[khushalsonawat]

Stacked on #225 (TH-8804). This PR's base is feat/TH-8804-source-subfolder, so its diff shows only these fixes. Two commits: the fixes, then the repeated tests removed at review.

Summary

This answers the review's three code comments on src/fi/alk/harness/framework_detection.py, which reached #225 from TH-8805 (#229). Each comment was reproduced at 3e4e3ec5 before it was fixed.

• JavaScript counts as other source. .js, .jsx, .mjs and .cjs now count, as do .mts and .cts, like .ts. A Python folder beside them reads unknown instead of none. Files under test, example and similar folders still don't count, and a named framework still wins. File names are matched case-sensitively, as before.
• Any blocker comes before the chat-model rule. A chat model names langchain only when nothing the check couldn't read might be building the agent. Before, a folder with a chat model and, for example, an unread file beside it read langchain. It now reads unknown.
• A repository-origin process is scanned in the source. framework_folders_for takes the folder's base from source_origin, the same base the runtime builds the process from (process_runtime.py). Bundle-origin processes stay in the bundle. Every process the authoring step writes is bundle-origin, so only hand-written manifests change folder.

Ticket: TH-8805.

Known limits

• A LangChain agent that only uses a chat model, with no agent or graph builder, reads unknown whenever anything blocks. Triggers include test or eval files that import a framework, the app's own agents package, and JavaScript or TypeScript outside test and example folders. That includes config and built files such as tailwind.config.js, static/js/ and web/dist/. A StateGraph app still reads langgraph. This is the rule the review asked for. Narrowing it is possible later.
• .JS and .TS in upper case aren't matched, as before.

Verification

• tests/harness/test_framework_detection.py: 262 passed. tests/harness/test_hosted_entrypoint.py: 136 passed. tests/harness/test_scenario_source_provision_agent.py: 23 passed. All on Python 3.12.14, each file in its own run, on 9c187fe2 plus this commit. On Python 3.12.3, the sandbox's version, the detector and scenario-source files give 262 + 23.
• 10 planted bugs, one per changed rule and test, each caught by its named test.
• Over 640 synthetic folder findings, the decision changes in 14 rows. Every one goes from langchain to unknown, with a chat model and one blocker. No answer changes in the 101 design-review fixtures, the 19 examples/frameworks or the 2 harness fixtures.
• One cold change review, clear (High 0, Medium 2, Low 4). Both Mediums were about feat(harness): build an agent from a repo's subfolder #225's description text. Applied after that review, not reviewed: two test rows (.jsx under tests/, .mjs under examples/), one test comment, the description text, and the removal of the repeated tests asked for in this PR's review.
• ruff check passes on the three files.

🤖 Generated with Claude Code

[cdileep23]

Please check the comments.

[cdileep23]

This is now the same rule as build_process_trees in process_runtime.py (context.bundle_dir if process.source_origin == "bundle" else source_root). The comment above it exists because the two copies have to stay in step, and nothing enforces that. If a third origin is added, or the rule changes in the runtime, this copy will drift without a failing test.

Optional, since two copies is still within the rule of three: a small method on SourceProcess, for example def root(self, *, source: Path, bundle_dir: Path) -> Path, would let both call sites use one rule, and the comment could go.

[khushalsonawat]

Agreed that the two copies can drift. I'd do this as a follow-up together with the typing comment: the method needs the tests' SimpleNamespace processes replaced by real SourceProcess models, which is that change. I'm leaving this thread open until that's settled.

[cdileep23]

Not caused by this PR, but it now depends on it: framework_folders_for takes manifest: Any (and job: Any), so process.source_origin and process.working_directory are read with no type to check them against. A typo or a renamed field would only show up at runtime.

The manifest is an EnvironmentBundleV2, so typing the parameter as that would let a type checker see these reads. The tests pass SimpleNamespace fakes, so they would need the real model or a cast. Fine as a follow-up.

[khushalsonawat]

Agreed. I'd do this as a follow-up, together with the shared SourceProcess method from the other comment, since both need the tests' fakes replaced by the real model. No change in this PR. I'm leaving the thread open until that's settled.