T075, 10.2 and 10.2a — an installed entry point serves all 42 bundle files; intent-feed.js is not owed (plan 034)

[brettheap]

Lane: openxfactory-4 (openXfactory-4-openDox_extraction)

Arc: neutral-product-standalone-operability

Plan 034 (specs/034-opendox-standalone-operation/tasks.md, read at openxFactory main 2140f5a7), phase-3 slice P3-E, the door, its first task:

• T075 (#1144's 10.2 and 10.2a): "The entry point serves the 42-file bundle, reachable in a browser from an openDox-only install. intent-feed.js is not owed (10.2a is a declaration, recorded in the PR)."
• Falsifier: F10.1's fetch, as T007 batch H amends it (openxFactory#1206 → f99a2097): a .[local] install and --local, which start the bundled server. Widened here to every one of the 42 files.
• After: T056, T038, T063, T069, T072, T007 (batch H). T063 has landed (#1218 → a883bbf6); T072 landed (T072, 13.1: the bundled PostgreSQL server, the local install's own child (plan 034) #69 → 5e7ab00). This PR now targets main. It was stacked on T072, 13.1: the bundled PostgreSQL server, the local install's own child (plan 034) #69's branch (build/034-p3i-t072-bundled-postgres) and branched at T072, 13.1: the bundled PostgreSQL server, the local install's own child (plan 034) #69's merge-from-main head 19e32f0c, which carries main 047bb4fa, and T072, 13.1: the bundled PostgreSQL server, the local install's own child (plan 034) #69's later heads are merged in, never rebased: 6eb0bbdb as a7cf2665, fedfa75d as cac12d91, 21bde1af (the adversarial-review round, which pins pixeltable-pgserver>=0.6.0,<0.7 in the same pyproject.toml) as a3cf5e9a, 28e195b9 (fix round 13) as 0457b383, and 085ba0b0 (fix rounds 14 and 15, plus main 2fc714d2 after T070, 13.4-13.6: OPENDOX_INSTALL_MODE and generate-and-open --local (plan 034) #67 T070, T078, 16.1: the OpenAI-compatible dialect joins DIALECTS (plan 034) #61 T078 and T079, 16.2: a model field the provider receives (plan 034) #62 T079 landed) as d7511388, 3ebccb3c (fix round 16) as 1e642b21, and 52a2b8dc (main 9a490405, after T081, 16.4: "no model configured" is a state, shown before any turn (plan 034) #74 T081 landed, which edits the bundle's views/doxbench-chat.js) as 0b0d762c, and c04690f8 (fix rounds 17 and 18, T072, 13.1: the bundled PostgreSQL server, the local install's own child (plan 034) #69's final pre-squash head) as bbd91a3f. Those merges were clean plain merges. The two pyproject.toml hunks are apart, so that merge needed no resolution and keeps both. After T072, 13.1: the bundled PostgreSQL server, the local install's own child (plan 034) #69's squash 5e7ab003 landed on main (over 1130e996, T080, 16.3: the credential stays a reference, and a raw key is refused (plan 034) #63 T080), main was merged with git merge-tree --merge-base c04690f8, recording both parents (308523d9), never rebased. c04690f8..5e7ab003 changes exactly T080, 16.3: the credential stays a reference, and a raw key is refused (plan 034) #63's five files, so against main this diff is T075's four files alone.

Ruled:

• R1Q22 (a), 5817152735;
• batch H's F10.1 amendment: R1Q15 (b), R1Q16 (iii), 5850003126;
• the drafting word, 5960162524 ("Draft all of them now (Recommended)").

Claimed on openxFactory#656 in 5960231968. Authored ahead as a DRAFT under 5960162524. T063 and T072 have since landed. The holder posts READY.

What the measurement found, at #69's head 19e32f0c

The suite's own runs use an editable install, which puts src/ on the path, so each of them serves the tree's 42 files. pip install ".[local]" installs a wheel, and the wheel is what was measured, from a directory outside the checkout:

• The wheel carried 41 of the tree's 42 web files. vendor/.gitkeep was missing. setuptools expands package data with the standard library's glob, and web/** never matches a name that starts with a dot. The census (tests/fixtures/web_boundary_census.yaml) counts that file, and pyproject.toml's own note says the glob "ships exactly what that census counts". Slice S5 had measured the same omission and let it stand as a directory keeper (tests/test_gate_loop_contributed.py). 10.2 counts it among the 42, so it ships.
• The installed opendox generate-and-open --local --no-open served the other 41, each with the tree's bytes, and answered GET /vendor/.gitkeep with 404. F10.1's own fetch of / passed. The content types were text/javascript, text/css and text/html. The bundled server started, migrated ['0001', '0002'], and stopped on SIGTERM with exit 0.
• Nothing else was missing. The static route (SimpleHTTPRequestHandler, over opendox/web/ beside the installed opendox/cli.py) already serves whatever the wheel carries. So serve.py and cli.py are NOT edited, and this PR joins neither single-writer order.

What changes

• pyproject.toml: opendox = ["web/**", "web/**/.*"], and the note above the line now counts 42 files across two subdirectories. A pattern whose last part starts with a dot is the one that matches such a name, at any depth under web/. Measured: 42 files from pip's isolated build, 42 from --no-build-isolation over the test extra's setuptools 84.0.0, and 42 from a wheel built out of the sdist (whose tarball carries src/opendox/web/vendor/.gitkeep). This is T072's file in phase 3's single-writer table, and this PR is stacked after T072.
• tests_runtime/test_served_bundle.py (new, 3 cases, the in-suite falsifier):
  1. the wheel built from this checkout carries every file under src/opendox/web/, and nothing else there;
  2. F10.1's fetch, widened. The wheel is installed under a prefix of its own (T072's recipe), and its opendox console script runs generate-and-open --local --no-open --port 0 from a directory that is not a checkout. The preconditions are checked first: none of the four siblings is importable, opendox --help exits 0, and the WEB_DIR it serves is the wheel's copy. Then / answers with <html, and each of the 42 files answers 200 with the tree's bytes and a type a browser accepts (a JavaScript MIME type for .js, text/css, text/html). The module graph a browser walks from / closes inside the bundle. The walk follows the page's src/href, then each module's static imports, then every dynamic import() target the bundle carries (views/projection-index.js is reached that way). Each module's imports are read with its comments removed.
  3. The walk's import reader, over the tree, with no server: its edges equal an independent line-by-line reading of every relative specifier, which is 80 edges. The edge a commented multi-line clause declares (views/doxbench-editor.js → views/doxbench-state.js) is named, and an in-memory mutant of only that specifier is caught. The process stops on SIGTERM with exit 0, and its bundled server is gone.
  • The child's environment drops every OPENDOX_*, PG*, GIT_* and XF_* name (the suite exports XF_GATE_PRINCIPALS). The fixture repository gets an explicit identity. The child is stopped by SIGTERM, which the local lifecycle reads as its interrupt, so a runner that ignores SIGINT (a nohup job) changes nothing. Under CI, a missing bundled server FAILS the module rather than skipping it, as test_bundled_postgres.py does. The short /tmp state directory's removal is retried, bounded, and asserted (71d24af6). A red mutant run had once left a postgres/data/pg_wal/ behind, because a backend was still writing after its postmaster went.
• tests/test_gate_loop_contributed.py and tests/test_validator_input_set.py: each pinned the bundle's package-data line verbatim, and each now pins the new line. The recursive-glob guarantee and the setuptools floor are unchanged. S5's measurement in the first docstring (40 entries, 36 under views/) is kept as S5's. T075's re-measurement sits beside it: 41 entries and 37 views from web/** alone, and 42 with both patterns.

10.2a: intent-feed.js is not owed (the declaration)

views/intent-feed.js stays at openxFactory under RULED OQ-F (not_moved / stays_openxfactory_adapter), and it is not owed to openDox. openDox's replacement is views/intent-binding.js. That module reaches ./intent-feed.js only by a dynamic import(), so an absent file is an absent binding, not a broken bundle. It is not counted as a missing file. The 42 do not include it, and the census has no row for it. Case 2 holds the declaration: across the whole bundle, the one module the graph reaches for and does not carry is views/intent-feed.js, reached only dynamically, and the installed server answers it 404.

The falsifier: before and after

f10-1-widened.sh is #1144's F10.1 as batch H amends it: pip install ".[local]", and opendox generate-and-open --local …, with every other line as written. T075's widening is appended after its last line. Each run starts from a clean checkout, in a FRESH venv, with every OPENDOX_*/PG*/GIT_*/XF_* variable unset. Two host preconditions apply, and neither is a change to the command:

• OPENDOX_STATE_DIR=$HOME/.local/state/p3-t075. This host's /workspace/projects is mode 777, and the bundle refuses a state directory under a world-writable, non-sticky ancestor, by name.
• A python → python3 shim on PATH, for F10.1's python -m venv.

the script

# #1144 F10.1, as T007 batch H amends it (`pip install ".[local]"`, `generate-and-open --local`),
# with T075's widening appended after its last line: the fetch, made over every
# file of the bundle. Run with bash, from a CLEAN checkout of openDox-code only.
set -euo pipefail
W=$(mktemp -d)                                        # scratch space, resolved at run time (never a host path)
python -m venv --clear "$W/v10"                      # a FRESH environment: nothing already installed stands in
. "$W/v10/bin/activate"
pip install ".[local]"
if python -c "import openxdox" 2>/dev/null; then echo "FAIL: sibling present"; exit 1; fi
opendox --help >/dev/null                       # the console script MUST exist
export GIT_AUTHOR_NAME=fixture GIT_AUTHOR_EMAIL=fixture@example.invalid GIT_COMMITTER_NAME=fixture GIT_COMMITTER_EMAIL=fixture@example.invalid
R=$(mktemp -d)/plain-documents
cp -r tests/fixtures/plain-documents "$R"   # a FRESH repository (preamble)
git -C "$R" init -q
git -C "$R" add -A
git -C "$R" commit -qm fixture
opendox generate-and-open --local --repo-root "$R" --repository fixture --no-open --port 8080 &
SERVER=$!
trap 'kill "$SERVER" 2>/dev/null || true' EXIT   # cleanup cannot mask the verdict
ready=0
for _ in $(seq 1 30); do
  if curl -sf http://127.0.0.1:8080/ >/dev/null; then ready=1; break; fi
  sleep 1
done
test "$ready" -eq 1                              # a server that never started FAILS here
curl -sf http://127.0.0.1:8080/ > "$W/bundle.html"   # no pipeline: curl's status is the status
grep -qi '<html' "$W/bundle.html"                # and it is really the bundle
# ---- T075's widening: every one of 10.2's 42 files, by the same fetch ----
find src/opendox/web -type f | sed 's|^src/opendox/web/||' | sort > "$W/bundle.txt"
test "$(wc -l < "$W/bundle.txt")" -eq 42         # 10.2's count, from the tree
while read -r f; do
  curl -sf "http://127.0.0.1:8080/$f" > "$W/served" || { echo "NOT SERVED: $f"; exit 1; }   # curl's status is the status
  cmp -s "$W/served" "src/opendox/web/$f" || { echo "NOT THE TREE'S BYTES: $f"; exit 1; }
  echo "served $f"
done < "$W/bundle.txt"
test "$(curl -s -o /dev/null -w '%{http_code}' http://127.0.0.1:8080/views/intent-feed.js)" = 404   # 10.2a: not owed
echo "F10.1 (as amended), widened to 42 files: PASS"

Before, at #69's head 19e32f0c: F10.1's own fetch of / passes, and the widening stops at the first file the wheel does not carry (rc 1):

  database $HOME/.local/state/p3-t075/postgres/run (bundled, pid 1837880, migrations applied now: ['0001', '0002'])
  validation: opendox-snapshot: 0 violations, by opendox.validator, over its packaged copy opendox-snapshot (sha256 f9e3e111af1d)
  serving http://127.0.0.1:8080/index.html
served app.js
served index.html
served styles.css
NOT SERVED: vendor/.gitkeep
rc=1

After, at 67b8d9da (rc 0). Re-run at d61bb9e9, a7cf2665, cac12d91, a3cf5e9a, 0457b383, d7511388, 1e642b21, 0b0d762c and the final head 308523d9: rc 0 each time, 42 served, F10.1 (as amended), widened to 42 files: PASS.

A finding for T077, about F10.1's fixed --port 8080: the first run at a3cf5e9a read NOT THE TREE'S BYTES: views/doxbench-chat.js. The body it fetched was 101660 B, while the installed file is 98535 B and equal to the tree's. That run's log has no serving line from its own server. Another run on this shared host already answered 127.0.0.1:8080 while this run's bundled server was still starting, and the readiness loop cannot tell whose server answered. The re-run adds one marked T075 guard line before the start, which is not F10.1 text: if curl -s -o /dev/null http://127.0.0.1:8080/; then echo "FAIL: port 8080 already serves"; exit 1; fi. Its log shows its own serving http://127.0.0.1:8080/index.html before the first fetch. The in-suite case binds --port 0 and reads the URL its own server prints, so it is not exposed to this.

  database $HOME/.local/state/p3-t075/postgres/run (bundled, pid 2340484, migrations applied now: ['0001', '0002'])
  validation: opendox-snapshot: 0 violations, by opendox.validator, over its packaged copy opendox-snapshot (sha256 f9e3e111af1d)
  serving http://127.0.0.1:8080/index.html
served app.js
served index.html
served styles.css
served vendor/.gitkeep
served vendor/markdown-it.min.js
served views/account-menu.js
served views/board.js
served views/bullseye.js
served views/canvas-model.js
served views/canvas.js
served views/composed-model.js
served views/display.js
served views/doc-wheel.js
served views/docs.js
served views/doxbench-chat-model.js
served views/doxbench-chat.js
served views/doxbench-editor.js
served views/doxbench-save.js
served views/doxbench-state.js
served views/edit.js
served views/explorer.js
served views/funnel.js
served views/grouping.js
served views/helpers.js
served views/intent-binding.js
served views/lens-model.js
served views/lens.js
served views/lineage.js
served views/model.js
served views/notebook.js
served views/outline-model.js
served views/projection-index.js
served views/repo-selector-model.js
served views/repo-selector.js
served views/settings.js
served views/staging-workbench-model.js
served views/staging-workbench.js
served views/swb-model-intake.js
served views/view_extension.js
served views/viewer.js
served views/wheel-model.js
served views/wheel.js
F10.1 (as amended), widened to 42 files: PASS
rc=0

The in-suite falsifier, LANG=C.UTF-8 python -m pytest -q tests_runtime/test_served_bundle.py:

• before (this commit's pyproject.toml reverted to T072, 13.1: the bundled PostgreSQL server, the local install's own child (plan 034) #69's): 2 failed, case 1 with assert ['vendor/.gitkeep'] == [], and case 2 with the installed entry point does not serve: [('vendor/.gitkeep', 'status 404')];
• after: 2 passed in 7.70s.

Mutants: 8 applied, 8 killed (at 0457b383, and again at d7511388)

Each mutant was applied to the committed tree, tests_runtime/test_served_bundle.py was run, and the tree was restored.

mutant	killed by
M1: the dotfile pattern dropped (opendox = ["web/**"], #69's line)	both cases: ['vendor/.gitkeep'] omitted, then status 404
M2: the recursive pattern made single-level ("web/*")	both cases: vendor/markdown-it.min.js and every views/ file omitted, then 404
M3: the static route serves .js as text/plain (a DashboardHandler.extensions_map override)	case 2: ('app.js', "served as 'text/plain'"), and every other module
M4: app.js statically imports ./views/board-missing.js	case 2's module graph: ('views/board-missing.js', 404). The per-file fetch alone passes this one.
M5: a views/intent-feed.js shipped in the bundle (10.2a broken)	both cases: 43 == 42, and the declared absence set() == {'views/intent-feed.js'}
M6: a fake openxdox on the child's path	case 2's precondition: a sibling is importable: importable siblings: ['openxdox']
M7: views/projection-index.js, a carried dynamic target, statically imports ./missing-below-dynamic.js	case 2's walk: ('views/missing-below-dynamic.js', 404). It survived the walk at d61bb9e9, which did not follow dynamic targets (Copilot's "previously missed" finding).
M8: in views/doxbench-editor.js, only the commented multi-line clause's specifier points at ./doxbench-state-missing.js (Copilot, r4170756761)	case 2: ('views/doxbench-state-missing.js', 404); case 3: the named edge absent. It survived a3cf5e9a's reader (2 passed).

The suite and CI

• Locally, at the final head 308523d9, in a venv installed as validate installs it (--only-binary :all: -c constraints-cpython312-linux.txt -e ".[runtime,test]"), with LANG=C.UTF-8: 3443 passed, 177 skipped, 0 failed (3230 / 177 / 0 at bbd91a3f, 3225 / 177 / 0 at 0b0d762c, 3176 / 177 / 0 at 1e642b21, 3172 / 177 / 0 at d7511388, 3066 / 177 / 0 at 0457b383, 3061 / 177 / 0 at a3cf5e9a, 3035 / 177 / 0 at cac12d91, 3029 / 177 / 0 at a7cf2665). At 67b8d9da it was 3020 passed and 177 skipped, which is T072, 13.1: the bundled PostgreSQL server, the local install's own child (plan 034) #69's 3195 selected plus the 2 new cases. The 177 skips are the DB-backed cases on a machine with no OPENDOX_TEST_DATABASE_URL. CI runs them against its postgres:16 service, and EXPECT_SKIPPED stays 11, because the new module never skips under CI.
• CI: validate green at 67b8d9da, d61bb9e9, a7cf2665, cac12d91, 71d24af6, a3cf5e9a, 0457b383, d7511388, 1e642b21 and the final head 0b0d762c. SonarCloud green at each since d61bb9e9.
• SonarCloud: at 67b8d9da the quality gate failed on 3.4% duplication on new code (required ≤ 3%). Its one block was the pipe reader, copied from test_bundled_postgres.py:160-175. In 65b2058f the entry point's streams go to files, and _served_url() polls the stdout file for the flushed URL line. The gate passed at d61bb9e9.
• Copilot: round 1 (at 67b8d9da) raised 2 low findings, the stale "41 files" note and a precondition that did not name the sibling it found. Both are fixed in d61bb9e9, and both threads were answered and resolved. Round 2 (at 65b2058f) had no new finding. It listed 2 "previously missed" ones, the dynamic-import walk and S5's stale counts. Both are fixed in 4df679b6 and answered in this comment. Round 3 (at 4df679b6) was "Approval recommended", with no finding. Rounds 4, 5 and 6 (at a7cf2665, cac12d91 and 71d24af6) were "Needs a closer look", with no finding, for the stacked, unlanded T072, 13.1: the bundled PostgreSQL server, the local install's own child (plan 034) #69 alone. Round 7 (at a3cf5e9a) was "Changes recommended" on one finding: the import pattern missed a multi-line clause with commented quotes (r4170756761). It is fixed in d99e5ccc (M8 above), and the thread was answered and resolved. Rounds 8 through 11 (at 0457b383, d7511388, 1e642b21 and 0b0d762c) were "Needs a closer look", with no finding. Round 8 listed that thread as resolved. 0 unresolved review threads.

Not in this PR

• T077 (the F10.1 run itself, quoted in its own PR) and T076 (the root README.md).
• One observation, not a gap: the static route takes its content types from the platform's mimetypes table. On Linux and on CI that table answers text/javascript for .js. A host whose table does not would fail case 2 (M3 shows it would be caught). Pinning the types would be a serve.py edit, so it is left to the holder.

🤖 Generated with Claude Code

Summary by Sourcery

Package and verify the complete standalone web bundle so an installed openDox entry point serves all 42 files while keeping intent-feed.js explicitly out of scope.

Bug Fixes:

• Ensure installed wheels include the complete 42-file web bundle, including the tracked vendor dotfile.

Enhancements:

• Add runtime coverage verifying that an installed local entry point serves the full browser-loadable bundle and preserves the declared absence of intent-feed.js.
• Validate bundle module-graph closure, content types, package contents, process lifecycle, and import discovery from an isolated installation.

Build:

• Update setuptools package-data configuration to include dotfiles beneath the web bundle.

Tests:

• Add end-to-end tests for wheel contents and serving all web assets from an installed openDox entry point.
• Update package-data contract tests to enforce the recursive and dotfile glob patterns.

[Copilot]

Copilot review overview

🔵 Needs a closer look

It remains a blocked draft stacked on the unmerged draft PR #69.

Review effort: Balanced
Findings: None

[sonarqubecloud]

Quality Gate passed

Issues
1 New issue
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud

[Copilot]

Copilot review overview

🟢 Approval recommended

The packaging fix is focused and comprehensively exercised without unresolved correctness issues.

Review effort: Balanced
Findings: None

[brettheap]

READY at 308523d — phase 3 is open (T063 landed, #1218 → a883bbf6); T072 landed (#69 → 5e7ab00). The holder checked: validate and SonarCloud green; Copilot's latest review at exactly 308523d is "Approval recommended", Findings: None; 0 unresolved threads; no closing keywords in the body or any commit. #73 contains #69's final pre-squash head c04690f and merges main 5e7ab00 (merge-tree --merge-base c04690f); its diff against main is T075's four files. T075's After line (T056, T038, T063, T069, T072, T007 batch H) is met; T073 (#72) is a sibling, not a predecessor. Brett: draft phase 3 ahead, land in plan order when green. Lane: openxfactory-4 (openXfactory-4-openDox_extraction)

[sourcery-ai]

Sorry @brettheap, you've used your own review budget of 250,000 diff characters for the last 7 days.

You can request another review in 6 hours and 31 minutes by commenting @sourcery-ai review. Upgrade to get a review now.

[brettheap]

Lane: openxfactory-4 (openXfactory-4-openDox_extraction)

LANDED — lane openxfactory-4, 2026-10-03T15:00:18Z, PR #73 → 9197ccd (opensoft/openDox-code main; plain gate)

Brett: land phase 1 / phase 2 PRs when green