Testing Nexus

Context: every Jest lane resolves obsidian to a hand-written mock — see moduleNameMapper in jest.config.js. A green suite therefore proves the code agrees with tests/mocks/obsidian/, not with Obsidian. Three ranking defects shipped past that suite. Everything below exists to close the gap between what Jest can prove and what the running plugin does.

Workflow

1. Pick the job and open its protocol. Work from the protocol, not from this router — the router names procedures, it does not contain them.

   Job	Protocol
   Add or change a test	protocols/write-a-test.md
   Prove a change works in the running plugin	protocols/live-loop.md
   No Obsidian here (container / cloud session)	protocols/headless-obsidian.md
   Run a gated live-smoke lane or the eval harness	protocols/run-gated-lanes.md
   shippedGuidanceCommands or ToolManagerCliSyntax is red	protocols/fix-shipped-docs-drift.md
   Merge a PR (what to run first, how to gate on CI)	protocols/merge-a-pr.md
   A symptom you cannot place	references/troubleshooting.md

2. Derive every list from the tree, never from a document. This skill states no test counts, no lane inventory and no env-var table on purpose, and you MUST NOT add one — those rot silently and a rotted list is worse than none. references/lanes.md holds the discovery commands.

3. Before calling a testing change done, run both checks from the repo root and fix anything they print:

   python3 .claude/skills/nexus-testing/scripts/check_live_lane_gates.py
   python3 .claude/skills/nexus-testing/scripts/check_catalog_target.py
   

4. NEVER claim a change is verified in the app on the strength of Jest alone. If the change touches load order, rendering, hydration or anything that only exists once the plugin is running, protocols/live-loop.md is the proof and nothing else is.

5. At the end of a session that used this skill, run protocols/self-refine.md.

Map

• protocols/ the procedures: write-a-test, live-loop, headless-obsidian, run-gated-lanes, fix-shipped-docs-drift, self-refine.
• references/ read on demand: lanes.md (where a test goes and how to find what exists), mock-honesty.md (the two habits that make a test able to fail), troubleshooting.md (symptom → cause).
• scripts/ the mechanical checks named in step 3. Run them; do not reimplement them.
• refinement-log.md what past sessions changed here and why.

Siblings — do not duplicate them here

• nexus-eval-harness — configuring, extending and debugging the eval harness. This skill covers only running it and the two knobs that surprise people.
• nexus-model-eval — grading specific models.
• nexus-tool-schemas — regenerating the tool catalog (no running vault needed).
• nexus-agents — the two-tool contract and the useTools payload shape a test has to build.
• nexus-mobile-compat — the mobile failure class the in-app loop cannot see.
• nexus-storage — resetting persisted state when a plugin reload is not enough.

Status of the in-app loop

Exercised 2026-08-14 against Obsidian 1.13.7 in a headless Linux container and 2026-08-21 against Obsidian 1.12.7 on macOS. Setup, vault-targeted plugin:reload, dev:errors, dev:console, dev:screenshot, dev:debug, and synchronous eval are confirmed working. The first run found a startup-ordering bug that every Jest lane was blind to; the macOS run confirmed explicit vault=<name> targeting and clean repeated reloads.