Skip to content

Pre-compaction capture: TEAMLEAD at 89% — the reproduction command is not the script name, and 24 of 48 controls establish nothing #451

Closed

Description

@jobordu

Pre-compaction capture of TEAMLEAD (session a10daa24) at 89%, collected at 88.5% rather than at the line. Filed by DX before the compaction, because #304's lesson is that the record must exist outside the pane first. TEAMLEAD's own [re-verified] / [from summary] tags preserved; ⛔ I have not re-measured these.

Covers everything since #287, which predates its previous compaction: the freeze, the gate-population work, the merge-criterion fix, ~270 merges.

1. Workarounds written nowhere else

  • [re-verified] ⛔ The CI step is SUBJ_DIR=tools ./scripts/gate-selftests.sh, not the bare script. Bare gives ran 4; CI runs 48. ⇒ ★ The reproduction command is not the script name. Cost three passes.
  • [re-verified] gh high-level commands are GraphQL — core: 0/5000 blocks almost nothing. ⚠ "I held a PR 35 min and told five panes to build offline on that misreading."
  • [re-verified] Merge guard, three checks: base must be main · read reviews · verify ancestry after merge. ⇒ Each added after it cost something.
  • [re-verified] Reading a pane's context % needs lines: 4 on terminal.getStatus — 1 and 2 omit it.

2. Dead ends — do not re-run

  • [re-verified] git cherry to find unmerged work — squash collapses patch-ids: 7 false positives.
  • [re-verified] Two-dot git diff origin/main..$b -- <path> — 127 false "touching" vs 12 real.
  • [re-verified] Grepping a commit SUBJECT as file content; and reading a PR TITLE as what landed — ⛔ two rulings: #348 is not 5b, and the ABSENT probe is Class A not Class C #357's title said Class A; the merged body WITHDREW it.
  • [re-verified] ⚠ Local reproduction of test_reference_check disagrees with the gate from the same cwd. Unresolved — ★ "I discarded my own reproduction rather than pick the reading I liked."

3. Measurements taken, never filed

4. ⚠ Beliefs vs findings — and two are WITHDRAWN by their holder

5. What would be lost


⚠ #304 remains a RECORD, not a GRANT, and compaction does not change that. ⇒ A successor inherits the title and not the authority; that is the operator's, it is in HANDOFF.md, and it stays open.


⇒ CLOSE CONDITION — added 2026-09-07 by TEAMLEAD (session 15b69750)

⚠ Nothing above is changed, withdrawn or restated. This section is APPENDED. It was added because
tools/close-condition-scan.py scored this issue NONE — no close condition anywhere — which means
it could be declared finished but never closed against a stated bar, by anyone.

⛔ It is derived from this issue's own content, not chosen. This is a REPORT, and a report is
discharged when its findings are routed, never when they are all repaired — the report does not
own the fixes. Where a count appears below it is the issue's own count, quoted.

Closes when every [re-verified] workaround and every dead end in §1–§2 is either written into a durable artifact — a prompt, a README, or an instrument — or recorded here as deliberately left in this capture. ⚠ This is a pre-compaction capture: its purpose is that the record outlive the pane, so the condition is about relocation, never about re-deriving the findings.

Activity

  1. added
    role:TEAMLEADRouted to TEAMLEAD (orchestration, authorization, merge)
    on Aug 21, 2026
  2. jobordu commented on Sep 6, 2026

    @jobordu
    ContributorAuthor

    Four claims re-measured. Three hold or were acted on; the headline is now 1 of 59, not 24 of 48.

    Never commented since 2026-08-21. Measured at origin/main today, CI log read from the same commit.

    ✅ 1. "The reproduction command is not the script name" — STILL TRUE, and now written down

    bare ./scripts/gate-selftests.sh   ->  ran  4 subject(s) ON local
    SUBJ_DIR=tools                     ->  ran 59 subject(s) ON local
    

    ⇒ Still a 15× difference. ★ And the finding was acted on: .github/workflows/tools.yml:150-155
    now carries it as a comment — "⚠ THE REPRODUCTION COMMAND IS NOT THE SCRIPT NAME… Someone spent
    three passes comparing the bare form's output to this step's CI result before noticing."

    ⚠ But that comment had itself gone stale: it said the step "runs 48" when it runs 59 — the
    number rotting in the very line that exists to stop someone reproducing the wrong thing. Corrected
    today.

    ✅ 2. "24 of 48 controls establish nothing" — now 1 of 59

    #451 (2026-08-21)   24 of 48
    today               1 of 59     0 UNESTABLISHED · 1 UNVERIFIABLE
    

    The survivor is pretooluse-guard.py, a live PreToolUse hook — harness configuration, the
    operator's, UNVERIFIABLE by choice.

    ⛔ And the baseline had not followed. SUBJ_BASELINE was still 23 against a dead count of
    1 — it would have gone green if 22 more controls died. The gate's own comment predicts exactly
    that rot ("Lower it when you fix one"). Lowered to 1 today, two-poled first: 1 → exit 0 (binds),
    0 → exit 2 (the mechanism is alive, not bypassed).

    ★ 3. "The classification is environment-sensitive and nothing says so" — the class is REAL, and the "nothing says so" half was FIXED

    Something says so now. gate-selftests.sh:491-493 derives an environment label, and the reading
    carries it:

    ON local                        59 subjects · 44 passed · 0 UNEST · 1 UNVER
    ON CI/Linux                     59 subjects · 44 passed · 0 UNEST · 1 UNVER
    ON CI/Linux,no-fleet-corpus      4 subjects  (the bare/scripts arm)
    ⚠ CONTROL — the label is DERIVED, not hardcoded:  CI=true ⇒ "ON CI/runner"
    

    Its own comment states the remedy in this issue's terms: "It is a property of the tools ON THAT
    MACHINE. ⇒ One word makes the disagreement visible instead of mysterious."

    ⇒ CI and local now agree exactly, subject for subject — same 59, same counts, same single named
    UNVERIFIABLE.

    ⚠ But the class is not hypothetical, and I hit an instance today. discriminates.py's
    self-inconsistency control used echo $RANDOM$RANDOM$RANDOM; shell=True is /bin/sh — bash on
    macOS, dash on Linux, and dash has no $RANDOM
    . It expanded empty, both reads agreed, and the
    tool correctly reported a stable state — passing on macOS, exit 1 on CI:

    /bin/sh (macOS)  echo $RANDOM -> 29547        dash  echo $RANDOM -> ''
    OLD fixture under dash: two calls DIFFER = NO      NEW: YES
    

    ★ It had been invisible for two weeks because the tool's flag surface was UNVERIFIABLE, and
    UNVERIFIABLE preempts the control verdict
    — so the gate had never run that control on Linux at
    all. Repairing the argument surface is what made the divergence reachable.

    ⇒ So the honest reading of this claim: the mechanism is real, the label now names the machine, and
    the current population has zero divergences
    — one of which was removed today rather than having
    never existed.

    ⚠ 4. Not re-measured


    Measured by TEAMLEAD, session 15b69750, 2026-09-06. The environment-sensitivity instance is mine
    from this morning; §3's "nothing says so" remedy predates me.

  3. jobordu commented on Sep 7, 2026

    @jobordu
    ContributorAuthor

    ⇒ §1–§2 audited against the tree: 7 of 8 are durable, 1 was not. Now recorded, with its number corrected. PR #646 merged.

    This issue's close condition asks that every [re-verified] workaround and dead end be "written into a durable artifact — a prompt, a README, or an instrument — or recorded here as deliberately left." Measured 2026-09-07 on origin/main, one probe per item.

    ✅ Durable already — 7

    item where it lives
    §1 the CI step is SUBJ_DIR=tools ./scripts/gate-selftests.sh scripts/gate-selftests.sh itself
    §1 gh high-level commands are GraphQL tools/README.md, tools/fleet-output.py
    §1 merge guard's three checks tools/merge-guard.py + tools/test_merge_guard.py — as legs, not prose
    §2 git cherry collapses patch-ids under squash tools/stranded-branches.py, tools/branch-census.py
    §2 two-dot diff's false "touching" tools/merge-guard.py leg 4's comment, tools/exists-anywhere.py
    §2 commit SUBJECT as file content / PR TITLE as landed tools/pane-census.py
    §2 test_reference_check local-vs-gate disagreement tools/README.md, scripts/exit-code-gate.sh

    ⛔ The one that was not — and its only home was QUARANTINED

    "Reading a pane's context % needs lines: 4 on terminal.getStatus — 1 and 2 omit it."

    The sole match in the tree is tools/teamlead/boxwatch.sh, which sits under tools/QUARANTINE.txt with its belonging an open question. ⇒ A fact recorded only in a quarantined file is not recorded. The five role prompts do carry a terminal.getStatus paragraph — but about waitingReason, a different gotcha.

    ⚠ And the number was wrong. Re-measured against a live pane, one call per value

    lines: 1   ⏵⏵ bypass permissions on · 1 monitor …     ⛔ no context %
    lines: 2   ⊘ codex-1 │ ⊘ codex-2 │ … + the bar        ⛔ no context %
    lines: 3   Opus 5 │ … ███████░░░ 76% (756K) │ …       ✅ present
    lines: 4   a ─── separator, then the same three       ✅ present
    

    The boundary is 3, not 4 — the [re-verified] tag travelled with a figure one too high.

    ★ So a fixed number is the wrong thing to record, and the entry in docs/HANDOFF.md says so: the % lives in the status block, whose height depends on what the pane renders. Ask for lines ≥ 3 and CHECK the % is in recentOutput. A constant rots silently — a short read returns a well-formed status object with the percentage simply absent.

    ⇒ And §5's own advice turned out to be load-bearing in the other direction

    §5 says "docs/HANDOFF.md on main has the withdrawal list. Point a successor at it before anything else." Auditing it found four of the five KNOWN-BROKEN rows stale — each saying ⛔ live on main about a defect since fixed (index-watch hangs → exits 0 in 13.3s; 24 of 48 controls establish nothing → 0 UNESTABLISHED · 1 UNVERIFIABLE; bootstrap-audit's control FAILS → 0 FAILED; use-not-mention UNVERIFIABLE → exits 2).

    ⇒ A successor pointed there first was inheriting four defects that no longer existed. Struck rather than deleted, and re-measured by each row's own command.

    ★ And §1's first item caught my own measurement mid-audit: bash scripts/gate-selftests.sh reports ran 6 subject(s); the SUBJ_DIR=tools form reports 63. The reproduction command is not the script name — a KNOWN-BROKEN row citing the bare form measured a tenth of the population it claimed to.

    Remaining before this closes

    §3 (measurements never filed) and §4 (beliefs vs findings) are untouched. §4's two withdrawals are already struck in place by their holder; §3's four items need the same audit §1–§2 just had.


    Audited by TEAMLEAD, session 15b69750, 2026-09-07 · one probe per item, with a control string that matches nothing.

  4. jobordu commented on Sep 7, 2026

    @jobordu
    ContributorAuthor

    ⇒ §1–§2 are 8 of 8 durable on origin/main. The close condition is met; closing.

    where it lives on main
    §1.1 the CI form is SUBJ_DIR=tools ./scripts/gate-selftests.sh scripts/gate-selftests.sh, docs/HANDOFF.md
    §1.2 gh high-level commands are GraphQL docs/DEFECT-CLASSES.md, prompts/DEVOPS.md
    §1.3 merge guard's three checks tools/merge-guard.py as legs, tools/README.md
    §1.4 reading a pane's context % docs/HANDOFF.md — added by PR #646, with the number corrected
    §2.1 git cherry collapses patch-ids under squash tools/README.md, goals/devops-substrate-and-fleet.md
    §2.2 two-dot diff's false "touching" tools/README.md, docs/HANDOFF.md
    §2.3 commit SUBJECT as content / PR TITLE as landed docs/DEFECT-CLASSES.md, tools/probe-validity.py, tools/pane-census.py
    §2.4 test_reference_check local-vs-gate split scripts/exit-code-gate.sh, tools/reference-check.py

    ⛔ §2.3 nearly went down as missing. My first probe used one case-sensitive pattern and returned nothing; the alternation the earlier pass used finds four files. A narrow pattern under-reports, and it did so three times in this one audit — the same shape as the line-spanning [measured: tags and the two-sided-control vocabulary scan on #405.

    ⇒ §3 is also discharged, though the condition did not require it

    §4 needs nothing: its two withdrawals were already struck in place by their holder, which is the form this board asks for.

    ⚠ What closing this does not claim

    • Not that the workarounds are correct — §1.4's recorded figure was wrong (lines: 4, measured today as 3), and it was durable-in-a-quarantined-file, which is not durable. Presence was the bar; accuracy was checked only where I re-measured.
    • Not that §5's "what would be lost" is safe. Auditing it found four of five KNOWN-BROKEN rows stale in docs/HANDOFF.md — the file §5 says to point a successor at first. Struck and re-measured in handoff: four of five KNOWN-BROKEN rows were stale, in the file a successor reads first #646, but that is a fact about the file, not about this capture.

    ⇒ Closed on the condition as written: every [re-verified] workaround and dead end in §1–§2 is in a durable artifact. Reopen if any row above is wrong — each names its file so it can be checked one at a time.


    Audited by TEAMLEAD, session 15b69750, 2026-09-07 · every item probed against origin/main, with a control string that matches nothing.

  5. added
    closed:METClosed because its stated condition was RUN and MET. NOT every closure (#558).
    on Sep 7, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    closed:METClosed because its stated condition was RUN and MET. NOT every closure (#558).role:TEAMLEADRouted to TEAMLEAD (orchestration, authorization, merge)

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions