Strategy Selection Test Plan

Testing chooseContextInjectionStrategy — does the plugin correctly pick inject-full-content vs retrieval based on context window occupancy?

2.1 — Full-content injection

Setup: Small model context (e.g. 4K). Attach Quarterly Performance Report.txt alone, fresh chat. Expect: File + prompt tokens fit under the 70%-of-remaining-context threshold → strategy = inject-full-content. Full file text gets wrapped into the prompt.

2.2 — Retrieval triggered by size

Setup: Same small context. Attach OCR Korea AI Plan.pdf (long OCR output). Expect: Token count exceeds the threshold → strategy = retrieval. Citations returned instead of full text.

2.3 — Strategy flips as context fills up

Setup: Have a long back-and-forth conversation to raise contextOccupiedPercent, then attach a mid-size file that earlier in the same chat would have fit. Expect: The same file that would get inject-full-content early in the conversation switches to retrieval once context is fuller (available token budget shrinks).

2.4 — Follow-up with no new attachment

Setup: Attach a file, ask a question about it. Then send a follow-up message with no new attachment. Expect: Files from history (not the new message) get stripped and routed through retrieval, not full injection.

2.5 — No files at all

Setup: Plain text question, no attachments anywhere in the chat. Expect: Plugin is a no-op — message passes through unchanged.