Reframe FM instructions positively and trim few-shots to two

[FuJacob]

Summary

Apple's WWDC25 prompt-design guidance is to use a short positive identity plus a small number of demonstrations rather than a long list of prohibitions, because the chat-tuned system model responds more reliably to "this is who you are" than to "do not do X." The previous FM session instructions were 10 lines of mostly negative rules with a 5-pair few-shot block. This rewrite collapses the rules to four positive lines and trims the few-shot block to two pairs (one prose, one code), shrinking the instruction prefix that lands in Apple's 4096-token shared context and re-pointing the chat prior at the actual task.

Stacked on feat/fm-streaming. Use the eval suite from the parent stack PR to compare drift / mid-word truncation rates against the previous instruction text.

Validation

xcodebuild test ... -only-testing:CotabbyTests/FoundationModelPromptRendererTests -only-testing:CotabbyTests/SuggestionEngineRouterTests CODE_SIGNING_ALLOWED=NO
→ Executed 9 tests, with 0 failures (0 unexpected)

swiftlint lint --quiet → no violations.

Manual: pick prefixes that historically drifted ("Hey Jacob, ", "Thanks for ", "Hi Sarah,\n\n") and verify the suggestion continues the sentence in voice instead of greeting the user back.

Linked issues

Refs the FM-quality investigation.

Risk / rollout notes

• Pure prompt-policy change. No protocol, request shape, or engine wiring change.
• Three test assertions changed shape: the old ones pinned on phrases ("text-continuation engine", "Do not repeat or quote the existing text.") that no longer appear. The new ones lock in the positive identity ("complete partially-typed text"), the output contract anchor ("Output the continuation only:"), specific forbidden-content tokens ("no greeting", "no markdown"), and the style line.
• A new test (test_sessionInstructions_includeExactlyTwoContinuationExamples) pins the few-shot count to two so the trim cannot silently regrow.
• The instruction-prefix cache from PR 2 (perf/fm-session-reuse-prewarm) is keyed on the rendered instructions string. After this lands, the first request on each session rebuilds the cache once with the shorter instructions; subsequent requests reuse it. No special migration needed.

Greptile Summary

This PR replaces the FM session instruction block with a shorter, positively-framed identity following Apple's WWDC25 prompt-design guidance, and trims the few-shot demonstration set from five pairs to two (one prose salutation, one code). No request shape, engine wiring, or output-normalizer logic changes.

• FoundationModelPromptRenderer.swift: the base rules array shrinks from 10 lines to 5 positive identity lines; continuationExampleLines shrinks from 10 strings (5 pairs) to 4 strings (2 pairs), keeping the prose-salutation and code cases that cover the two principal failure modes.
• CotabbyTests/PromptPolicyTests.swift: three tests are renamed and rewritten to match the new instruction text, and a new test_sessionInstructions_includeExactlyTwoContinuationExamples is added to pin the few-shot count so a silent regrowth would fail fast.

Confidence Score: 5/5

Safe to merge — this is a pure prompt-text change with no logic, protocol, or wiring modifications.

Every changed line is a string constant or a test assertion that mirrors it. The instruction-prefix cache rebuild on first request after landing is acknowledged and requires no migration. The new tests are accurate, and the few-shot trim is deliberate and well-reasoned.

No files require special attention.

Important Files Changed

Filename	Overview
Cotabby/Support/FoundationModelPromptRenderer.swift	Rewrites FM session instructions from 10 negative prohibition lines + 5 few-shot pairs to 5 positive identity lines + 2 few-shot pairs; production logic is unchanged — the diff is purely string constants.
CotabbyTests/PromptPolicyTests.swift	Three test renames/rewrites plus a new count-pinning test; assertions are accurate for the new instruction text, and the coverage split across tests is appropriate.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[SuggestionRequest] --> B[sessionInstructions]
    B --> C[5 positive identity lines]
    C --> C1["1. Positive identity: 'You complete partially-typed text'"]
    C --> C2["2. Output contract: no greeting/sign-off/quotes/markdown/labels/explanation"]
    C --> C3["3. Anti-echo guard: Continue from position after existing text"]
    C --> C4["4. Style match: language, register, casing, punctuation"]
    C --> C5["5. Context usage: clipboard/screen only when helpful"]
    B --> D{languageInstruction?}
    D -- yes --> E[Append language hint]
    D -- no --> F[Skip]
    E --> G["Examples header + 2 few-shot pairs"]
    F --> G
    G --> G1["Pair 1: prose salutation (anti-restart)"]
    G --> G2["Pair 2: code prefix → code continuation"]
    G --> H{customRules?}
    H -- yes --> I["'Your style preferences:' + rules + subordination line"]
    H -- no --> J[Skip]
    I --> K[lines.joined separator newline]
    J --> K
    K --> L[Instructions string → Apple FM channel]

Loading

_{Reviews (4): Last reviewed commit: "Restore no-echo rule to prevent empty co..." | Re-trigger Greptile}