Skip to content

fix(desktop): restore channel header hidden behind shared backdrop - #2139

Merged
thomaspblock merged 1 commit into
mainfrom
fix-channel-name
Jul 19, 2026
Merged

fix(desktop): restore channel header hidden behind shared backdrop#2139
thomaspblock merged 1 commit into
mainfrom
fix-channel-name

Conversation

@morgmart

@morgmart morgmart commented Jul 19, 2026

Copy link
Copy Markdown
Contributor

Problem

#2108 (focused thread mode) added z-0 to the channel content <section>. That created a stacking context which flattened the header's z-30 beneath the sibling z-30 shared header backdrop — so the backdrop painted over the entire channel header: name, badges, and action buttons.

Fix

Remove the z-0. It was unnecessary for the focus drawer's layering — the drawer's z-41 already paints above the section's contents (max z-40) without forcing a stacking context. Both focus-mode e2e tests still pass.

Regression guard

Added a paint-order assertion to thread-focus-mode.spec.ts that compares the header's and backdrop's stacking-context branches directly. Neither toBeVisible() nor elementFromPoint can catch this occlusion (the backdrop is pointer-events-none, so hit-testing skips it). Verified the guard fails with the z-0 re-added and passes with the fix.

Testing

  • pnpm exec playwright test thread-focus-mode — 2 passed with fix; guard correctly fails when z-0 is reintroduced
  • Screenshot capture of /channels/general confirms the full header (title, description, members, actions) is visible again

@morgmart
fix(desktop): restore channel header hidden behind shared backdrop
21991e3 
PR #2108 added z-0 to the channel content section, creating a stacking
context that flattened the header's z-30 beneath the sibling z-30
shared header backdrop — the backdrop painted over the channel name,
badges, and action buttons. The z-0 was unnecessary: the focus drawer's
z-41 already wins over the section's contents without it.

Adds a paint-order regression guard to the focus-mode e2e spec that
compares the header's and backdrop's stacking branches directly, since
neither toBeVisible() nor elementFromPoint (backdrop is
pointer-events-none) can detect this occlusion.
@morgmart
morgmart requested a review from a team as a code owner July 19, 2026 19:55
@thomaspblock
thomaspblock merged commit 916c67d into main Jul 19, 2026
25 checks passed
@thomaspblock
thomaspblock deleted the fix-channel-name branch July 19, 2026 20:24
@wesbillman wesbillman mentioned this pull request Jul 20, 2026
Merged
BradGroux pushed a commit to BradGroux/buzz that referenced this pull request Aug 23, 2026
@morgmart
fix(desktop): restore channel header hidden behind shared backdrop (b…
9e7e31b 
…lock#2139)
tucktuck101 pushed a commit to tucktuck101/buzz that referenced this pull request Sep 17, 2026
@serina-mcfall
docs(plans): record the re-dispatch observable decision (block#2139)
404e095 
Serina's call, 2026-09-15: each pass stamps a fresh per-pass run
identifier alongside its verdicts, and a verdict carrying a prior pass's
identifier is a replay rather than a re-run.

Chosen over comparing the two passes' raw stdout, which fails against
the exact thing block#2139 asks it to detect - a deterministic verifier
returning byte-identical output twice is the healthy case, so matching
bytes cannot distinguish a real re-run from a cache hit.

The limitation is recorded rather than hidden: a fresh identifier proves
a dispatch happened, not that it was independent of the first pass's
reasoning. STEP 2 now requires the skill to say so.

Signed-off-by: test <test@example.com>
tucktuck101 pushed a commit to tucktuck101/buzz that referenced this pull request Sep 17, 2026
@serina-mcfall
feat(the-professor): STEP 2 — prove the second pass re-dispatched (bl…
04c31fd 
…ock#2139)

Adds section 4a to verify-claims/SKILL.md. The run-twice rule was
decided 2026-09-04 but had no observable, so a pass that re-checked
every claim and a pass that replayed the first pass's verdicts were
indistinguishable from outside - and caching those verdicts is the
obvious optimisation.

The observable, per Serina's 2026-09-15 decision: a per-pass run
identifier, generated fresh before the pass dispatches anything and
stamped on every verdict it produces. A verdict carrying the previous
pass's identifier was replayed, not re-checked.

Comparing the two passes' raw stdout was rejected when the decision was
taken: a deterministic verifier returning byte-identical output twice is
the healthy case, so matching bytes cannot distinguish a re-run from a
cache hit.

The identifier alone is not enough, so it is bound to evidence - one
recorded invocation per CITED CLAIM per pass, not one per pass. A pass
over six claims that records one invocation has not re-checked the other
five whatever its verdicts are stamped with, which is the evasion the
round-2 review found the earlier wording allowed. An inspector counts
cited claims against invocation records and requires equality.

Invocation records identify their claim by location and citation rather
than reproducing its text, matching the discipline check-page's findings
already follow, so this new evidence channel does not become a way for
draft content to outlive the draft.

The limitation is recorded rather than hidden: a fresh identifier and a
matching record prove a dispatch HAPPENED, not that it was INDEPENDENT.
An agent could re-dispatch carrying the first pass's reasoning and still
produce a well-shaped record. Isolation stays a property of how step 2
is followed.

Verification: python3 launchpad/agents/the-professor/tools/check_professor.py
--offline -> ALL NETWORK-FREE CHECKS PASSED (the invocation CI uses).
All seven local section references (2, 2a, 2b, 2c, 3, 4, 4a) resolve to
real headings.
Not verified: procedure text, which the harness asserts nothing about.
STEP 3 exercises the contract; the replay-negative case lands in STEP 8.

Signed-off-by: test <test@example.com>
tucktuck101 pushed a commit to tucktuck101/buzz that referenced this pull request Sep 17, 2026
@serina-mcfall
feat(the-professor): STEP 4 — wire the third gate and the final pass …
d7d1cff 
…into draft-page (block#2138, block#2139)

draft-page mentioned verify-claims exactly once before this, inside
baseline mode's numbered step 5. The ordinary drafting path went
check-page -> screen-sensitive -> write, so the gate Serina declared
mandatory and unskippable on 2026-09-04 was wired into one mode and
absent from the path almost every page takes.

Restructured the tail of the procedure. Step 8 no longer authorises the
write - it authorises step 9. Two new steps and a renumber:

  9.  the claim gate. Third, always after the other two, never beside
      them. Any verdict other than SUPPORTED on any claim stops the
      write entirely, as does a dispatch that produced no verdict at
      all. No partial-write path and no write-the-supported-sections
      path. Findings use screen-sensitive's shape so review has one
      place to look. A draft with zero behaviour claims still runs the
      gate and records that it found none - it never skips, and never
      invents a claim to have something to verify.
  10. the final independent pass, and ALL THREE gates run in it, not
      just verify-claims. A fix made in response to step 9 is an edit
      after check-page and screen-sensitive already ran, so re-running
      only the claim gate leaves those two holding verdicts about a page
      that no longer exists - a repair could introduce a secret or break
      a citation after screening and still reach the target path. Any
      edit after this sequence invalidates it and requires the whole
      sequence again from check-page.
  11. the existing post-write ledger and index hand-off, unchanged.

Baseline mode now points at steps 9 and 10 by number rather than naming
the gate before either had a step to point at, and takes no carve-out
from the edit-invalidates-the-sequence rule.

Forward references corrected in the same pass: step 6 said the ledger
append happens 'in step 8' (now 11) and that 'both gates below' need the
complete draft (now three).

Verification: python3 launchpad/agents/the-professor/tools/check_professor.py
--offline -> ALL NETWORK-FREE CHECKS PASSED. verify-claims now appears 8
times, across the step 9 and 10 procedure text, baseline mode and the
checklist, against 1 before - and that 1 was not in the ordinary path.
Not verified: procedure text, which the harness asserts nothing about.

Signed-off-by: test <test@example.com>
tucktuck101 pushed a commit to tucktuck101/buzz that referenced this pull request Sep 17, 2026
@serina-mcfall
feat(the-professor): STEP 5 — wire the third gate and the final pass …
a33adde 
…into update-page (block#2138, block#2139)

update-page contained zero occurrences of verify-claims before this. It
went check-page -> screen-sensitive -> overwrite the real file.

Same shape as STEP 4's draft-page change, restructured 6-9: step 6 is
the sensitivity gate and no longer authorises the overwrite, step 7 is
the claim gate, step 8 is the final pass over all three gates, step 9 is
the existing post-write ledger append.

ONE DECISION THIS STEP HAD TO MAKE, flagged in the file as needing
confirmation. Neither block#2138 nor the plan says whether the claim gate
inherits step 5's deliberate whole-page scope. check-page runs page-wide
because a section-scoped edit can break a page-level rule. Dispatching
every claim on the page would multiply every update's cost by the page's
total claim count, doubled again by run-twice, for claims nothing in the
edit could have invalidated.

Split, on the grounds that it is not a weakening:

  - the UNSOURCED check runs over the WHOLE page. It needs no dispatch,
    so it is free, and a section-scoped edit genuinely can strip the only
    citation backing a claim elsewhere - the same cross-section failure
    step 5 already scopes page-wide for.
  - the per-claim dispatch covers only the rewritten section. A claim in
    an untouched section is checked against an untouched citation;
    rewriting a different section cannot change whether that citation
    supports it. A citation that has since gone stale is scan-repo's job
    to detect, not this gate's to re-litigate on every unrelated edit.

The free half runs page-wide and the paid half runs where the content
changed. The alternative is strictly more expensive with no failure mode
identified that it catches and this does not.

Also fixed: a cross-reference to draft-page section 8's retry-safety
note, which STEP 4 moved to section 11.

Verification: python3 launchpad/agents/the-professor/tools/check_professor.py
--offline -> ALL NETWORK-FREE CHECKS PASSED. Headings run 0, 1, 1a, 2-9
contiguously.
Not verified: procedure text, which the harness asserts nothing about.

Signed-off-by: test <test@example.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@thomaspblock thomaspblock thomaspblock approved these changes

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@morgmart @thomaspblock