fix: load models on GitHub Pages sub-path deployments (issue #13)

[konard]

Fix sub-path (GitHub Pages project-site) model loading — issue #13

Fixes #13.

The bug

On the deployed site https://link-assistant.github.io/model-in-browser/ (a
GitHub Pages project site, served under the /model-in-browser/ sub-path)
the app failed with:

Error: no available backend found. ERR: [wasm] importing a module script failed

No model ever loaded; Retry Load reproduced the failure every time.

Root cause

The inference worker derived ONNX Runtime Web's WASM wasmPaths from
self.location.origin, which is scheme + host + port only — it drops the
/model-in-browser/ sub-path. ORT then tried to import() its glue module
ort-wasm-simd-threaded.jsep.mjs from https://host/ort/… instead of
https://host/model-in-browser/ort/…, which 404s. The failed dynamic import
surfaces inside ORT as the generic "no available backend found" message.

It shipped unnoticed because the e2e suite ran against vite preview, which
serves at the origin root — where origin-based resolution is accidentally
correct, so the sub-path layout that triggers the bug was never exercised.

The fix

Resolve the ORT base from the deployment base, not the origin, in both
threads (defence in depth):

• App.tsx uses import.meta.env.BASE_URL (Vite's configured base, sub-path-aware)
  and posts it to the worker.
• worker.ts prefers that value; its fallback resolves ../ort/ relative to
  the worker's own location (<base>assets/… → <base>ort/), so it is correct
  under any sub-path with no origin involved.

Codebase-wide audit (R6)

The ORT wasmPaths derivation was the only origin-rooted asset path. The
Candle WASM engine, app JS/CSS bundles, and model weights are all resolved
relatively (Vite, base: './') or fetched from absolute remote URLs, so they are
already sub-path-safe. Full audit table in the case study.

Debug / verbose mode (R4)

Added ?debug=1 (or localStorage.mib_debug = '1'): the app and worker now log
the resolved ORT base path, so any future asset-path problem is self-diagnosing.

Tests (R1, R2)

• browser-commander e2e — web/e2e/browser-commander/subpath-load.mjs
  (npm run test:e2e:bc): boots a sub-path server, drives Chromium via
  browser-commander,
  loads SmolLM2 135M Instruct (q8), and asserts the ORT glue is served from
  the sub-path with no backend error. Passing.
• Playwright sub-path spec — web/e2e/subpath.spec.ts run by a new
  chromium-subpath project against web/e2e/subpath-server.mjs (reproduces the
  GitHub Pages project-site layout with production COOP/COEP headers). Fails on
  the pre-fix code, passes on the fix. Passing (9.9s).
• Manual Playwright verification: model loads and generates text end-to-end under
  the sub-path.

Case study (R3)

docs/case-studies/issue-13/README.md — timeline, full requirements table
(R1–R7), root-cause analysis, the fix, codebase-wide audit, the
report-upstream decision (R5: N/A, bug is in this repo), online research with
citations, and before/after screenshots.

Upstream report (R5)

Not applicable — the root cause is in this repository's code. ORT/transformers.js
behaved correctly (they requested the configured URL); the supported
env.backends.onnx.wasm.wasmPaths override is exactly what the fix uses.

Screenshots

Reported failure	After fix — ready	After fix — generation

Release

Adds a patch changelog fragment (changelog.d/20260610_150000_fix_subpath_ort_backend.md).

[konard]

⏳ Usage Limit Reached

The automated solution draft was interrupted because the Anthropic Claude Code usage limit was reached.

📊 Limit Information

• Tool: Anthropic Claude Code
• Limit Type: Usage limit exceeded
• Reset Time: in 2h 7m (Jun 10, 4:50 PM UTC)
• Session ID: 27a57714-f399-41b0-b055-a5ed9517b38f

🔄 How to Continue

Auto-resume is enabled. The session will automatically resume (with context preserved) when the limit resets.

🤖 Models used:

• Tool: Anthropic Claude Code
• Requested: opus
• Model: Claude Opus 4.8 (claude-opus-4-8)

📎 Execution log uploaded as Gist (3462KB)

• View complete execution log

---

This session was interrupted due to usage limits. The session will automatically resume when the limit resets.

[konard]

⏰ Auto Resume (on limit reset)

Auto-resuming automated work session at 2026-06-10T17:01:03.574Z

This session automatically resumed after the usage limit reset, continuing with the previous context preserved.

The PR has been converted to draft mode while work is in progress.

This is an auto-resumed session. Please wait for the session to finish, and provide your feedback.

[konard]

Working session summary

All requirements are addressed and PR #14 is finalized. Summary of what was completed:

Issue #13 — "It does not work" — Resolved

Root cause: On the GitHub Pages project-site deployment (https://link-assistant.github.io/model-in-browser/), the inference worker derived ONNX Runtime Web's WASM wasmPaths from self.location.origin — which is host-only and drops the /model-in-browser/ sub-path. ORT then 404'd on its glue module ort-wasm-simd-threaded.jsep.mjs (requested from /ort/… instead of /model-in-browser/ort/…), surfacing as the generic "no available backend found … importing a module script failed". It shipped unnoticed because the e2e suite ran against vite preview at the origin root, where origin-based resolution is accidentally correct.

The fix: Resolve the ORT base from the deployment base in both threads — App.tsx uses import.meta.env.BASE_URL; worker.ts falls back to a worker-relative ../ort/. Neither involves origin, so it's correct under any sub-path.

Requirements delivered:

• R1 — browser-commander e2e test (npm run test:e2e:bc) loading SmolLM2 135M Instruct under the sub-path — passing locally and in CI.
• R2 — Playwright sub-path regression spec + manual verification; root cause fixed.
• R3 — docs/case-studies/issue-13/README.md: timeline, requirements table, root-cause analysis, codebase audit, online research with citations, before/after screenshots.
• R4 — ?debug=1 / localStorage.mib_debug verbose mode logging the resolved ORT path.
• R5 — N/A (bug is in this repo's code; documented why).
• R6 — Audited all asset paths; ORT wasmPaths was the only origin-rooted one.
• R7 — All in PR fix: load models on GitHub Pages sub-path deployments (issue #13) #14.

Verification: All three CI workflows pass (CI/CD Pipeline, Build and Deploy, E2E Tests including the new browser-commander step). Working tree clean, PR marked ready, changelog fragment added.

PR: #14

---

This summary was automatically extracted from the AI working session output.

[konard]

🔄 Solution Draft Log (auto resume on limit reset)

This log file contains the complete execution trace of the AI solution draft process.

💰 Cost estimation:

• Public pricing estimate: $10.088122
• Calculated by Anthropic: $4.607969
• Difference: $-5.480153 (-54.32%)

📊 Context and tokens usage:

Claude Opus 4.8: (3 sub-sessions)

1. 113.5K / 1M (11%) input tokens, 23.6K / 128K (18%) output tokens
2. 114.1K / 1M (11%) input tokens, 32.2K / 128K (25%) output tokens
3. 75.4K / 1M (8%) input tokens, 20.9K / 128K (16%) output tokens

Total: (25.0K new + 488.7K cache writes + 9.9M cache reads) input tokens, 76.7K output tokens, $10.060093 cost

Claude Haiku 4.5:

• 15.4K / 200K (8%) input tokens, 518 / 64K (1%) output tokens

Total: 15.4K input tokens, 518 output tokens, $0.028029 cost

🤖 Models used:

• Tool: Anthropic Claude Code
• Requested: opus
• Main model: Claude Opus 4.8 (claude-opus-4-8)
• Additional models:
  • Claude Haiku 4.5 (claude-haiku-4-5-20251001)

Note: This session was automatically resumed after a usage limit reset, with the previous context preserved.

📎 Log file uploaded as Gist (1623KB)

• View complete solution draft log

---

Now working session is ended, feel free to review and add any feedback on the solution draft.

[konard]

✅ Ready to merge

This pull request is now ready to be merged:

• All CI checks have passed
• No merge conflicts
• No pending changes

---

Monitored by hive-mind with --auto-restart-until-mergeable flag