Morning check of the 2026-10-22 away window

docs/retrospectives/2026-10-23-morning.md

@@ -0,0 +1,213 @@
+# 2026-10-23 morning check: the away window of 22 October, 13:48 to 19:48 UTC
+
+Times are UTC. Window: entered 2026-10-22 13:48:44, set to end 19:48:44,
+grants `worktree-worktree-ar50-vegetation` (increment 91, the land-cover
+forest and open-land split) and `worktree-lsm-energy-balance` (increment 92,
+the one-layer snow energy balance) (`.git/harness/windows.jsonl`,
+`unattended.json`). It ran out on its own; Ola's next prompt came at
+2026-10-23 08:18:14. Sources: the main session's transcript
+`c38521c2-eceb-4357-9777-0b262dbe53b1.jsonl` ("line N"), its subagent
+records under `c38521c2-…/subagents/`, `.git/harness/queue.jsonl`,
+`.claude/current-task/session.md`, and `git log` of the three branches named
+below. Ola told the main session he would be out 5 h (line 1438); the window
+he set was 6 h.
+
+## 0. Duties owed
+
+`python3 tools/duties.py`: "morning check (1 away window ended since,
+latest 2026-10-22 19:48 UTC)". Owed from 19:48:44; this check was spawned at
+2026-10-23 08:23:42 (line 3843): **12 h 35 min**, of which 12 h 30 min fell
+before Ola's return. Nothing woke the idle session at 19:48 (proposal 4
+below). After Ola's return the spawn came 5 min later, behind the two pushes
+and the enqueue he approved at 08:20:43 (line 3624). The weekly check of the
+private ideas file is next due 2026-10-26.
+
+## 1. Idle timeline
+
+Agent spans from the subagent records (first and last timestamp of each);
+commits from `git log` of `worktree-lsm-energy-balance`,
+`worktree-worktree-ar50-vegetation` and `worktree-lsm-forest-plan`.
+
+| Stretch | Length | What ran |
+|---|---|---|
+| 13:48 to 16:49:45 | 3 h 01 min | 91: green, code review r1, red fix, green fix, master merge, prose edits, code review r2 (APPROVED `dac60f4d`), record `8efa120b`. 92: red suite, design r3, red update, green (3 stages), code review r1, tiny-pack fix through design, red and green, code review r2 (APPROVED `2434aac1`), `@perf` acceptance (station test failed, `331215d5`), diagnosis (`c565d34a`). Triangle-budget steep-terrain probe. Forest-case plan, its review and wordings (`ec144949`). The design of increment 93, the three-layer snow pack (`f9c3d977`), its review and one fix. At no point did more than 60 s pass with no agent running. |
+| **16:50:28 to 19:48:44** | **2 h 58 min** | **Nothing.** No agent, no commit on any ref (`git log --all --since 16:51 --until 19:49` is empty). The Stop hook fired at 16:50:21 (line 3520); the main session set `NOW: idle` (line 3536). |
+| 19:48:44 to 08:18:14 | 12 h 30 min | Outside the window; waiting for Ola. |
+
+**A fallback existed at 16:50.** `.claude/REQUIRED-READING.md`, *Unattended
+mode*: "idle is accepted only when no such item exists" (an item that needs
+no ruling and writes no governed path). `session.md`'s own `QUEUE` line, the
+one that says "Fallback for the last hour: none left", lists eight lessons
+under "Next retro". Recording them is `@orchestrator`'s work in
+`docs/retrospectives/`, which `guard_governance.py`'s `GOVERNED` lists do not
+cover, and it needs no ruling; with every other item waiting only on Ola,
+none counted toward the cap of three. This check records them now, 15 h
+later. Second night running: `2026-10-22-night.md` §1 found the same, a
+queue said to be empty that was not.
+
+## 2. Refusals and false positives
+
+Four hook refusals in the main session (two with one cause), one in a
+subagent; no guard false positive.
+
+1. **13:48:04, `python3 tools/away.py --help`** (line 1447), by the main
+   session, refused: "Only Ola enters or leaves unattended mode". The rule
+   says no agent runs `away.py`, whatever the flag; the main session then read
+   the source with `grep` and `sed`. A true refusal, and an act outside the
+   main session's role, harmless.
+2. **14:08:42, `@tester` (agent `aaf46bd4`), a compound command**: a shell
+   heredoc edit of `tests/python/energy_reference.py`, three gates, then
+   `git commit -q --amend --no-edit -a`, chained with `;` and `&&`
+   (`queue.jsonl` entry 1; subagent line 1106). The guard's reason: an amend
+   of an unpublished commit passes only last after `cd` and `git add`, or as
+   `git -C <absolute path> commit --amend …` alone; "it may be redone in that
+   form". **Not a false positive**: an amend rewrites history, and the
+   command was not in the allowed form. The tester redid the edit with Edit
+   (an ordinary test file) and the amend as
+   `git -C /Users/…/lsm-energy-balance commit --amend --no-edit -a` (line
+   1202), as the rule allows; its handback says so (main line 1660).
+3. **15:20:01, `git push origin 8efa120b:refs/heads/worktree-worktree-ar50-vegetation`**,
+   by the main session (line 2745). **The cause is the short sha**:
+   `tools/push_grant.py`'s `granted_act` passes a push only when
+   `SHA.fullmatch(sha)`, with `SHA = re.compile(r"[0-9a-f]{40}")`; an
+   8-hex sha makes it return `None`. **The guard's stated reason is wrong
+   for this case**: it prints `NOT_BARE`, "the grant takes only one bare
+   command, every word written out (no variable, no &&, ;, pipe, redirect or
+   heredoc)", none of which the command had. The main session diagnosed it
+   from the mismatch and did not retry (line 2750); Ola's 08:20 yes, then
+   the full-sha push at 08:21:00 (line 3648). Cost: 91 waited from 15:20 to
+   08:21 for a grant it already had. Retro #38's Q-A (the refusal names the
+   failing word), still in `session.md` as an `ASK OLA` line, would have
+   named it.
+4. **15:23:03 and 15:52:11, spawn refused: "brief: edited; paste the output
+   of tools/brief.py unchanged"** (lines 2937, 3125). Both times the main
+   session had completed the brief's `Review rounds recorded: …; last:` line,
+   which `tools/brief.py:205` cuts at `MAX_VERDICT = 200` characters with no
+   mark, here mid-path ("`docs/increments/92-`"). The diff of each refused
+   prompt against the `brief.py` output is that one line. A true refusal,
+   but the cut looks like damage and invited the edit twice.
+
+## 3. Work out of role or beyond a ruling
+
+1. **The main session started increment 93 without Ola's yes.** At 16:27:00
+   (line 3340) it briefed `@architect` to write the three-layer design,
+   reading 92's §4.1 (the layers named as the fallback if the station test
+   failed) as covering it. Ola's ruling named the "one-layer snowpack energy
+   balance"; `.claude/REQUIRED-READING.md`: "A question whether to build at
+   all waits for Ola." `@reviewer` blocked it (line 3414, CHANGES REQUESTED
+   `f9c3d977`, finding 1); the main session agreed (line 3417) and wrote no
+   tests or code. Cost: three spawns, about 23 min. Ola said yes at 08:20
+   (line 3624, answer 3), so the design is not wasted. Same day as the sun
+   position moved under "radiation" without asking (Ola's memory note on
+   rulings covering what they name).
+2. **Review records written by `@architect`**: `1ac21210` (92 design r3) and
+   `dac60f4d` (91 code r1), at the main session's request, while
+   `docs/increments/README.md` step 4 says "the main session itself, not an
+   agent it spawns". Ola said yes to letting the owner write them at 13:32:37
+   (line 1023, answer 2, after the question at line 1019). The main session
+   then wrote `8efa120b`, `f35d6012` and `2fd8861b` itself, citing the rule
+   (line 2693). Rule text and practice disagree until proposal 1 lands.
+3. **`@developer` (agent `aba4af0d`) ran three `until ! pgrep -f …` wait
+   loops** that matched their own shell and never ended (its lesson, main
+   line 2381). Every brief forbids a polling loop.
+4. **`tools/duties.py --status` hung for 55 min** (13:48:58 to 14:43:44,
+   lines 1490 to 2197). `--status` is the status-line mode and reads the
+   session JSON from stdin (`tools/duties.py:216`); the Bash tool leaves
+   stdin open, so `stdin.read()` blocks. Checked: `sleep 30 | python3
+   tools/duties.py --status` is still blocked at 5 s; with `</dev/null` it
+   prints "Duties: unknown (JSONDecodeError…)" at once. The `session.md`
+   write chained after it never ran. Plain `python3 tools/duties.py` is the
+   agents' form.
+
+## 4. Reports and records
+
+- **Clock times written by hand were wrong.** `session.md`'s `ASK OLA` line
+  for the 91 push says "refused 18:2x"; it was written at 15:20:26 (line
+  2790) about the 15:20:01 refusal. `NOW` lines said "16:5x UTC" at 13:48
+  and "17:4x" at 14:43. The main session corrected itself at 16:18 (line
+  3310, "My earlier stamps were local time"); local time was UTC+2, so not
+  even that. The wrong stamp still stood in `session.md` this morning.
+- **`session.md` form**: the recap warns "0 QUEUE lines" (the line starts
+  `QUEUE, in order:`), and lines of 509 and 319 characters against 300.
+  Repeat of `2026-10-22-merges-335-337.md`.
+- **The morning report** (line 3620) said "From 14:00 to about 16:50: work
+  ran"; it ran from 13:48. It did report the 14:08 refusal as unchecked
+  rather than guess, and owned the 93 overstep. Otherwise it matches the
+  record above.
+
+## 5. Lessons recorded
+
+From handbacks in the window (main transcript line of each) and the main
+session's list:
+
+1. **(`@architect`, line 1160.) The tester's scratch-copy rule is unclear in
+   scope.** `.claude/agents/tester.md@acef2200:76`: "When the design names the fix by
+   code location, apply it to a scratch copy". A whole new feature listing
+   code locations is not "a fix"; the architect read it as bug fixes only.
+   Proposal 1 (c).
+2. **(`@reviewer`, line 942.) A cap computed as `min(Σ sw, …)` over bins
+   mixes areas**: the sum is per m² of triangle, the heat capacity per m² of
+   snow, so a bin could refreeze more water than it holds. Check that each
+   side of a `min` is per the same area. Also: a function reused by name
+   ("as snowfall") can carry side effects; read its comment.
+3. **(`@reviewer` line 1618, `@architect` line 2482.) A design's list of
+   prose edits was not checked off at green**; the code review found it.
+   Proposal 1 (b).
+4. **(`@architect`, line 2065.) `tools/check_citations.py` on a private
+   branch compares against public master** and flags 144 citations as at
+   risk, 2 of them the change's own.
+5. **(`@tester` line 3019, `@developer` line 3055.) The private repository's
+   build and scratch copy differ**: `tools/scratch_copy.py` sets
+   `PYTHONPATH` for `tin_engine`, not `rasputin_lsm` under `src/`; the
+   extension rebuilds with `uv pip install --python .venv/bin/python -e .`;
+   C++ tests need `rasputin_DIR`. Retro #38's Q-C covers part of this.
+6. **(`@reviewer`, line 3414.) It could not write its note file: no Write
+   tool.** Repeat of `2026-10-14-merge-46.md` finding 8, still open.
+7. **(`@tester`, line 2428.) Keep a throwaway implementation as a patch in
+   the scratchpad until the increment merges**; a mutant inside `#if` needs
+   `[[maybe_unused]]` under `-Werror`.
+8. **(`@perf` line 3209, `@architect` lines 1777, 2972, 3317, 3469.)**
+   Check the first output against observed values; a dropped coupled
+   component may have provided stability; look for every division by the
+   small quantity; a station test mixes snowfall with melt; a bound from a
+   difference states its direction.
+9. **(Main session.) `tools/brief.py`'s cut verdict line** (§2.4), **the
+   short-sha refusal** (§2.3), **the hung `--status`** (§3.4).
+
+## 6. Rule text and proposals
+
+`python3 tools/rule_sizes.py`: 9,494 words, unchanged since
+`2026-10-22-merges-335-337.md` (CLAUDE.md 1,187; REQUIRED-READING 1,884;
+increments README 1,007; PRINCIPLES 437; briefs/common 236; agents:
+architect 495, developer 211, orchestrator 482, perf 635, reviewer 570,
+tester 813; skills 346, 249, 308, 634). Its proposals 2 and 3 (−33) are not
+yet ruled. No new anecdote found in the rule files.
+
+1. **Rule text, net −24 words.**
+   - (a) `docs/increments/README.md` step 4: "`@reviewer` is read-only, so the
+     main session itself, not an agent it spawns, copies" becomes "The
+     persona that owns the file, or the main session, copies" (−3), and "The
+     main session applies them exactly" becomes "The recorder applies them
+     exactly" (−1). Ola's yes of 13:32; §3.2.
+   - (b) Step 3 gains "A design's list of prose edits is `@architect`'s,
+     made before code review." (+12). Lesson 3.
+   - (c) `.claude/agents/tester.md@acef2200:76`: "names the fix" becomes "names a bug fix" (+1).
+     Lesson 1.
+   - (d) `.claude/REQUIRED-READING.md`, *The harness*: cut "Ola's status
+     line is `tools/duties.py --status`." (−6); it is what invited §3.4.
+   - (e) Same section: cut "`guard_push.py` and `guard_governance.py` judge
+     what a command writes or pushes, as parsed by `tools/shell_scan.py`; a
+     line they cannot read is judged by its text." (−24), and "all the same"
+     after it (−3); the code states it.
+2. **`tools/brief.py`**: cut the verdict line at a word and end it with
+   "(cut)", so it reads as cut, not damaged. Evidence §2.4; about 2 lines.
+3. **A new increment file needs a quoted ruling**: `tools/brief.py` refuses
+   an `@architect` brief whose `--increment` does not yet exist unless an
+   `--ola` quote is given. Evidence §3.1; about 5 lines.
+4. **Before going idle in a window, schedule a wake-up at its end** to run
+   the morning check, so its retrospective is ready on Ola's return.
+   Evidence §0 and §1 (12 h 30 min owed while idle). No rule text if the
+   Stop hook's message names it.
+5. **`check_citations.py` compares against `private/master`** on a branch
+   holding private commits, as `push_grant.py` already tells them apart.
+   Lesson 4.