code-coordinator 0.5.145__tar.gz → 0.5.146__tar.gz

code_coordinator-0.5.146/.github/workflows/acceptance-web.yml

@@ -0,0 +1,126 @@
+name: Acceptance (web)
+
+# #2389: split out of test.yml's `acceptance` job — it was running on EVERY
+# push/PR to this repo regardless of what changed, unconditionally installing
+# Node + a real Chromium (`npx playwright install chromium --with-deps`)
+# before running the web-playwright (ms-51) acceptance route. A pure `tui/`
+# Rust change (already covered by cargo-test.yml's own tui-tuidriver route,
+# `--for-path tui/src/main.rs`, gated on its own `paths:` filter) had no
+# business paying that cost on every review round — this job is genuinely
+# webapp-specific (`--for-path coord/dashboard/webapp/src/App.tsx`), so it
+# gets the same `paths:`-gated treatment cargo-test.yml already established
+# for the Rust side. Not in the branch-protection required-checks list
+# (only test (3.12)/(3.13), thin-client-strict, e2e, no-gh-on-path are), so
+# this staying un-triggered on an unrelated PR is not a merge blocker.
+on:
+  push:
+    branches: [main]
+    paths:
+      - 'coord/dashboard/webapp/**'
+      - 'tests/acceptance/**'
+      - '.github/coord-ci-acceptance.yml'
+      # Self-trigger so edits to this workflow are exercised by their own PR.
+      - '.github/workflows/acceptance-web.yml'
+  pull_request:
+    branches: [main]
+    paths:
+      - 'coord/dashboard/webapp/**'
+      - 'tests/acceptance/**'
+      - '.github/coord-ci-acceptance.yml'
+      - '.github/workflows/acceptance-web.yml'
+  workflow_dispatch:
+
+jobs:
+  # #1950: the SEALED oracle-loop acceptance suite (tests/acceptance/ms-NN/,
+  # playwright.acceptance.config.ts) was run by no automatic gate anywhere —
+  # only `coord acceptance run/record`, invoked once while a milestone was
+  # being driven and never again. ms-51 went red at #1547 (the responsive
+  # shell removed the `<h1>coord</h1>` §1a pins) and stayed red through
+  # #1548/#1550/#1551/#1818 with nothing noticing. This job is the cheapest
+  # fix named in #1950: run the accumulated sealed suite whenever anything it
+  # covers changes, same shape as `e2e` (test.yml) (Python for the `coord`
+  # CLI on PATH — the acceptance config's `webServer` shells out to `coord
+  # web --fixture ... --dist dist`, see playwright.acceptance.config.ts's
+  # module docstring — plus Node + a real chromium).
+  #
+  # `npm run test:acceptance` with NO `{ms}` filter (unlike a single
+  # milestone's `-- ms-NN`) walks the whole `tests/acceptance/` tree — the
+  # accumulated suite across every closed milestone, i.e. exactly `coord
+  # acceptance run --all` (Gate C) would assemble, just invoked by CI's
+  # cadence instead of a milestone-close event.
+  #
+  # #2180: the `npm run test:acceptance` step now runs THROUGH `coord
+  # acceptance run --all --ci` (#2164) instead of calling Playwright
+  # directly. Same suite, same reporter under the hood — the wrapper's
+  # web-playwright driver shape appends `--reporter=json` and reads
+  # PLAYWRIGHT_JSON_OUTPUT_FILE itself (coord/acceptance_drivers.py) — but
+  # now a test-id listed in a manifest's `expected_red:` that fails does not
+  # redden this job (a sealed slice authored red before its fix exists,
+  # #2164's whole point), while one that unexpectedly PASSES does, loudly
+  # (#1965's vacuous-assertion case). Before this, ms-51 going red for any
+  # reason failed this job exactly the same as every other kind of red —
+  # there was no way to merge a red-by-design slice without `--force-merge`.
+  # .github/coord-ci-acceptance.yml's header explains why this points at a
+  # CI-local config instead of the fleet's real ~/.coord/coordinator.yml.
+  acceptance:
+    runs-on: ubuntu-latest
+    timeout-minutes: 20
+    defaults:
+      run:
+        working-directory: coord/dashboard/webapp
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+      - working-directory: .
+        run: pip install -e ".[dev]"
+      - uses: actions/setup-node@v4
+        with:
+          node-version: '20'
+      - run: npm ci
+      # #2393: same fix as test.yml's `e2e` job — see that job's comment.
+      - name: Cache Playwright browsers
+        id: playwright-cache
+        uses: actions/cache@v4
+        with:
+          path: ~/.cache/ms-playwright
+          key: ${{ runner.os }}-playwright-${{ hashFiles('coord/dashboard/webapp/package-lock.json') }}
+      - name: Install Playwright + OS deps (cache miss)
+        if: steps.playwright-cache.outputs.cache-hit != 'true'
+        # #2393 follow-up: the Chromium download from Playwright's CDN has hung
+        # (not merely failed) multiple times in one day, past what caching alone
+        # can paper over on a cache-miss run (a version bump, or the very first
+        # run ever). A short per-attempt timeout + retry turns "hang for the
+        # whole 20-minute job budget" into "fail fast and try again" — transient
+        # CDN/mirror flakiness usually clears within a couple of attempts.
+        run: |
+          for attempt in 1 2 3; do
+            echo "Playwright install attempt $attempt/3..."
+            if timeout 300 npx playwright install chromium --with-deps; then
+              echo "Playwright install succeeded on attempt $attempt"
+              exit 0
+            fi
+            echo "Playwright install attempt $attempt failed or timed out after 5m"
+            sleep 10
+          done
+          echo "Playwright install failed after 3 attempts"
+          exit 1
+      - name: Install Playwright OS deps only (cache hit)
+        if: steps.playwright-cache.outputs.cache-hit == 'true'
+        run: npx playwright install-deps chromium
+      - name: coord acceptance run --all --ci (web-playwright route, ms-51)
+        working-directory: .
+        run: >-
+          coord acceptance run --repo claude-coordinator --all --ci
+          --for-path coord/dashboard/webapp/src/App.tsx
+          --config .github/coord-ci-acceptance.yml --path .
+      - name: A red result here is attributable, not a mystery
+        if: failure()
+        run: |
+          echo "::error::The sealed acceptance suite (tests/acceptance/ms-NN/) failed."
+          echo "The failing spec's path and 'describe' title name the milestone (ms-NN) and the issue that slice was authored for (see tests/acceptance/ms-NN/manifest.yml)."
+          echo "Two causes, and only one fix is right:"
+          echo "  1. The app regressed against an still-accurate contract -> fix the app."
+          echo "  2. Legitimate app evolution made the contract stale (e.g. #1547 vs ms-51) -> amend the contract: coord acceptance mock <repo> <tracking_issue> --amend '<what changed and why>'"
+          echo "Do not silence this job (continue-on-error, skip, etc.) instead of doing one of the above -- that is exactly how ms-51 rotted unnoticed through #1547-#1818 (#1950)."

{code_coordinator-0.5.145 → code_coordinator-0.5.146}/.github/workflows/test.yml

@@ -252,6 +252,12 @@ jobs:
 
   e2e:
     runs-on: ubuntu-latest
+    # #2389: this is the SAME `npx playwright install chromium --with-deps`
+    # step that hung 34+ minutes on `acceptance` with no automatic recovery
+    # (a job with no `timeout-minutes` defaults to GitHub's 360-minute
+    # ceiling) — required, so it stays in this file, but gets the same
+    # bounded timeout rather than the implicit 6-hour one.
+    timeout-minutes: 20
     defaults:
       run:
         working-directory: coord/dashboard/webapp
@@ -272,90 +278,58 @@ jobs:
         with:
           node-version: '20'
       - run: npm ci
-      - run: npx playwright install chromium --with-deps
-      - run: npm run test:e2e
-
-  # #1950: the SEALED oracle-loop acceptance suite (tests/acceptance/ms-NN/,
-  # playwright.acceptance.config.ts) was run by no automatic gate anywhere —
-  # only `coord acceptance run/record`, invoked once while a milestone was
-  # being driven and never again. ms-51 went red at #1547 (the responsive
-  # shell removed the `<h1>coord</h1>` §1a pins) and stayed red through
-  # #1548/#1550/#1551/#1818 with nothing noticing. This job is the cheapest
-  # fix named in #1950: run the accumulated sealed suite on every push/PR,
-  # same shape as `e2e` above (Python for the `coord` CLI on PATH — the
-  # acceptance config's `webServer` shells out to `coord web --fixture
-  # ... --dist dist`, see playwright.acceptance.config.ts's module
-  # docstring — plus Node + a real chromium).
-  #
-  # `npm run test:acceptance` with NO `{ms}` filter (unlike a single
-  # milestone's `-- ms-NN`) walks the whole `tests/acceptance/` tree — the
-  # accumulated suite across every closed milestone, i.e. exactly `coord
-  # acceptance run --all` (Gate C) would assemble, just invoked by CI's
-  # cadence instead of a milestone-close event.
-  #
-  # Deliberately NOT added to scripts/coord-test-runner.sh (the Test gate's
-  # engine, coordinator.yml's `test_command`): that script also runs as a
-  # systemd *user* unit on the daemon host (#1814's PATH problem) and would
-  # need node/npm/chromium provisioned there for every future
-  # coord/dashboard/webapp/** branch's Test stage, not just this
-  # already-`browser`-capability-gated CI runner. CI is the tighter, cheaper
-  # loop for this suite; revisit the Test-gate route if CI's cadence turns
-  # out not to be enough (#1950's own "decide deliberately").
-  #
-  # ATTRIBUTABILITY (#1950 item 2): a red result here names a spec file
-  # under `tests/acceptance/ms-NN/` and, by that suite's own convention
-  # (tests/acceptance/ms-51/manifest.yml), a `describe` block carrying the
-  # milestone's tracking issue — e.g. `ms-51/home-active.spec.ts › ms-51
-  # Home Active tab (#1544) › ...`. Playwright's `list` reporter prints that
-  # full path/title on failure, so the ms-NN + issue are already in the
-  # job's own output; the failure-hint step below just spells out the two
-  # possible causes so whoever triages it doesn't have to rediscover them.
-  #
-  # #2180: the `npm run test:acceptance` step now runs THROUGH `coord
-  # acceptance run --all --ci` (#2164) instead of calling Playwright
-  # directly. Same suite, same reporter under the hood — the wrapper's
-  # web-playwright driver shape appends `--reporter=json` and reads
-  # PLAYWRIGHT_JSON_OUTPUT_FILE itself (coord/acceptance_drivers.py) — but
-  # now a test-id listed in a manifest's `expected_red:` that fails does not
-  # redden this job (a sealed slice authored red before its fix exists,
-  # #2164's whole point), while one that unexpectedly PASSES does, loudly
-  # (#1965's vacuous-assertion case). Before this, ms-51 going red for any
-  # reason failed this job exactly the same as every other kind of red —
-  # there was no way to merge a red-by-design slice without `--force-merge`.
-  # .github/coord-ci-acceptance.yml's header explains why this points at a
-  # CI-local config instead of the fleet's real ~/.coord/coordinator.yml.
-  acceptance:
-    runs-on: ubuntu-latest
-    defaults:
-      run:
-        working-directory: coord/dashboard/webapp
-    steps:
-      - uses: actions/checkout@v4
-      - uses: actions/setup-python@v5
+      # #2393: `npx playwright install chromium --with-deps` was the step
+      # that hung 30+ minutes with no automatic recovery (2026-08-18, this
+      # job and `acceptance`'s) — a fresh ~300MB Chromium download from
+      # Playwright's CDN on EVERY run, no caching, so every run was fully
+      # exposed to any network hiccup on that path. Caching the browser
+      # binary means most runs skip the download-heavy half entirely; only
+      # a cache miss (a `@playwright/test` version bump) pays the network
+      # cost. `install-deps` (apt-only, no download) still runs on a cache
+      # hit so a runner image without chromium's OS libs preinstalled
+      # doesn't break.
+      - name: Cache Playwright browsers
+        id: playwright-cache
+        uses: actions/cache@v4
         with:
-          python-version: "3.12"
-      - working-directory: .
-        run: pip install -e ".[dev]"
-      - uses: actions/setup-node@v4
-        with:
-          node-version: '20'
-      - run: npm ci
-      - run: npx playwright install chromium --with-deps
-      - name: coord acceptance run --all --ci (web-playwright route, ms-51)
-        working-directory: .
-        run: >-
-          coord acceptance run --repo claude-coordinator --all --ci
-          --for-path coord/dashboard/webapp/src/App.tsx
-          --config .github/coord-ci-acceptance.yml --path .
-      - name: A red result here is attributable, not a mystery
-        if: failure()
+          path: ~/.cache/ms-playwright
+          key: ${{ runner.os }}-playwright-${{ hashFiles('coord/dashboard/webapp/package-lock.json') }}
+      - name: Install Playwright + OS deps (cache miss)
+        if: steps.playwright-cache.outputs.cache-hit != 'true'
+        # #2393 follow-up: the Chromium download from Playwright's CDN has hung
+        # (not merely failed) multiple times in one day, past what caching alone
+        # can paper over on a cache-miss run (a version bump, or the very first
+        # run ever). A short per-attempt timeout + retry turns "hang for the
+        # whole 20-minute job budget" into "fail fast and try again" — transient
+        # CDN/mirror flakiness usually clears within a couple of attempts.
         run: |
-          echo "::error::The sealed acceptance suite (tests/acceptance/ms-NN/) failed."
-          echo "The failing spec's path and 'describe' title name the milestone (ms-NN) and the issue that slice was authored for (see tests/acceptance/ms-NN/manifest.yml)."
-          echo "Two causes, and only one fix is right:"
-          echo "  1. The app regressed against an still-accurate contract -> fix the app."
-          echo "  2. Legitimate app evolution made the contract stale (e.g. #1547 vs ms-51) -> amend the contract: coord acceptance mock <repo> <tracking_issue> --amend '<what changed and why>'"
-          echo "Do not silence this job (continue-on-error, skip, etc.) instead of doing one of the above -- that is exactly how ms-51 rotted unnoticed through #1547-#1818 (#1950)."
+          for attempt in 1 2 3; do
+            echo "Playwright install attempt $attempt/3..."
+            if timeout 300 npx playwright install chromium --with-deps; then
+              echo "Playwright install succeeded on attempt $attempt"
+              exit 0
+            fi
+            echo "Playwright install attempt $attempt failed or timed out after 5m"
+            sleep 10
+          done
+          echo "Playwright install failed after 3 attempts"
+          exit 1
+      - name: Install Playwright OS deps only (cache hit)
+        if: steps.playwright-cache.outputs.cache-hit == 'true'
+        run: npx playwright install-deps chromium
+      - run: npm run test:e2e
+
+  # #2389: the SEALED web-playwright acceptance suite (ms-51) moved to its
+  # own workflow, acceptance-web.yml, with a `paths:` filter scoped to
+  # `coord/dashboard/webapp/**`/`tests/acceptance/**` — mirroring how
+  # cargo-test.yml already gates the tui-tuidriver route on `tui/**`. This
+  # job was unconditionally installing Node + a real Chromium on every
+  # push/PR to this repo regardless of what changed, including pure-Rust
+  # `tui/` diffs that have nothing to do with the webapp (#1950's original
+  # "cheapest fix" traded diff-scoping for coverage-by-default; #2389
+  # restores diff-scoping now that the Rust side has its own equivalent
+  # gate). Not in this repo's required-checks list, so this move changes
+  # nothing about what blocks a merge.
 
   # #1550: `coord/dashboard/webapp/src/api/generated.ts` is produced by
   # `scripts/codegen.py` from the dashboard's OpenAPI spec

{code_coordinator-0.5.145 → code_coordinator-0.5.146}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: code-coordinator
-Version: 0.5.145
+Version: 0.5.146
 Summary: Coordinate Claude Code workers across multiple machines and repos.
 Author-email: John Donaghy <johnfdonaghy@gmail.com>
 License-Expression: LicenseRef-FSL-1.1-MIT

{code_coordinator-0.5.145 → code_coordinator-0.5.146}/code_coordinator.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: code-coordinator
-Version: 0.5.145
+Version: 0.5.146
 Summary: Coordinate Claude Code workers across multiple machines and repos.
 Author-email: John Donaghy <johnfdonaghy@gmail.com>
 License-Expression: LicenseRef-FSL-1.1-MIT

{code_coordinator-0.5.145 → code_coordinator-0.5.146}/code_coordinator.egg-info/SOURCES.txt

@@ -16,6 +16,7 @@ pyproject.toml
 .githooks/post-merge
 .github/coord-ci-acceptance.yml
 .github/ISSUE_TEMPLATE/bug_report.md
+.github/workflows/acceptance-web.yml
 .github/workflows/auto-release.yml
 .github/workflows/cargo-test.yml
 .github/workflows/publish.yml

{code_coordinator-0.5.145 → code_coordinator-0.5.146}/code_coordinator.egg-info/scm_file_list.json

@@ -9,6 +9,7 @@
     ".githooks/post-merge",
     ".github/ISSUE_TEMPLATE/bug_report.md",
     ".github/coord-ci-acceptance.yml",
+    ".github/workflows/acceptance-web.yml",
     ".github/workflows/auto-release.yml",
     ".github/workflows/cargo-test.yml",
     ".github/workflows/publish.yml",

{code_coordinator-0.5.145 → code_coordinator-0.5.146}/code_coordinator.egg-info/scm_version.json

@@ -1,7 +1,7 @@
 {
-  "tag": "0.5.145",
+  "tag": "0.5.146",
   "distance": 0,
-  "node": "g2c4967683cf16fc253e1af6018babca87d6eb43a",
+  "node": "geabea71cd1e375065c4baa33f806f7f473662f82",
   "dirty": false,
   "branch": "HEAD",
   "node_date": "2026-08-18"

{code_coordinator-0.5.145 → code_coordinator-0.5.146}/coord/test_author.py

@@ -60,6 +60,13 @@ TEST_AUTHOR_DENY_COMMANDS: list[str] = [
     "Bash(gh *)",
 ]
 
+# #1172 follow-up: how many versioned forks (`-v2`, `-v3`, ...) to probe
+# before giving up on finding a JIT slice's next re-authoring branch. Kept
+# small and finite — a real loop here means something is genuinely wrong
+# (not a legitimate string of re-authoring dispatches), and #1172's own
+# guard should surface that as a loud, readable error rather than spin.
+MAX_REAUTHOR_BRANCH_ATTEMPTS = 20
+
 TEST_AUTHOR_SYSTEM_PROMPT = """\
 You are an INDEPENDENT acceptance-test author dispatched by the coordinator.
 
@@ -538,20 +545,39 @@ def dispatch_test_author(
     branch = target_branch or f"issue-{tracking_issue}-{_slugify(assignment_title)}"
 
     if github_ops.pr_is_merged(repo_cfg.github, branch):
-        raise RuntimeError(
-            f"branch {branch!r} already has a merged PR — dispatching would "
-            "push new commits onto a dead branch with nothing left to open "
-            "a PR against them (#1172). "
-            + (
-                f"Issue #{issue_number}'s JIT slice already landed; if it "
-                "needs more tests, this needs a fresh branch (not "
-                "auto-forked yet) — do not retry as-is."
-                if issue_number is not None else
-                "The milestone's Gate-A suite PR already merged; if the "
-                "suite needs more work, open a fresh branch by hand — do "
-                "not retry as-is."
+        if issue_number is not None:
+            # #1172's own stated fix scope was "refuse (or fork a fresh
+            # branch)" — only the refuse half shipped. This is the other
+            # half: a genuine re-authoring dispatch (e.g. a Gate-A contract
+            # `--amend`, docs/ORACLE_LOOP.md's "amend the contract -> the
+            # test-author updates the affected slice") has nowhere else to
+            # push once the slice's own PR has merged, so fork onto the next
+            # unmerged versioned branch instead of failing outright.
+            base_branch = branch
+            for suffix in range(2, MAX_REAUTHOR_BRANCH_ATTEMPTS + 1):
+                candidate = f"{base_branch}-v{suffix}"
+                if not github_ops.pr_is_merged(repo_cfg.github, candidate):
+                    branch = candidate
+                    target_branch = candidate
+                    break
+            else:
+                raise RuntimeError(
+                    f"branch {base_branch!r} and its next "
+                    f"{MAX_REAUTHOR_BRANCH_ATTEMPTS - 1} versioned forks "
+                    "(-v2..-v"
+                    f"{MAX_REAUTHOR_BRANCH_ATTEMPTS}) all already have "
+                    "merged PRs (#1172) — that many re-authoring rounds on "
+                    "one slice is not a normal amend/re-sync cycle; open a "
+                    "fresh branch by hand and investigate."
+                )
+        else:
+            raise RuntimeError(
+                f"branch {branch!r} already has a merged PR — dispatching "
+                "would push new commits onto a dead branch with nothing "
+                "left to open a PR against them (#1172). The milestone's "
+                "Gate-A suite PR already merged; if the suite needs more "
+                "work, open a fresh branch by hand — do not retry as-is."
             )
-        )
 
     deny_commands = test_author_deny_commands(config, repo_name)
 

{code_coordinator-0.5.145 → code_coordinator-0.5.146}/tests/test_ci_acceptance_gate_1950.py

@@ -5,8 +5,9 @@ milestone is actively being driven, so a slice that goes red after its
 milestone closes (ms-51 at #1547, silently, through #1548/#1550/#1551/#1818)
 is never re-checked.
 
-This is the regression guard for the fix: `.github/workflows/test.yml` gets
-an `acceptance` job that runs the sealed suite on every push/PR (see that
+This is the regression guard for the fix: `.github/workflows/acceptance-
+web.yml` (originally an `acceptance` job inside `test.yml` — see #2389
+below) runs the sealed suite whenever anything it covers changes (see that
 job's own header comment for why CI rather than
 scripts/coord-test-runner.sh). Parsing the real workflow file — not a
 fixture — means this test fails the moment someone removes/disables/
@@ -24,6 +25,15 @@ wrapper invocation; a second test (`test_acceptance_job_uses_the_2164_ci_
 wrapper_not_the_raw_driver`) pins the specific `--all --ci` contract #2180's
 own acceptance criteria require ("through coord acceptance run --all --ci,
 not the raw driver command").
+
+#2389 update: the `acceptance` job moved out of `test.yml` into its own
+`acceptance-web.yml`, `paths:`-gated to `coord/dashboard/webapp/**` /
+`tests/acceptance/**` — it was unconditionally installing Node + a real
+Chromium on every push/PR regardless of what changed, including pure-Rust
+`tui/` diffs, mirroring the `paths:` gating `cargo-test.yml` already has for
+the tui-tuidriver route. `ACCEPTANCE_WORKFLOW_PATH` (not `WORKFLOW_PATH`) is
+what every acceptance-job test below now parses; `WORKFLOW_PATH` (`test.yml`)
+still backs the `test`/`e2e` job assertions, which did not move.
 """
 
 from __future__ import annotations
@@ -34,23 +44,24 @@ import yaml
 
 REPO_ROOT = Path(__file__).resolve().parents[1]
 WORKFLOW_PATH = REPO_ROOT / ".github" / "workflows" / "test.yml"
+ACCEPTANCE_WORKFLOW_PATH = REPO_ROOT / ".github" / "workflows" / "acceptance-web.yml"
 CARGO_WORKFLOW_PATH = REPO_ROOT / ".github" / "workflows" / "cargo-test.yml"
 CI_ACCEPTANCE_CONFIG_PATH = REPO_ROOT / ".github" / "coord-ci-acceptance.yml"
 
 
-def _load_workflow() -> dict:
+def _load_workflow(path: Path = WORKFLOW_PATH) -> dict:
     # YAML's `on:` key parses to the boolean `True` under the default
     # SafeLoader (YAML 1.1 treats the bare word `on` as a boolean) — harmless
     # here since this test never reads that key, but noted so a future
     # reader isn't surprised the loaded dict's key isn't the string "on".
-    return yaml.safe_load(WORKFLOW_PATH.read_text())
+    return yaml.safe_load(path.read_text())
 
 
-def _job_steps(workflow: dict, job_name: str) -> list[dict]:
+def _job_steps(workflow: dict, job_name: str, path: Path = WORKFLOW_PATH) -> list[dict]:
     jobs = workflow.get("jobs") or {}
     job = jobs.get(job_name)
     assert job is not None, (
-        f"expected a {job_name!r} job in {WORKFLOW_PATH} — jobs present: "
+        f"expected a {job_name!r} job in {path} — jobs present: "
         f"{sorted(jobs)}"
     )
     return job.get("steps") or []
@@ -65,8 +76,8 @@ def _step_runs(steps: list[dict], needle: str) -> dict | None:
 
 
 def test_acceptance_job_exists_and_runs_the_sealed_suite() -> None:
-    workflow = _load_workflow()
-    steps = _job_steps(workflow, "acceptance")
+    workflow = _load_workflow(ACCEPTANCE_WORKFLOW_PATH)
+    steps = _job_steps(workflow, "acceptance", ACCEPTANCE_WORKFLOW_PATH)
     test_step = _step_runs(steps, "coord acceptance run")
     assert test_step is not None, (
         "no step in the 'acceptance' job runs `coord acceptance run` — "
@@ -82,7 +93,7 @@ def test_acceptance_job_does_not_neuter_a_red_result() -> None:
     step (or the whole job) — the same "green screenshot" failure mode
     #1950 describes, just relocated from "never runs" to "runs but can't
     fail"."""
-    workflow = _load_workflow()
+    workflow = _load_workflow(ACCEPTANCE_WORKFLOW_PATH)
     jobs = workflow.get("jobs") or {}
     job = jobs.get("acceptance")
     assert job is not None
@@ -107,8 +118,8 @@ def test_acceptance_job_uses_the_2164_ci_wrapper_not_the_raw_driver() -> None:
     registry — a sealed slice authored red by design would fail this job
     exactly like any other red result, forcing `--force-merge` (#2164's
     whole reason for existing)."""
-    workflow = _load_workflow()
-    steps = _job_steps(workflow, "acceptance")
+    workflow = _load_workflow(ACCEPTANCE_WORKFLOW_PATH)
+    steps = _job_steps(workflow, "acceptance", ACCEPTANCE_WORKFLOW_PATH)
     test_step = _step_runs(steps, "coord acceptance run")
     assert test_step is not None
     run = test_step["run"]
@@ -213,8 +224,8 @@ def test_acceptance_job_installs_coord_cli_on_path() -> None:
     timeout instead of a real (or real-absence-of) assertion failure. Same
     requirement `e2e` already has for `live-update-fixture.spec.ts`
     (#1551)."""
-    workflow = _load_workflow()
-    steps = _job_steps(workflow, "acceptance")
+    workflow = _load_workflow(ACCEPTANCE_WORKFLOW_PATH)
+    steps = _job_steps(workflow, "acceptance", ACCEPTANCE_WORKFLOW_PATH)
     install_step = _step_runs(steps, "pip install")
     assert install_step is not None, (
         "the 'acceptance' job never installs the `coord` CLI (`pip install "
@@ -224,11 +235,121 @@ def test_acceptance_job_installs_coord_cli_on_path() -> None:
 
 
 def test_acceptance_job_installs_a_real_browser() -> None:
-    workflow = _load_workflow()
-    steps = _job_steps(workflow, "acceptance")
+    workflow = _load_workflow(ACCEPTANCE_WORKFLOW_PATH)
+    steps = _job_steps(workflow, "acceptance", ACCEPTANCE_WORKFLOW_PATH)
     browser_step = _step_runs(steps, "playwright install")
     assert browser_step is not None, (
         "the 'acceptance' job never runs `npx playwright install` — "
         "chromium (or whichever project(s) playwright.acceptance.config.ts "
         "declares) won't be present to actually run the suite"
     )
+
+
+def test_acceptance_workflow_is_paths_gated_to_webapp_changes() -> None:
+    """#2389: this job used to live in `test.yml` with no `paths:` filter at
+    all, so it ran unconditionally on every push/PR — including a pure-Rust
+    `tui/` diff with nothing to do with the webapp — always installing
+    Node + a real Chromium first. Mirrors `cargo-test.yml`'s own `paths:`
+    gate for the tui-tuidriver route."""
+    workflow = _load_workflow(ACCEPTANCE_WORKFLOW_PATH)
+    # YAML 1.1's bareword parsing turns the `on:` key into the boolean
+    # `True`, not the string "on" — see `_load_workflow`'s docstring.
+    on = workflow.get(True) or workflow.get("on") or {}
+    for trigger in ("push", "pull_request"):
+        paths = (on.get(trigger) or {}).get("paths") or []
+        assert "coord/dashboard/webapp/**" in paths, (
+            f"acceptance-web.yml's {trigger!r} trigger isn't paths-gated to "
+            f"coord/dashboard/webapp/**: {paths!r}"
+        )
+
+
+def test_acceptance_job_caches_playwright_browsers() -> None:
+    """#2393: neither `acceptance` nor `e2e` cached the Playwright browser
+    binary, so every single CI run re-downloaded ~300MB of Chromium from
+    Playwright's CDN from scratch — full exposure to any network hiccup on
+    that path, every time, not just occasionally. Confirmed live
+    2026-08-18: this hung 3 separate PRs' worth of runs in one afternoon.
+    A cache hit skips the download-heavy `install --with-deps` entirely in
+    favor of the apt-only `install-deps`."""
+    workflow = _load_workflow(ACCEPTANCE_WORKFLOW_PATH)
+    steps = _job_steps(workflow, "acceptance", ACCEPTANCE_WORKFLOW_PATH)
+    cache_step = next(
+        (s for s in steps if (s.get("uses") or "").startswith("actions/cache")),
+        None,
+    )
+    assert cache_step is not None, (
+        "the 'acceptance' job has no actions/cache step for Playwright's "
+        "browser binaries — every run re-downloads Chromium from scratch"
+    )
+    assert cache_step.get("with", {}).get("path") == "~/.cache/ms-playwright", (
+        f"cache step doesn't target Playwright's browser cache dir: "
+        f"{cache_step.get('with')!r}"
+    )
+    # The full (network-heavy) install must be conditioned on a cache MISS —
+    # otherwise the cache exists but is never actually consulted.
+    install_step = _step_runs(steps, "playwright install chromium --with-deps")
+    assert install_step is not None
+    assert "cache-hit != 'true'" in (install_step.get("if") or ""), (
+        "the full `playwright install --with-deps` step isn't gated on a "
+        f"cache miss, so the cache added above is never actually used: "
+        f"{install_step.get('if')!r}"
+    )
+
+
+def test_acceptance_job_retries_playwright_install_on_transient_hangs() -> None:
+    """#2393 follow-up, confirmed live 2026-08-18: caching alone doesn't help
+    a cache-miss run (the very first run on a branch, or any run after a
+    `@playwright/test` version bump) — and PR #2392's own first CI run hit
+    exactly that, hanging the full 20-minute job budget on one install
+    attempt with nothing to fall back on. The install must instead be
+    wrapped in a bounded per-attempt timeout with more than one attempt, so
+    a transient CDN/mirror hang fails fast and gets retried instead of
+    burning the whole job timeout on a single stuck attempt."""
+    workflow = _load_workflow(ACCEPTANCE_WORKFLOW_PATH)
+    steps = _job_steps(workflow, "acceptance", ACCEPTANCE_WORKFLOW_PATH)
+    install_step = _step_runs(steps, "playwright install chromium --with-deps")
+    assert install_step is not None
+    run = install_step["run"]
+    assert "timeout " in run, (
+        f"the cache-miss install step has no per-attempt timeout, so one "
+        f"stuck attempt can still consume the entire job budget: {run!r}"
+    )
+    assert "for attempt in" in run and "1 2 3" in run, (
+        f"the cache-miss install step doesn't actually loop over multiple "
+        f"attempts — a single attempt is exactly what hung 4 separate times "
+        f"on 2026-08-18: {run!r}"
+    )
+
+
+def test_e2e_job_retries_playwright_install_on_transient_hangs() -> None:
+    """Same fix, same reason, as the 'acceptance' job's version above — the
+    `e2e` job (test.yml) hit the identical hang on 2026-08-18 and got the
+    identical caching fix in #2393, so it needs the identical retry
+    follow-up too."""
+    workflow = _load_workflow()
+    steps = _job_steps(workflow, "e2e")
+    install_step = _step_runs(steps, "playwright install chromium --with-deps")
+    assert install_step is not None
+    run = install_step["run"]
+    assert "timeout " in run, (
+        f"e2e's cache-miss install step has no per-attempt timeout: {run!r}"
+    )
+    assert "for attempt in" in run and "1 2 3" in run, (
+        f"e2e's cache-miss install step doesn't loop over multiple attempts: {run!r}"
+    )
+
+
+def test_acceptance_job_has_a_bounded_timeout() -> None:
+    """#2389: a job with no `timeout-minutes` defaults to GitHub's 360-minute
+    ceiling — confirmed live 2026-08-18, `npx playwright install chromium`
+    hung 34+ minutes with no automatic recovery. A bounded timeout doesn't
+    prevent a hang, but it guarantees one can't silently occupy a runner (and
+    keep a PR looking "still checking") for hours."""
+    workflow = _load_workflow(ACCEPTANCE_WORKFLOW_PATH)
+    jobs = workflow.get("jobs") or {}
+    job = jobs.get("acceptance") or {}
+    timeout = job.get("timeout-minutes")
+    assert isinstance(timeout, int) and 0 < timeout <= 60, (
+        f"'acceptance' job's timeout-minutes is {timeout!r} — expected a "
+        "small positive bound, not GitHub's implicit 360-minute default"
+    )