDS-INVESTOR open)` row is still the investor's.
- **Scope — a tool and two suites, no app code, so no reviewer gate and no promote**: paths changed are `tools/cache-show` (new), `tools/REGISTRY.md` (its section + the Live figure refresh), `tests/test_cache_show.sh` (new), `tests/test_registry_coverage.sh` (three scope refreshes), `CHANGELOG.md` (this entry) and `agent-logs/PROGRESS.md` (its carrier). Nothing under `app/src/php` → dev and prod both stay **0.4.28**. **No DNS write, no mail sent, no unit restarted, `crontab -l` untouched, spend 0.00** of the 5.00 October allowance (`budget-show` → Allowance 5.00 / Spent 0.00 / Remaining 5.00), free `*-free` model only (`opencode/mimo-v2.6-flash-free`), no key configured, **no secret or PII in any prompt, file or commit**.

### Queue
Authoritative next-candidate queue lives in `agent-logs/PROGRESS.md`, in the newest run entry's `Queue — next small step` bullet — it is not duplicated here, and has not been since `[0.4.67]`. To read the live queue, open `PROGRESS.md`; to know whether that is still true, run `tools/queue-source-check`.
> build · mimo-v2.6-flash-free
$ sleep 60; wc -l /tmp/opencode/regcov2.out; tail -3 /tmp/opencode/regcov2.out
244 /tmp/opencode/regcov2.out
PASS: C34b plant: and it still carries a number, so there is still something to be stale
PASS: C34c plant: and the OLD ^- Live spelling cannot see it at all - the (120) defect
PASS: C34d plant: and it now has no date anywhere on its line
> build · mimo-v2.6-flash-free
$ date -u; sleep 90; wc -l /tmp/opencode/regcov2.out; tail -2 /tmp/opencode/regcov2.out
Thu Oct  8 04:00:23 UTC 2026
317 /tmp/opencode/regcov2.out
PASS: C40a non-vacuity: the definition line was located in this suite
PASS: C40b plant: the comment beside the definition is the stale one now
> build · mimo-v2.6-flash-free
$ sleep 150; wc -l /tmp/opencode/regcov2.out; tail -2 /tmp/opencode/regcov2.out
466 /tmp/opencode/regcov2.out

=== Results: 464 passed, 0 failed ===
> build · mimo-v2.6-flash-free
Registry coverage green: **464 / 0**. Now the full regression run in background while I prepare the PROGRESS entry.
$ cd /data/repo && ./tools/regression-run > /tmp/opencode/regrun786.out 2>&1; echo "RC=$?"; tail -5 /tmp/opencode/regrun786.out
Command moved to the background (shell ID: sh_119af2d7a001A0BuvTZd3F8nAD).
Output is streaming to: /root/.local/share/opencode/shell/fa3e85b2145a60c127b5c9261e76a7ba6d520299/sh_119af2d7a001A0BuvTZd3F8nAD.out
$ cd /data/repo && grep -n "^## \|^### Queue" agent-logs/PROGRESS.md | tail -8; echo "==="; grep -n "Queue — next small step" agent-logs/PROGRESS.md | tail -3; echo "==="; wc -l agent-logs/PROGRESS.md
12713:## Builder — run 59 — `PY30X`'s two `mine()`-side INPUTS given teeth: M48/M49 landed ([0.4.248], step (ao))
12964:## Builder — run 60 — `PY30SRC`'s `BODIES` filter given teeth: M50/M51 landed ([0.4.249], step (ap))
13236:## Builder — run 62 — queue item (aq) INDEPENDENTLY RE-MEASURED AND CLOSED: its code, registry and `[0.4.250]` reached HEAD through two OTHER commits while its own queue record never landed (the shift rate-limited out at 13:11+02:00), so this entry carries the record and clears `queue-source-check`'s two violations (R8 by landing, R10 by landing in the same commit as the changelog it records) — every (aq) number re-taken on the committed bytes (`--mutations` **566/0**, defaults **340/0** plain · **337/0** live-free, reader **759/0**), both shapes re-planted from needles derived HERE (**313/9**, **336/4**), and the pre-install blindness reproduced (**528/10**, **551/5**) — closing changelog **`[0.4.251]`**, no suite, tool or code file changed
13483:## Builder — run 63 — queue item (ar) LANDED: `LINE`, the `DIAGNOSIS` format string BOTH step-2 heredocs define byte-identically (`PY30PLANT` line 915, `PY30SWAP` line 970), the last un-named input of either writer, given teeth — **M54** a disclosed **nine-red** and **M55** a disclosed **four-red**, planted as MULTI-RED because the probe's **ENUMERATION** shows no single-red shape exists for either input, and the teeth read the other way in the **twenty-fourth** reading (two superseded teeth runs and M53's pin-check compose shadow disclosed, not smoothed over) — closing changelog **`[0.4.252]`**, `--mutations` **566 → 576/0** plain · **563 → 573/0** live-free, defaults unchanged at **340/0** plain · **337/0** under `IPDRIFT_NO_LIVE=1`, teeth **574/2 (rc 1)**
13788:## Builder — run 64 — queue item (as) LANDED: `ORDER`, the `public_report_class` return order `('public', 'private-or-cgnat', 'unparseable')` this file's step-2 heredoc (`tests/test_ip_drift_cron.sh:914`) carries, the last un-named input of `PY30PLANT`'s `ALSO:` join, judged by measurement BEFORE it was planted — the pure reorder **SURVIVES** (recorded as a survivor, not planted, the way `M44` was) and the **entry drop** reddens **2**, landed as **M56** — closing changelog **`[0.4.253]`**, `--mutations` **576 → 581/0** plain · **573 → 578/0** live-free, defaults unchanged at **340/0** plain · **337/0** under `IPDRIFT_NO_LIVE=1`, teeth **580/1 (rc 1)**
14047:## Builder — run 765 — queue item (at) CLOSED BY RECORD: `IPDRIFT_MUT_SUBSET` — the `--mutations` subset gate under which the reader suite's plant roster actually runs, the **eleven** mutant labels `tests/test_system_status_ip_drift.sh` names in its own `--help`, whose CODE the loop's sweep commit `3fe2ff1` carried with no record anywhere — so the step was **the record itself** (plus run 64's two stale items measured and closed) — closing changelog **`[0.4.254]`**, plain **759/0** · `--mutations` **824/0** (65 `ok` sites across 13 plant sites), teeth **822/2 (rc 1)**, all three files record-only, no code changed
14160:## Builder — run 784 — queue item (au) LANDED: **the (at) step-(4) TEETH reading moved in-file, beside the two pins it proves, and the reader suite's stale header lead corrected** — `nine mutants of THIS FILE's own source` → `eleven … : nine` (nine src pins M14–M22 **plus** the two ORDER pins M23/M24 = the eleven `--help` has always said), **comments/wording only** — closing changelog **`[0.4.255]`**, default **759/0** · `--mutations` **824/0** (65 `ok`, **0** FAIL) re-taken *after* both edits, teeth re-run by this run → clean **759/0** + battery **822/2** with both reds on the redirected pins, targets still lines **974**/**1358**
14311:## Builder — run 785 — queue item (av) LANDED: **the disk reclaimed before 100 % — an access-time cut of the Go build cache pulled `/` from 97.1 % / 2.9 G free to 75.2 % / 24.7 G free**, working set kept and every gate re-proved afterwards — closing changelog **`[0.4.256]`**, `find /tmp/gocache -type f -atime +2 -delete` → **108,378 files / 20,783 MiB** gone, `go test ./... -count=1 -short` **rc 0 / 28 s** *after* the cut, `disk-show` **rc 0** (was **rc 3 BREACH**), `system-status` **rc 0** with `go-tests [OK]` + `go-compile [OK]`, `version-check` **0.4.28 = 0.4.28**
===
13393:**Queue — next small step (read this first):** one item, the (aq) bullet's successor by
13663:**Queue — next small step (read this first):** one item, (ar) is DONE (M54/M55 are in the
13941:**Queue — next small step (read this first):** one item, (as) is DONE (M56 is in the
===
14448 agent-logs/PROGRESS.md
> build · mimo-v2.6-flash-free
→ Read repo/agent-logs/PROGRESS.md [offset=14311, limit=60]
> build · mimo-v2.6-flash-free
$ cd /data/repo && ./tools/regression-run --list | wc -l; git rev-list --count origin/main..main; git log --oneline -1; git status --porcelain
95
15
cb50cdd run 785: PROGRESS close readings for f95a3ae — qsc rc 0 [0.4.256] (R8+R9 6 path tokens), test_queue_source 278/0, repo-lint 261/0, system-status rc 0 git-tree OK go-compile OK (commit f95a3ae), red-watch 23/7 unchanged post-commit, disk-show rc 0 (75.3%), STEP 0 re-read 0/0, no push (main ahead 14)
 M CHANGELOG.md
 M tests/test_registry_coverage.sh
 M tools/REGISTRY.md
?? tests/test_cache_show.sh
?? tools/cache-show
$ cd /data/repo && ./tools/disk-show 2>&1 | tail -6; df -B1 / | tail -1
pip-cache                absent  /root/.cache/pip
  apt-archives             0 B  /var/cache/apt/archives
  loop-scratch             2467231325 B  /tmp/opencode
  reclaimable total: 33561245353 B across 13 paths (0 unmeasured)

OK: all 2 watched path(s) below 90% (worst / at 75.6%)
/dev/mapper/ubuntu--vg-ubuntu--lv 105089261568 75336736768 24367038464  76% /
> build · mimo-v2.6-flash-free
$ cd /data/repo && ./tools/regression-run --list | head -3; echo ...; ./tools/regression-run --list | tail -3; echo "==="; ./tools/regression-run --help 2>&1 | head -20
regression-run  repo=/data/repo  tests=/data/repo/tests  2026-10-08T04:04:49Z
  test_app_contrast.php
  test_app_contrast_meta.php
...
  test_vhost_ui_smoke.sh
  test_webmail_session_routing.php
result: 93 suite(s) discovered, none run (exit 0)
===
usage: regression-run [-h] [--format {human,json}] [--tests-dir TESTS_DIR] [--only SUBSTR] [--timeout SECONDS] [--list] [--log-dir DIR] [--baseline FILE] [--save-baseline FILE] [--note TEXT]

Run every test suite in the repo and total the result.

options:
  -h, --help            show this help message and exit
  --format {human,json}
                        output format (default: human)
  --tests-dir TESTS_DIR
                        directory of suites (default: <repo>/tests, or
                        $GLADEX_TESTS_DIR)
  --only SUBSTR         run only suites whose filename contains SUBSTR
  --timeout SECONDS     per-suite timeout (default: $REGRESSION_TIMEOUT or
                        600)
  --list                discover and list suites, run nothing (exit 0)
  --log-dir DIR         write each suite's raw output to DIR/<suite>.log
  --baseline FILE       compare this run against a file written by --save-
                        baseline
  --save-baseline FILE  record this run's per-suite counts for a later
                        --baseline
> build · mimo-v2.6-flash-free
$ cd /data/repo && grep -n "verify38" -r tests tools PROGRESS.md agent-logs 2>/dev/null | head -10
tools/REGISTRY.md:6409:`[0.4.96]`'s fix, committed instead of left at `/tmp/opencode/verify38.sh` where
tools/go-compile-drift-verify:8:# five-scenario harness that was left at /tmp/opencode/verify38.sh - outside the
agent-logs/identity-sofia.md:15011:preceding 90 minutes**: `repo_copy` (5.0 G, 27 Sep), `verify38` (761 M, 27 Sep),
agent-logs/identity-sofia.md:15549:preceding 90 minutes**: `repo_copy` (5.0 G, 27 Sep), `verify38` (761 M, 27 Sep),
agent-logs/identity-sofia.md:33005:2872:## [0.4.97] - 2026-09-27 — queue item (40): the five-scenario harness that *proves* `[0.4.96]`'s drift fix lived only at `/tmp/opencode/verify38.sh`, outside the repo, undocumentable and deleted by the next reboot with the proof — promoted to a committed on-demand tool, `tools/go-compile-drift-verify` (`--help`, `--list`, `--only`, `--format human|jsonl`, exit codes `2 > 3 > 1 > 0`), which is **not** in `tests/` because `regression-run` would otherwise pay five clones on every run
agent-logs/identity-leon.md:11733:# five-scenario harness that was left at /tmp/opencode/verify38.sh - outside the
agent-logs/PROGRESS.md:5675:- **Verification — five scenarios, all in throwaway clones, `/data/repo` never written to by the harness** (`/tmp/opencode/verify38.sh`): **S1 normal → 82/0, 0 notes** · **S2 forced drift (the pre-fix scenario, re-run) → 82/0 + note** (`first read ae04bd2, row 2f96896 … both judge 2f96896`) so same-scenario pre/post is **81/1 → 82/0** · **S3 row names an older existing commit → same-commit FAIL, no note** · **S4 row names a newer broken commit** — ground truth measured first (`Y(broken) ok=False`, `X(clean) ok=True`) → drift fires and **assertion 1 goes red** (`got ok=False, exit 1`), i.e. the pinned re-read cannot bless a broken commit · **S5 row names a nonexistent commit → same-commit FAIL, no note** (`cat-file -e`). S3–S5 each total **79 / 3**: the other two reds are the plant's own assertions in earlier sections (the mutation rewrites the row everywhere by construction), not weakened checks. Standalone on the real repo **82 / 0, no note**; `bash -n` clean; suite md5 **`dcfc9e54fcb54f1c032a1c44ca30a54a`**.
agent-logs/PROGRESS.md:5683:- **Next-candidate queued, not actioned**: carry items (2)–(8), (14), (15), (17), (18), (20), (22), (23), (25), (28), (32), (36) and (39) from the 16:45Z entry unchanged — item **(38) struck — ACTIONED by `[0.4.96]`** (captured FAIL line `… (commit 343037d)', linter: e7221f3`, both reads moved ahead of the assertions, drift branch with `cat-file -e` + `merge-base --is-ancestor` and a pinned re-read carrying `LIVE_SHA`/`LIVE_OK`/`LIVE_RC`; suite held at **82**, same-scenario pre/post **81/1 → 82/0**, five scenarios, regression **54 / 4593 / 0**). **(40) new from this run**: the five-scenario harness that *proves* the fix lives only at `/tmp/opencode/verify38.sh` (and the race probe at `/tmp/opencode/race38.sh`) — it clones the repo five times and takes ~5 minutes, so it is neither in `tests/` (where `regression-run` would discover and run it every time) nor documented anywhere but here, and a reboot deletes it along with the proof. Worth a decision: commit it as an on-demand tool under `tools/` with `--help` and a `REGISTRY.md` entry (my preference — it is a verifier, not a unit test), or shrink it to one deterministic in-suite drift case, which would move the assertion count off 82 and therefore needs the arithmetic redone. Not actioned this run because the step was already a capture, a fix, a self-defect, five scenarios, a regression and two docs.
agent-logs/PROGRESS.md:5705:## 2026-09-27T18:32Z main-loop run — STEP 0 answered FIRST (0 unread at start; fresh-probe reply **dev 90 / prod 56**), then **the investor's order "finish Immich setup via API NOW" arrived mid-run and was executed the same run — IMMICH NOW HAS 8 ACCOUNTS, ALL VERIFIED** (admin + the seven identities, admin role proved 200/403/401, report **dev 92 / prod 57**); queue item **(40) EXECUTED**: the five-scenario proof left at `/tmp/opencode/verify38.sh` became a committed on-demand tool `tools/go-compile-drift-verify` (`--help`, `--list`, `--only`, `--format human|jsonl`, exit `2 > 3 > 1 > 0`) that is deliberately **not** in `tests/` — full run **5 selected / 5 met / 0 unmet / 0 unverified / 289s / exit 0**; CHANGELOG **`[0.4.97]`** + **`[0.4.98]`**, regression **54 / 4593 / 0, exit 0**
agent-logs/PROGRESS.md:5713:- **The step (queue item 40), and why `tools/` and not `tests/`**: `[0.4.96]` closed with its own note that the five-scenario harness proving the fix lived only at `/tmp/opencode/verify38.sh` — outside the repo, undocumented, deleted by the next reboot along with the proof. It is a **verifier**, not a unit test: five clones and **289s**, and `regression-run` discovers `<repo>/tests` only, so putting it there would charge every run and move the totals every entry quotes. `tools/go-compile-drift-verify` (extensionless, `#!/usr/bin/env bash` → `repo-lint` dispatches as `bash`, md5 **`21bcdc94e6c2604aa24573f7bd5b123b`**, 429 lines) takes `--help` / `--list` / `--only S1[,S2…]` / `--format human|jsonl` / `--keep` / `--out DIR` and exits `0` all met / `1` an expectation unmet / `2` bad arguments / `3` could not verify, precedence **`2 > 3 > 1 > 0`** — the same shape `regression-run` uses, because a run in which part did not execute is not a verdict about the part that did.
$ cd /data/repo && grep -o "(aw)" agent-logs/PROGRESS.md | head -3; echo "---ay"; grep -o "(ay)" agent-logs/PROGRESS.md | head -3; echo "---queue-source-check on tokens"; ./tools/queue-source-check 2>&1 | head -5
(aw)
---ay
---queue-source-check on tokens
queue-source-check: FAIL - 1 violation(s): [0.4.257] is not named in agent-logs/PROGRESS.md - the newest CHANGELOG entry has no record in the authoritative file
  - [0.4.257] is not named in agent-logs/PROGRESS.md - the newest CHANGELOG entry has no record in the authoritative file
> build · mimo-v2.6-flash-free
$ cd /data/repo && grep -n "(aw)" agent-logs/PROGRESS.md | head -5
14283:  **(aw)** if a compile is in flight: the teeth family's next file, **`tests/test_healthcheck.sh`
$ cd /data/repo && grep -rn "(aw)" --include=*.md . | grep -v identity- | head -10
./agent-logs/PROGRESS.md:14283:  **(aw)** if a compile is in flight: the teeth family's next file, **`tests/test_healthcheck.sh`
> build · mimo-v2.6-flash-free
→ Read repo/agent-logs/PROGRESS.md [offset=14262, limit=48]
> build · mimo-v2.6-flash-free
$ cd /data/repo && sed -n '1,40p' tools/cache-show
#!/usr/bin/env python3
"""
cache-show - Report the Go build cache, the access-time cut it is waiting for,
and how long the root filesystem has left (read-only watchdog).

WHY THIS EXISTS (queue item (ax), 2026-10-08): `/` filled to 97.1 % on
2026-10-08 and the reclaim was a hand-run `find /tmp/gocache -type f -atime
+2 -delete` - a rule derived from scratch by the run that took it, because no
file owned it. The cache regrows (~7 GB/day measured that day), so the cut is a
stopgap until the rule has a home. This tool IS that home: it measures what a
trim would reclaim, reports the growth that undoes it and projects the day the
watched mount reaches the threshold. The three questions a next run would
otherwise re-derive - how big, how much is cold, how long until 90 % - are one
command here.

    cache-show never deletes, moves or writes anything - it only reports.
    Reclaiming is a separate, deliberate action: the prune command is PRINTED
    as text (never executed), so the decision stays with whoever reads it.

One `find` pass over the cache directory yields every cache number: apparent
bytes per file plus its access time and modification time. Access time is what
a cache trim needs (relatime on this box bumps atime on the first access after
24 h, so "not accessed in 48 h" means "nobody read it in two days"), and
modification time is what the growth rate needs (bytes WRITTEN per day, which
needs no second sample and therefore no state file). The walk is bounded by
WALK_TIMEOUT_S; a walk that does not finish is `cannot verify`, never a pass.

Usage:
    cache-show [--format {json,human}] [--threshold PCT] [--cache-dir DIR] [--mount PATH]

Exit codes (the caller keys off these, not off the text):
    0 - ok: the cache and the mount were both measured and the mount is below
        the threshold (`ok` in the JSON is true iff this code is returned)
    1 - unexpected internal error, or an unreadable test-hook file
    2 - invalid arguments (unknown flag, missing value, --format outside
        json|human, --threshold outside 1-100)
    3 - threshold breached: the watched mount is at or above it
    4 - cannot verify: the watched mount path does not exist, or the cache
        walk failed or timed out - NOT a pass, never reported as one
    Precedence: 2 > 1 > 4 > 3 > 0.
$ ls -la /tmp/opencode/verify38.sh /tmp/opencode/race38.sh 2>&1; du -sh /tmp/opencode 2>/dev/null; du -sh /tmp/gocache 2>/dev/null; df -B1 / | tail -1
-rw-r--r-- 1 root root 1680 Sep 27 19:07 /tmp/opencode/race38.sh
-rw-r--r-- 1 root root 5566 Sep 27 19:20 /tmp/opencode/verify38.sh
2.5G	/tmp/opencode
26G	/tmp/gocache
/dev/mapper/ubuntu--vg-ubuntu--lv 105089261568 75341828096 24361947136  76% /
> build · mimo-v2.6-flash-free
$ cd /data/repo && grep -n "path token\|resolve\|new tool\|new suite" tools/queue-source-check | head -30
67:        marks it new (`new tool` / `new suite` / `(new)`)
128:        merely QUOTES `new tool` is excused for every token it carries - the
216:                      resolves to /data/repo - the same trap repo-lint
261:NEW_MARKERS = ("new tool", "new suite", "(new)")
414:    without the anchor git would resolve `CHANGELOG.md` against the caller's
754:                "it new in the bullet (`new tool` / `new suite` / `(new)`) or "
792:        "%d PROGRESS bullet(s), %d path token(s)"
832:    the diagnosis. That is CHANGELOG [0.4.62]'s defect, and a new tool does not
> build · mimo-v2.6-flash-free
→ Read repo/agent-logs/PROGRESS.md [offset=14438, limit=11]
> build · mimo-v2.6-flash-free
> build · mimo-v2.6-flash-free

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

[Thu Oct  8 06:12:57 2026] 127.0.0.1:59394 Accepted
[Thu Oct  8 06:12:57 2026] 127.0.0.1:59394 Closing
[Thu Oct  8 06:12:58 2026] 127.0.0.1:59402 Accepted
[Thu Oct  8 06:12:58 2026] 127.0.0.1:59402 Closing
[Thu Oct  8 06:12:58 2026] 127.0.0.1:59404 Accepted
[Thu Oct  8 06:12:58 2026] 127.0.0.1:59404 Closing
[Thu Oct  8 06:12:58 2026] 127.0.0.1:59416 Accepted
[Thu Oct  8 06:12:58 2026] 127.0.0.1:59416 Closing
[Thu Oct  8 06:12:58 2026] 127.0.0.1:59424 Accepted
[Thu Oct  8 06:12:58 2026] 127.0.0.1:59424 Closing
[Thu Oct  8 06:12:58 2026] 127.0.0.1:59430 Accepted
[Thu Oct  8 06:12:58 2026] 127.0.0.1:59430 Closing
[Thu Oct  8 06:12:58 2026] 127.0.0.1:59446 Accepted
[Thu Oct  8 06:12:58 2026] 127.0.0.1:59446 Closing
[Thu Oct  8 06:12:58 2026] 127.0.0.1:59460 Accepted
[Thu Oct  8 06:12:58 2026] 127.0.0.1:59460 Closing
[Thu Oct  8 06:12:58 2026] 127.0.0.1:59462 Accepted
[Thu Oct  8 06:12:58 2026] 127.0.0.1:59462 Closing
[Thu Oct  8 06:12:58 2026] 127.0.0.1:59476 Accepted
[Thu Oct  8 06:12:59 2026] 127.0.0.1:59476 Closing
[Thu Oct  8 06:12:59 2026] 127.0.0.1:59482 Accepted
[Thu Oct  8 06:12:59 2026] 127.0.0.1:59482 Closing
[Thu Oct  8 06:12:59 2026] 127.0.0.1:59484 Accepted
[Thu Oct  8 06:12:59 2026] 127.0.0.1:59484 Closing
[Thu Oct  8 06:12:59 2026] 127.0.0.1:59486 Accepted
[Thu Oct  8 06:12:59 2026] 127.0.0.1:59486 Closing
[Thu Oct  8 06:12:59 2026] 127.0.0.1:59494 Accepted
[Thu Oct  8 06:12:59 2026] 127.0.0.1:59494 Closing
[Thu Oct  8 06:12:59 2026] 127.0.0.1:59500 Accepted
[Thu Oct  8 06:12:59 2026] 127.0.0.1:59500 Closing
[Thu Oct  8 06:12:59 2026] 127.0.0.1:59504 Accepted
[Thu Oct  8 06:12:59 2026] 127.0.0.1:59504 Closing
[Thu Oct  8 06:14:24 2026] 127.0.0.1:37324 Accepted
[Thu Oct  8 06:14:24 2026] 127.0.0.1:37324 Closing
[Thu Oct  8 06:14:24 2026] 127.0.0.1:37332 Accepted
[Thu Oct  8 06:14:24 2026] 127.0.0.1:37332 Closing
[Thu Oct  8 06:14:24 2026] 127.0.0.1:37336 Accepted
[Thu Oct  8 06:14:24 2026] 127.0.0.1:37336 Closing
[Thu Oct  8 06:14:24 2026] 127.0.0.1:37342 Accepted
[Thu Oct  8 06:14:25 2026] 127.0.0.1:37342 Closing
[Thu Oct  8 06:14:25 2026] 127.0.0.1:41840 Accepted
[Thu Oct  8 06:14:25 2026] 127.0.0.1:41840 Closing
[Thu Oct  8 06:14:25 2026] 127.0.0.1:41848 Accepted
[Thu Oct  8 06:14:25 2026] 127.0.0.1:41848 Closing
[Thu Oct  8 06:14:25 2026] 127.0.0.1:41862 Accepted
[Thu Oct  8 06:14:25 2026] 127.0.0.1:41862 Closing
[Thu Oct  8 06:14:25 2026] 127.0.0.1:41866 Accepted
[Thu Oct  8 06:14:25 2026] 127.0.0.1:41866 Closing
[Thu Oct  8 06:14:25 2026] 127.0.0.1:41872 Accepted
[Thu Oct  8 06:14:25 2026] 127.0.0.1:41872 Closing
[Thu Oct  8 06:14:25 2026] 127.0.0.1:41878 Accepted
[Thu Oct  8 06:14:25 2026] 127.0.0.1:41878 Closing
[Thu Oct  8 06:14:25 2026] 127.0.0.1:41892 Accepted
[Thu Oct  8 06:14:25 2026] 127.0.0.1:41892 Closing
[Thu Oct  8 06:14:25 2026] 127.0.0.1:41896 Accepted
[Thu Oct  8 06:14:25 2026] 127.0.0.1:41896 Closing
[Thu Oct  8 06:14:25 2026] 127.0.0.1:41906 Accepted
[Thu Oct  8 06:14:25 2026] 127.0.0.1:41906 Closing
[Thu Oct  8 06:14:47 2026] 127.0.0.1:56628 Accepted