Stage 1 first: the shape of the day, a clock for the judges, and a readable gate line (#411, #412, #413)

[hugolytics]

Problem

On 2026-09-08, a real session showed three things wrong at once. First, the
ladder: an auto-started session asked "What do you want to get out of the
day?" (stage 2, Priorities) before any Stage 1 probe had run, then a Stage 1
question, then back to a stage-1-shaped question — 1 -> 2 -> 1, not the
1, 1, …, 2, 3 the design intends. Second, the clock: the user said "im
going to my own office in 2 hours, so thats a 30 minute commute" at 10:09
local, and the probe that came back asked "Will you be at your office by
9:30 AM to start work?" — forty minutes in the past relative to when they
were speaking, because neither judge knew what time it was. Third, once
several rows were open at once the gate line read as one criterion repeated
per row in a comma-separated run, unreadable as anything but noise.

What changed

#411 — Stage 1 before the hard user blocker. The run loop now consults
Stage 1 before first_hard_user_blocker() whenever the stage is still open,
so the shape of the day is elicited first and "what do you want out of the
day" is asked once Stage 1 closes. The branch this check used to skip is
unchanged once the stage has closed — a fully-open catalog still guards the
request and the frame as hard, user-owned requirements. The ladder is now
1, 1, …, 2, 3 by construction, matching the design (design doc addendum in
this branch, #411 named at the exact paragraph that states the ladder).

#412 — a clock for the judges. elicit() now requires a tz-aware now
and raises before any judge runs if it is naive. Both the coverage judge
(classify) and the probe judge (generate) receive now and
planning_day inside a "clock" object on every prompt, so a judgement
about whether a stated time is still reachable has something to compute
against. The regression gate for this is two deterministic unit tests that
read the sent JSON back and assert the "clock" key is present with the
right shape — not a stub that only proves an argument was passed.

#413 — a gate line grouped by row. _gate_line now groups every open
cell by its row and lists each row once, with its open criteria in
parentheses, instead of repeating the row's label once per open criterion.
It raises ValueError if a cell names a row or criterion outside the
catalog, rather than silently under-reporting what's open.

Evidence

• Ladder (timeboxing: the kernel asks the Priorities question before Stage 1 runs, so the ladder reads 1 → 2 → 1 and the shape of the day comes second #411): tests/unit/test_adaptive_stage1.py::test_the_shape_of_the_day_is_asked_before_the_priorities_question
  and ::test_the_priorities_question_is_still_asked_once_stage_one_closes
  assert the order directly.
• Clock, the gate (timeboxing: Stage 1 has no clock — 'in 2 hours' at 10:09 became 'by 9:30 AM' #412): tests/unit/test_elicitation_judges.py::test_classify_sends_the_clock_object
  and ::test_generate_sends_the_clock_object assert the sent prompt (not a
  stub's received args) carries "clock". These are what CI gates.
• Clock, the eval (timeboxing: Stage 1 has no clock — 'in 2 hours' at 10:09 became 'by 9:30 AM' #412), informational only: tests/evals/test_stage1_clock.py
  targets the exact mechanism of the live bug (fixed/contradictory, the
  real 09:30 work-start must rule, the office statement) and asks a
  non-contender model whether each probe reflects correct timing. Measured
  across seven resamples: 4/32 relevant probes aware (~12%) vs 1/20 blind
  (~5%). Directionally real — the aware side never scored below the blind
  side — but the absolute rate is low enough that a binomial at n=8 has a
  real chance of landing on zero for a correct build too (one aware resample
  did). This is reported as a finding, not a pass/fail gate: too weak a
  signal to gate on at any affordable sample size.
• Gate line (timeboxing: the Stage 1 gate line joins row and criterion with the same comma, so 'assumptions' reads as four separate needs #413): with every one of the 45 cells open, _gate_line
  produces a single 862-character line, one clause per row, each row named
  exactly once — under Slack's per-block text cap and with no cell dropped
  silently.
• A live Stage 1 turn, rendered by the real path: ran elicit() against
  the frozen fixture store (data/fixtures/stage1-20260905.db) for the
  working_tuesday fixture day (2026-09-08), a DAY_FRAME of 07:00–23:00,
  and one ELICITED_STATEMENT — "im going to my own office in 2 hours, so
  thats a 30 minute commute." — with now pinned to 2026-09-08 10:09
  Europe/Amsterdam, contender judges on the production timeboxing_judge
  agent type (the flash pin). Built the AwaitingUser the kernel's
  _stage1_outcome produces from the resulting gate and top probe, ran it
  through map_outcome and render_stage_card, and posted the actual
  blocks (not hand-written ones) to #ff-e2e, threaded under
  1788780130.551149. The posted card (ts 1788954054.922099) asked "Will
  you be working remotely before you head to the office at 12:09?" — a
  probe about the office trip, naming a clock time 2 hours after 10:09,
  exactly the "in 2 hours" statement resolved forward rather than the stale
  9:30 the live bug produced. The gate line read Still need: how the day is bounded (assumptions, unstated knowledge) · what is fixed (assumptions, alternatives, clarity, unstated knowledge) · … — one clause per row, as
  designed. Read back via conversations.replies: the message came back
  whole, with the "Decided" line correctly showing "you said: im going to my
  own office in 2 hours, so thats a 30 minute commute."
  Finding, not part of these three tickets: while resampling this turn,
  a probe whose judge attached closed-set options (e.g. suggested times for
  the end-of-day planning block) produced an actions block with multiple
  buttons sharing one static action_id
  (ff_timebox_blocker_option) — Slack's chat.postMessage rejected the
  whole message with invalid_blocks / action_id already exists. This
  reproduces regardless of timeboxing: the kernel asks the Priorities question before Stage 1 runs, so the ladder reads 1 → 2 → 1 and the shape of the day comes second #411/timeboxing: Stage 1 has no clock — 'in 2 hours' at 10:09 became 'by 9:30 AM' #412/timeboxing: the Stage 1 gate line joins row and criterion with the same comma, so 'assumptions' reads as four separate needs #413 (multi-option Stage 1 probes exist
  independently of this branch) and existing unit tests never catch it
  because they assert decoded button values, never post to the real API.
  Worth its own ticket.

Human checklist

• Start a session (no prior facts) and confirm the first card the bot
  sends is a Stage 1 probe, not the Priorities question.
• Reply with something like "in 2 hours" to a Stage 1 probe and confirm
  the next probe's stated time, if it names one, is after the current
  time rather than a stale earlier hour.
• Confirm a gate line with several rows open reads as one clause per
  row, criteria in parentheses — not one row name repeated per
  criterion.

Status

Rebased onto origin/feat/stage-card-whole — #397 is still OPEN, so
this PR's diff includes #397's commits until it merges into main. Not
blocked from opening as a draft; will need a rebase onto main once #397
lands, same as any stacked branch.

🤖 Generated with Claude Code