Skip to content

ci(claude-review): replace plugin invocation with inline gh pr comment#71

Merged
adnaan merged 1 commit intomainfrom
fix-claude-review-workflow
Apr 15, 2026
Merged

ci(claude-review): replace plugin invocation with inline gh pr comment#71
adnaan merged 1 commit intomainfrom
fix-claude-review-workflow

Conversation

@adnaan
Copy link
Copy Markdown
Contributor

@adnaan adnaan commented Apr 15, 2026

Why

The Claude Code Review workflow on this repo had not been posting review comments on any PR — the workflow ran successfully (job returned exit 0) but no comment ever appeared. Investigation on #70 found two contributing problems.

Problem 1: Plugin-based prompt silently no-ops

The workflow used a plugin-based prompt invocation:

plugin_marketplaces: 'https://github.com/anthropics/claude-code.git'
plugins: 'code-review@claude-code-plugins'
prompt: '/code-review:code-review ${{ github.repository }}/pull/${{ github.event.pull_request.number }}'

The code-review@claude-code-plugins plugin appears to either suppress its review output entirely or use an internal posting path that does not work with the GitHub App token context the action provides. The job step returns success but no PR comment ever appears.

Compare to the livetemplate library workflow, which uses an inline prompt + explicit gh pr comment instruction. With that approach, Claude generates the review, then runs gh pr comment itself via its Bash tool (using the GH App installation token that the Claude Code Action injects into the runner's environment). This has been working continuously on the library repo and produced 4 review comments across the recent rounds of livetemplate/livetemplate#336.

Problem 2: Workflow validation against default branch blocks in-PR fixes

The Claude Code Action validates the PR's workflow file against the default branch's copy as a security mitigation. Any in-PR change to claude-code-review.yml itself fails with:

App token exchange failed: 401 Unauthorized — Workflow validation failed.
The workflow file must exist and have identical content to the version on
the repository's default branch.

This is a legitimate security boundary — without it, anyone with push access to the repo could open a one-line PR that modifies the workflow to inject gh repo set or arbitrary commands, then trigger the Claude Code Action to execute them with the GH App's elevated token.

The implication is that the fix has to land on main first via this dedicated PR before it takes effect on any other PR. We can't validate the change inside an existing feature PR.

Changes

  • Drop plugin_marketplaces, plugins, and the slash-command prompt
  • Add inline prompt with explicit "Use gh pr comment" instruction (verbatim from the library workflow, with REPO / PR NUMBER interpolation)
  • Add claude_args with allow-list for: gh issue view, gh search, gh issue list, gh pr comment, gh pr diff, gh pr view, gh pr list
  • Add allowed_bots: copilot-swe-agent (matches library config; lets Claude interact with Copilot comments on the same PR)

Trigger types are preserved unchanged (opened, synchronize, ready_for_review, reopened) — no functional reason to align them with the library's narrower set.

After this merges

  • Future PRs to livetemplate/examples will get real Claude review comments within ~5–8 minutes of each push
  • livetemplate/examples#70 (Session 3 patterns) will pick up the new workflow on its next push or workflow rerun, finally allowing Claude to review the Session 3 code (Copilot has already reviewed it; Claude has been silent because of this bug)

Test plan

This PR can't really be tested in isolation — the workflow is only triggered by pull_request events, and the only pull_request event for this PR is the PR-itself event. The first real validation will happen on the next PR to land after this merges.

🤖 Generated with Claude Code

The Claude Code Review workflow on this repo had not been posting
review comments on any PR — the workflow ran successfully (job
returned exit 0) but no comment ever appeared. Investigation on
#70 found two contributing problems:

1. The workflow used a plugin-based prompt invocation
   (`/code-review:code-review` via the `code-review@claude-code-plugins`
   plugin from the `anthropics/claude-code` marketplace) that silently
   fails to post review comments. The plugin appears to either suppress
   its output entirely or use an internal posting path that is not
   wired up correctly with the GitHub App token context the action
   provides.

2. The Claude Code Action validates the workflow file against the
   default branch's copy as a security mitigation, so any in-PR fix
   to this workflow itself fails with a 401 token-exchange error
   ("Workflow validation failed. The workflow file must exist and
   have identical content to the version on the repository's default
   branch"). This means the fix has to land on main first via a
   dedicated PR (this one) before it takes effect on any other PR.

The livetemplate library repo's claude-code-review.yml uses a
different approach that posts comments reliably:

- Inline prompt that explicitly tells Claude to call `gh pr comment`
  via its Bash tool
- `claude_args` allow-list whitelisting the relevant gh CLI
  subcommands (gh issue view, gh search, gh issue list,
  gh pr comment, gh pr diff, gh pr view, gh pr list)
- `allowed_bots: copilot-swe-agent` to interact with Copilot
  comments on the same PR

With that wiring, Claude generates the review, runs `gh pr comment`
itself (using the GH App installation token that the Claude Code
Action injects into the runner's environment), and the comment shows
up on the PR within 5–8 minutes. This has been working on the
library repo continuously, producing 4 review comments across the
recent rounds of the Session.TriggerAction PR
(livetemplate/livetemplate#336).

Changes in this PR:

- Drop `plugin_marketplaces`, `plugins`, and the slash-command prompt
- Add inline prompt with explicit "Use `gh pr comment`" instruction
- Add `claude_args` with the gh CLI tool allow-list
- Add `allowed_bots: copilot-swe-agent`

Trigger types are preserved unchanged (`opened, synchronize,
ready_for_review, reopened`) — no functional reason to align them
with the library's narrower trigger set, and shrinking them would
miss some PR state transitions.

After this lands on main, future PRs (and any rebased existing PR)
will pick up the new workflow on their next push and produce real
Claude review comments instead of the silent-no-op behaviour.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Copilot AI review requested due to automatic review settings April 15, 2026 01:59
Copy link
Copy Markdown

Copilot AI left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Pull request overview

Updates the repository’s Claude Code Review GitHub Actions workflow to reliably post PR review feedback by switching from a plugin-based invocation to an inline prompt that explicitly instructs Claude to post via gh pr comment.

Changes:

  • Remove plugin marketplace/plugin configuration and the slash-command prompt.
  • Add an inline review prompt that instructs Claude to post results using gh pr comment.
  • Add claude_args to restrict the allowed Bash tools to a specific gh allow-list, and set allowed_bots.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

@adnaan adnaan merged commit c6fa6c7 into main Apr 15, 2026
13 of 14 checks passed
@adnaan adnaan deleted the fix-claude-review-workflow branch April 15, 2026 02:10
adnaan added a commit that referenced this pull request Apr 15, 2026
Seven items from Claude Code Review on the rebased session-3 branch
(now that the workflow fix from #71 is on main, Claude is finally
posting reviews on examples PRs — this is the first one):

1. LazyLoad goroutines now use the documented cancellation pattern
   (handlers_loading.go)

   OnConnect and Reload previously ignored TriggerAction errors via
   `_ = session.TriggerAction(...)`. ProgressBarController.Start uses
   the documented `if err := session.TriggerAction(...); err != nil
   { return }` form to bail when the session disconnects mid-loop.
   Adopting the same form here makes the cancellation pattern
   consistent across the file and matches the proposal's Server Push
   pattern (#31). For the lazy-load case the goroutine is short
   enough that leaking it is harmless in practice, but consistency
   matters more.

2. Reconnect-during-loading note in OnConnect

   Claude observed that if the client reconnects within the 2s
   loading window, OnConnect fires again and a second goroutine
   spawns while the first is still asleep. With the err-returning
   pattern from #1, the first goroutine's TriggerAction returns an
   error (stale session) and exits cleanly — only the most recent
   goroutine completes. Added a comment explaining this implicit
   reliance on framework session-invalidation semantics.

3. Misleading comment on the OnConnect nil-session guard

   The previous comment claimed "the spinner-forever case is the
   documented JS-disabled fallback". This was wrong: JS-disabled
   clients never reach OnConnect at all (no WebSocket connection
   means OnConnect is never called). The actual JS-disabled
   fallback is created by Mount() returning Loading=true on the
   initial HTTP GET. The nil branch in OnConnect is purely a
   defensive guard against framework regressions — fixed the
   comment to say so.

4. TestLazyLoading/Reload_Refetches_Fresh_Content trivially-true
   inequality (patterns_test.go)

   The previous assertion `if firstContent == secondContent` was
   trivially false because the two strings have different prefixes
   ("Content loaded lazily at …" vs "Content reloaded at …") — they
   could never be equal regardless of timing. Replaced with positive
   assertions that firstContent contains the initial-load prefix and
   does NOT contain the reload prefix, and vice versa for
   secondContent. This actually proves both strings were generated
   by their respective controller paths (initial OnConnect vs
   Reload action), which was the original intent.

5. async-operations.tmpl: <mark> → <del> for error block
   (templates/loading/async-operations.tmpl, patterns_test.go)

   CLAUDE.md convention is `<del style="display:block;text-
   decoration:none">` for block-level error messages, while `<mark>`
   is reserved for highlighted/badge text. Switched the error detail
   block to <del> for consistency with the rest of the patterns
   example. Updated TestAsyncOperations selectors at three sites
   (Initial_Load presence check, Fetch_Transitions wait condition,
   outcome detection switch) to match.

6. Login regression test: narrow `ins` selector to id
   (login/login_test.go, login/templates/auth.html)

   The previous test selector `WaitForText("ins", ...)` matched any
   <ins> on the page. If the login template ever gains another
   styled <ins> (e.g., a generic success flash), the test could
   match the wrong element and pass spuriously. Added an
   `id="server-welcome-message"` to the auth.html template's
   ServerMessage <ins> and switched the test to use the explicit
   id selector. The id is stable and unambiguous.

7. Reload implicit-guard comment

   Bundled into #1's comment block. ProgressBarController.Start has
   an explicit `if state.Running { return }` guard against double-
   click stacking. LazyLoadController.Reload doesn't need one because
   the template hides the Reload button while Loading=true. Added a
   short comment explaining the asymmetry so future readers don't
   wonder why the patterns differ.

All Session 3 + login tests still pass against the published
livetemplate v0.8.18.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
adnaan added a commit that referenced this pull request Apr 15, 2026
…ss (#70)

* feat: patterns example Session 3 (patterns #14-16) — loading & progress

Implements the three Loading & Progress patterns from the patterns
example proposal (livetemplate#333):

- Pattern #14: Lazy Loading
- Pattern #15: Progress Bar
- Pattern #16: Async Operations

All three rely on `session.TriggerAction()` from a background goroutine
to push state updates to the client without polling. None of them
worked against the shipping livetemplate v0.8.17 because `ctx.Session()`
returned nil — the Session interface was declared but never wired into
any production code path. The library gap is fixed in
livetemplate#336 (merged as 4883481).

## What's new

**state_loading.go** — three state structs (LazyLoadState,
ProgressBarState, AsyncOpsState) following the Title/Category-first
convention from Sessions 1 and 2.

**handlers_loading.go** — three controllers + handler factories:
- `LazyLoadController.OnConnect` spawns a 2s goroutine that calls
  TriggerAction("dataLoaded", ...) once. Reload action re-triggers
  the same path. Mount() guards on `ctx.Action() == ""` so the Reload
  POST doesn't reset state.
- `ProgressBarController.Start` spawns a goroutine that ticks
  TriggerAction("updateProgress", {progress: int}) every 500ms for
  10 iterations. UpdateProgress reads the int via ctx.GetInt and
  emits a `success` flash on completion. The `Running` flag prevents
  double-click stacking within a session.
- `AsyncOpsController.Fetch` spawns a goroutine that sleeps 2s and
  calls TriggerAction("fetchResult", ...) with either success or
  error data (~33% simulated failure rate via math/rand). FetchResult
  routes to "success" / "error" via ctx.GetBool and emits the
  matching flash.

**templates/loading/{lazy-loading,progress-bar,async-operations}.tmpl**
— Pico CSS markup using `<progress>`, `<blockquote>`, `<mark>`, and
the FlashTag helper. The async-operations template uses {{with}} on
.Result and .Error to drop the redundant {{if eq .Status}}
boilerplate around the FlashTag calls — FlashTag is self-guarding
when the flash key isn't set, and {{with}} on the mutually-exclusive
Result/Error strings handles the success/error rendering branch
without referencing .Status.

**main.go** — three new mux.Handle registrations under /patterns/loading/.

**data.go** — flips Implemented:true on the three Loading & Progress
entries in allPatterns(). The index template iterates allPatterns()
data-driven, so no index.tmpl edits are required.

**patterns_test.go** — three new test functions:
- TestLazyLoading: Initial spinner, data arrives via push, Reload
  produces fresh content with a different timestamp.
- TestProgressBar: Initial state, Start runs to completion (ticks
  visible, success flash, "Run Again" button), Run Again restarts.
- TestAsyncOperations: Initial state, Fetch transitions through
  loading → success/error (random outcome accepted), matching
  flash element verified.

All tests use condition-based waits (e2etest.WaitForText, WaitFor
with real JS predicates) and the shared setupTest fixture. None
use chromedp.Sleep.

**login/login_test.go** — adds Server_Welcome_Message_via_WebSocket_Push
subtest to TestLoginE2E. Before the library fix, the login example's
sendWelcomeMessage goroutine silently no-op'd because ctx.Session()
returned nil; the existing "Successful Login" test passed only
because it asserted on the static "Welcome, testuser!" template
literal, not the server-pushed timestamped message. The new subtest
waits for "pushed from the server" text in the <ins> element with a
5s timeout — without the library fix the goroutine never fires and
this assertion times out.

## Dependencies

Requires livetemplate >= v0.8.18 (the next release after the merged
fix in livetemplate#336). go.mod will be bumped in a follow-up
commit on this branch once the library release lands. Until then,
the branch builds against the local main checkout via go.work for
manual testing.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* chore(deps): bump livetemplate to v0.8.18 for Session 3 patterns

v0.8.18 contains the Session.TriggerAction fix (livetemplate#336)
required by Session 3 patterns #14-16 (Lazy Loading, Progress Bar,
Async Operations). Previously the library declared the Session
interface but never wired ctx.WithSession() into any production code
path, so ctx.Session() returned nil and goroutine-based server
push silently no-op'd. v0.8.18 wires Session at all five NewContext
call sites (WS lifecycle, WS action, HTTP lifecycle, HTTP action,
PubSub server-action dispatch) plus the previously-missing
handleDispatchedAction and upload completion paths.

This bump also activates the regression test for the login example's
sendWelcomeMessage server push (added in 4f45ebe), which was
previously broken in the same way.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: address PR #70 review comments

Six items from Copilot review on the Session 3 patterns PR:

1. LazyLoadController.Reload — check session before mutating state
   (handlers_loading.go)

   Copilot: Reload set state.Loading=true and state.Data="" first,
   then early-returned if ctx.Session() was nil — leaving the page
   in a permanent loading spinner with no recovery path. Reordered
   so the session check happens first; if nil, return without
   touching state. With livetemplate v0.8.18+ the nil branch is
   unreachable (every action context now has WithSession wired),
   but the safer ordering means a future framework regression
   surfaces as "Reload does nothing" rather than "spinner stuck
   forever".

2. LazyLoadController.OnConnect — clarify nil-session intent
   (handlers_loading.go)

   Copilot raised the same "stuck loading forever" concern for
   OnConnect, but the existing structure was already correct
   (session check before goroutine spawn). Updated the doc comment
   to explicitly call out (a) why the defensive check stays even
   though v0.8.18+ guarantees non-nil session, and (b) that the
   "spinner-forever-with-JS-disabled" case is the documented
   fallback behaviour for the Lazy Loading pattern, not a bug.

3. ProgressBarController.Start — check session before Running=true
   (handlers_loading.go)

   Copilot: Start set Running=true before checking ctx.Session().
   If session were nil the goroutine wouldn't spawn, but Running
   would stay true forever and the Running guard at the top of
   Start would block all subsequent button clicks. Reordered to
   check session first.

4. AsyncOpsController.Fetch — check session before Status="loading"
   (handlers_loading.go)

   Same pattern as #3. The button is template-disabled when
   Status=="loading", so leaving Status pinned to "loading" with
   no goroutine to clear it would freeze the UI. Reordered.

5. math/rand "non-deterministic by design" comment — clarify
   (handlers_loading.go)

   Copilot incorrectly claimed math/rand top-level functions are
   deterministic unless seeded. They were until Go 1.19, but Go
   1.20 changed math/rand to auto-seed at program startup from a
   system source. Updated the comment to explicitly cite the Go
   1.20 behaviour change so the reasoning is clear and the comment
   doesn't read as wishful thinking.

6. TestAsyncOperations flash assertion strength
   (patterns_test.go)

   Copilot: the test only checked for presence of the
   output[data-flash="success"|"error"] element. An empty
   placeholder element with no text would satisfy the selector
   and silently mask a regression where SetFlash wasn't called.
   Strengthened the assertion to read element.textContent and
   verify it contains the expected literal flash text ("Fetch
   complete" or "Fetch failed"), matching what the controller's
   FetchResult method passes to ctx.SetFlash.

Items explicitly NOT changed:

- Copilot's suggestion to add synchronous data-load fallbacks for
  non-WebSocket clients in Lazy Loading. The proposal explicitly
  documents the JS-disabled limitation as expected behaviour for
  pattern #14 — adding a synchronous fallback would defeat the
  pattern's purpose of demonstrating server push, and the spinner-
  forever case is preferable to silently rendering empty content.

- Copilot's suggestion to add "live session unavailable" error
  flashes for Progress Bar / Async Ops. With v0.8.18+ the case is
  unreachable, so the speculative recovery UI would just be dead
  code that confuses future readers.

All Session 3 tests still pass against the published v0.8.18.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: address PR #70 round-2 Claude review comments

Seven items from Claude Code Review on the rebased session-3 branch
(now that the workflow fix from #71 is on main, Claude is finally
posting reviews on examples PRs — this is the first one):

1. LazyLoad goroutines now use the documented cancellation pattern
   (handlers_loading.go)

   OnConnect and Reload previously ignored TriggerAction errors via
   `_ = session.TriggerAction(...)`. ProgressBarController.Start uses
   the documented `if err := session.TriggerAction(...); err != nil
   { return }` form to bail when the session disconnects mid-loop.
   Adopting the same form here makes the cancellation pattern
   consistent across the file and matches the proposal's Server Push
   pattern (#31). For the lazy-load case the goroutine is short
   enough that leaking it is harmless in practice, but consistency
   matters more.

2. Reconnect-during-loading note in OnConnect

   Claude observed that if the client reconnects within the 2s
   loading window, OnConnect fires again and a second goroutine
   spawns while the first is still asleep. With the err-returning
   pattern from #1, the first goroutine's TriggerAction returns an
   error (stale session) and exits cleanly — only the most recent
   goroutine completes. Added a comment explaining this implicit
   reliance on framework session-invalidation semantics.

3. Misleading comment on the OnConnect nil-session guard

   The previous comment claimed "the spinner-forever case is the
   documented JS-disabled fallback". This was wrong: JS-disabled
   clients never reach OnConnect at all (no WebSocket connection
   means OnConnect is never called). The actual JS-disabled
   fallback is created by Mount() returning Loading=true on the
   initial HTTP GET. The nil branch in OnConnect is purely a
   defensive guard against framework regressions — fixed the
   comment to say so.

4. TestLazyLoading/Reload_Refetches_Fresh_Content trivially-true
   inequality (patterns_test.go)

   The previous assertion `if firstContent == secondContent` was
   trivially false because the two strings have different prefixes
   ("Content loaded lazily at …" vs "Content reloaded at …") — they
   could never be equal regardless of timing. Replaced with positive
   assertions that firstContent contains the initial-load prefix and
   does NOT contain the reload prefix, and vice versa for
   secondContent. This actually proves both strings were generated
   by their respective controller paths (initial OnConnect vs
   Reload action), which was the original intent.

5. async-operations.tmpl: <mark> → <del> for error block
   (templates/loading/async-operations.tmpl, patterns_test.go)

   CLAUDE.md convention is `<del style="display:block;text-
   decoration:none">` for block-level error messages, while `<mark>`
   is reserved for highlighted/badge text. Switched the error detail
   block to <del> for consistency with the rest of the patterns
   example. Updated TestAsyncOperations selectors at three sites
   (Initial_Load presence check, Fetch_Transitions wait condition,
   outcome detection switch) to match.

6. Login regression test: narrow `ins` selector to id
   (login/login_test.go, login/templates/auth.html)

   The previous test selector `WaitForText("ins", ...)` matched any
   <ins> on the page. If the login template ever gains another
   styled <ins> (e.g., a generic success flash), the test could
   match the wrong element and pass spuriously. Added an
   `id="server-welcome-message"` to the auth.html template's
   ServerMessage <ins> and switched the test to use the explicit
   id selector. The id is stable and unambiguous.

7. Reload implicit-guard comment

   Bundled into #1's comment block. ProgressBarController.Start has
   an explicit `if state.Running { return }` guard against double-
   click stacking. LazyLoadController.Reload doesn't need one because
   the template hides the Reload button while Loading=true. Added a
   short comment explaining the asymmetry so future readers don't
   wonder why the patterns differ.

All Session 3 + login tests still pass against the published
livetemplate v0.8.18.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* revert: keep <mark> for async-operations error detail

The previous commit (e32e703) switched the error block in
async-operations.tmpl from <mark> to <del style="display:block;
text-decoration:none">{{.}}</del> based on Claude's review nit
flagging the deviation from CLAUDE.md convention. Reverting that
specific change after follow-up discussion:

- <mark> is shorter and reads cleaner in the template
- <mark> is the more semantic choice for "highlighted text" which
  matches the error-detail role: the FlashTag above is the primary
  error indicator, and <mark> draws attention to the specific error
  string as a secondary highlight
- Pico's default yellow background for <mark> works fine as a
  visual cue without needing block-level error styling
- If a more error-specific look is wanted later, livetemplate.css
  can override mark[role="error"] or similar — keeping the markup
  tight today does not preclude a stronger style tomorrow

The CLAUDE.md guidance ("<del> for error messages") is preserved
unchanged for now since this is one example's deliberate deviation,
not a project-wide convention shift. If the team decides to allow
<mark> for inline error highlights more broadly, that's a separate
docs change.

Test selectors at three sites in TestAsyncOperations are reverted
back to `mark` to match the template.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: address PR #70 round-3 Claude review comments

Six items from Claude's two reviews on the recent commits:

1. AsyncOpsController.Fetch missing concurrent-call guard
   (handlers_loading.go)

   Real bug: ProgressBarController.Start guards against re-entry via
   `if state.Running { return }`, but AsyncOpsController.Fetch had no
   equivalent. The button is template-disabled during loading, but a
   direct WebSocket message bypassing the rendered UI could spawn two
   parallel goroutines, each calling TriggerAction("fetchResult") and
   producing duplicate state transitions and SetFlash calls. Added
   `if state.Status == "loading" { return state, nil }` at the top
   of Fetch, mirroring ProgressBarController's guard.

2. AsyncOpsController goroutine err-return for consistency
   (handlers_loading.go)

   The goroutine used `_ = session.TriggerAction(...)` on both branches
   while LazyLoadController and ProgressBarController use the
   `if err := session.TriggerAction(...); err != nil { return }` form.
   This is harmless for AsyncOps because the goroutine is single-shot
   (one TriggerAction call then exit), but inconsistent with the
   established pattern. Switched both branches to the err-return form
   so readers learning from this example see the idiomatic shape
   everywhere. Added a comment explaining that the err-return is
   defense-in-depth for a single-shot goroutine.

3. TestProgressBar/Start_Runs_To_Completion intermediate-tick check
   (patterns_test.go)

   The previous WaitFor only asserted `progress.value > 0`, which would
   also be satisfied by a regression where the goroutine instantly
   jumps to 100 without intermediate ticks. Strengthened to
   `progress.value > 0 && progress.value < 100`, matching the
   Run_Again_Restarts_Timer subtest. This proves the goroutine is
   actually ticking in 10% increments, not jumping straight to done.

4. progress-bar.tmpl: 100% complete label in Done state
   (templates/loading/progress-bar.tmpl)

   During Running, the template shows the progress bar plus a
   `<p><small>{{.Progress}}% complete</small></p>` label. When the job
   finishes and transitions to Done, the label disappeared, creating
   a small visual discontinuity (label visible at 10–90%, then vanishes
   at 100). Added `<p><small>100% complete</small></p>` to the Done
   branch for visual continuity.

5. aria-live on the error mark for screen reader announcement
   (templates/loading/async-operations.tmpl)

   Wrapped the error <mark> in `<p aria-live="assertive">` so screen
   readers announce the error when the page transitions from loading
   to error state. The success branch's <blockquote> doesn't need
   aria-live because the FlashTag above already has appropriate ARIA
   semantics (output[role="status"] from the FlashTag helper). The
   error path's FlashTag uses role="alert" but the additional error
   detail in the mark needed its own announcement signal.

6. Concurrent Fetch regression test
   (patterns_test.go)

   Added TestAsyncOperations/Concurrent_Fetch_Reaches_Single_Result.
   Sends two `fetch` actions in immediate sequence via direct
   WebSocket message (bypassing the disabled button), waits for the
   cycle to complete, and asserts exactly one result element
   (blockquote OR mark) is present. This is a smoke test for the
   user-visible invariant — concurrent Fetches don't break the page —
   rather than a direct test of the guard logic, because detecting
   "the second call was rejected" from the rendered HTML is hard
   when the state machine is idempotent in its final state.

Items NOT addressed:

- Test description string "mark element" → already accurate after
  the previous revert (template uses <mark>, description says "mark").
- Reload button disable while Loading → the framework auto-adds
  aria-busy on form buttons during the WS round-trip, AND the Reload
  button only renders in the {{else}} branch when Loading=false, so
  a separate disabled attribute would be redundant.
- TestLazyLoading no JS-disabled path test → not actionable, matches
  how all other patterns are structured. The JS-disabled spinner-
  forever case is documented in handlers_loading.go's OnConnect
  comment block.
- async-operations <mark> → <del> swap → user explicitly settled this
  in the previous revert (keeping <mark> as the more semantic choice
  for "highlighted error detail" with potential CSS override later
  in livetemplate.css).
- Reconnect goroutine comment polish → the existing comment block
  already explains the mechanics adequately.

All Session 3 + login tests still pass against published v0.8.18.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: address PR #70 round-4 Claude review comments

Two items from Claude's latest review:

1. Reload double-goroutine comment was inaccurate (handlers_loading.go)

   Claude correctly caught that my comment claimed
   "TriggerAction returns an error when DataLoaded clears Loading
   and a second Reload starts; the err check below handles it"
   — but TriggerAction errors only on session disconnect, not on
   state changes. The actual behaviour is: both goroutines run to
   completion, both call TriggerAction successfully, and the second
   simply overwrites state.Data with a newer timestamp. This is
   harmless for the demo but the explanatory comment was wrong.
   Rewrote the comment to describe the actual mechanism (concurrent
   completion, harmless overwrite) and noted the migration path
   if stricter single-flight semantics are wanted later.

2. Document the deliberate <mark> deviation from CLAUDE.md
   (templates/loading/async-operations.tmpl)

   Claude has flagged the <mark> usage as a CLAUDE.md convention
   violation on three consecutive reviews. The user explicitly
   settled this in an earlier round: <mark> is the more semantic
   choice for "highlighted error detail" (the FlashTag above with
   role="alert" is the primary error indicator; this is a secondary
   highlight of the error string), and Pico's default styling works
   without needing block-level error markup. To stop the review-
   comment loop without changing the underlying decision, added an
   inline {{/* ... */}} template comment directly above the line
   explaining (a) it's a deliberate deviation, (b) the semantic
   reasoning, and (c) the override path via livetemplate.css if a
   stronger error look becomes desired. Future Claude reviews
   should see the comment and recognise it as a documented
   deviation rather than an oversight.

Items NOT changed:

- map[string]interface{} → map[string]any cosmetic suggestion: the
  livetemplate library defines the TriggerAction signature using
  interface{}, so callers naturally use the same syntax for type
  uniformity. Switching only the example would create a stylistic
  inconsistency with the framework's own API surface.

All Session 3 tests still pass against published v0.8.18.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: address PR #70 round-5 Claude review

- Document why ProgressBar/AsyncOps controllers don't need OnConnect:
  state has no `lvt:"persist"` tags, so reconnects always produce fresh
  state via cloneStateTyped() rather than restoring stuck Running=true /
  Status="loading". The "stuck after reconnect" scenario the bot flagged
  cannot occur in ephemeral mode. LazyLoadController needs OnConnect
  because the spinner→data swap *is* the pattern; ProgressBar/AsyncOps
  patterns are about the goroutine push itself, and a mid-run disconnect
  cleanly ends the demo.
- map[string]interface{} → map[string]any throughout handlers_loading.go
  for consistency with linter expectations and bot review.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: address PR #70 round-6 Claude review

- TestProgressBar.Run_Again_Restarts_Timer now asserts the success flash
  re-appears on the second completion. Catches a regression where the
  controller forgot to call SetFlash on the re-completion path.
- Rewrote LazyLoadController.OnConnect reconnect-during-loading comment
  to accurately describe both possible outcomes (gap → goroutine errors;
  reconnected → both goroutines dispatch to new connection, second
  overwrites Data harmlessly). The previous version claimed framework
  session-invalidation semantics that don't actually exist — sessions
  are looked up by groupID, and groupID is stable across reconnects.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: address PR #70 round-6 Claude review

- Add Loading re-entrancy guard to LazyLoadController.Reload, symmetric
  with ProgressBarController.Start and AsyncOpsController.Fetch. The
  previous comment correctly noted that the template hides the button
  while loading, but a direct WS message could bypass that — the asymmetry
  with the other two controllers was a trap for readers pattern-matching
  from this file.
- TestAsyncOperations.Fetch_Transitions_Through_Loading_To_Result now
  WaitFor's the flash element before reading its text, and properly
  reports the chromedp.Run error instead of swallowing it. Also switches
  the JS string to fmt.Sprintf for readability.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: address PR #70 round-7 Claude review

- Add <noscript> note to lazy-loading.tmpl spinner branch. JS-disabled
  clients never establish a WebSocket, so the spinner never resolves;
  this fallback makes the constraint visible to readers in the rendered
  UI rather than only in Go comments. The pattern itself is still a
  patterns-demo trade-off (a production lazy-load would render content
  server-side first), so the noscript text says "this pattern requires
  JavaScript" rather than pretending it has a real fallback.
- Widen TestProgressBar.Run_Again_Restarts_Timer intermediate-tick
  timeout from 3s → 5s. The goroutine ticks every 500ms over 5s; on a
  loaded CI runner the first tick can be delayed past 3s and produce a
  spurious failure even when the goroutine is working correctly. The
  outer 'Run Again button appears' assertion already enforces full
  completion within 10s, so widening the inner check is safe.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs: sanction <mark> for secondary inline error details in CLAUDE.md

The async-operations.tmpl error string was using <mark> as a "deliberate
deviation" from CLAUDE.md, with an inline comment explaining the rationale.
This commit promotes that exception into the convention itself: <mark> is
now the documented choice for secondary inline error details (a specific
error string highlighted alongside a primary FlashTag alert that already
carries role="alert"). <del> remains the primary error alert.

Also updates the inline template comment to reference the new convention
instead of explaining a deviation that no longer exists.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs: comment FlashTag placement scope in progress-bar.tmpl

Document why FlashTag is rendered only inside the .Done branch (the
only render path that emits a "success" flash). The idle and Running
branches don't need a FlashTag slot because no controller code path
emits a flash while either is the active render state.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: address PR #70 round-8 Claude review

- Sharpen progress-bar.tmpl FlashTag comment to explain the failure
  mode if the FlashTag is moved outside the .Done branch (the flash
  would be consumed during a Running/idle render before Done is
  reached, and the user would never see it). The previous comment
  described the design but didn't make the load-bearing constraint
  obvious to a future maintainer.
- Bump TestProgressBar.Start_Runs_To_Completion intermediate-tick
  timeout from 3s → 5s for consistency with Run_Again_Restarts_Timer
  and to give loaded CI runners headroom before the goroutine
  completes the full 5s run.
- Shorten CLAUDE.md <mark> guideline from 60 words to ~30 while
  preserving the rule, the FlashTag pairing, the aria-live note,
  and the "primary vs secondary" rule of thumb.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: address PR #70 round-9 Claude review

Use {{.Progress}} instead of literal "100" in the .Done branch of
progress-bar.tmpl. The controller currently sets Progress=100 whenever
Done=true, so the rendered output is identical, but binding to .Progress
makes the template survive any future controller change to the terminal
value (e.g., capping at 99 if a follow-up adds a "still finalising"
state) without silently showing a stale 100.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* refactor: dedupe progress markup in progress-bar.tmpl

Factor the <progress> + percent label out of both the .Running and
.Done branches into a single {{if or .Running .Done}} block, and
restructure the form/button branches as {{if .Done}}{{else if not
.Running}}{{end}}. The rendered output is identical in all three
states (idle, running, done), but the template no longer duplicates
the <progress value="{{.Progress}}" max="100"> markup across two
branches — a future styling change only needs to touch one place.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: address PR #70 round-10 Claude review

- Move <noscript> in lazy-loading.tmpl outside the {{if .Loading}} block.
  The fallback is now unconditional at the top of the article so future
  template restructuring can't accidentally drop it. JS-disabled clients
  always see Loading=true so the rendered output is unchanged in practice;
  this is purely a robustness improvement for future maintenance.
- TestAsyncOperations.Fetch_Transitions: now checks the chromedp.Run error
  when reading `outcome`, and uses a single `flashSelector` variable across
  the WaitFor and Evaluate so the two reads can't drift. The race window
  the bot flagged was theoretical (subtests are sequential) but the
  defensive change is trivially safer.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants