RL-019 Explicit implementation of a proposed draft improvement

Observation

After the model suggested “Implement a more robust navigation system,” the user replied: “yes implement a more robust navigation system. make sure you keep a checklist.” The harness classified the message as ordinary chat. The model returned another four-item prose plan and asked “Would you like me to proceed with implementing this?” No code changed and no host checklist existed.

Baseline

• Intent: chat.
• Operation: chat, with draft revision 3 supplied only as metadata.
• Reply: 299 characters of planning and repeated confirmation.
• Draft after the turn: unchanged.
• Host checklist: absent.

Hypotheses

• H1: With an active draft, an agreement-prefixed request beginning “yes implement” and naming a page part routes to draft editing.
• H2: The existing code-edit path displays and completes its host-owned draft/check checklist and applies a valid exact edit.
• H3: The same wording without an active draft does not become a file or task mutation.

Intervention

• The bounded draft-edit recognizer now accepts implement as an edit verb.
• It removes a short agreement prefix (yes, yes please, okay, ok or sure) before matching a named page part such as navigation.
• The same agreement-prefix normalization now applies to bounded code follow-ups, so the later short form “yes, implement this” reaches the edit path instead of returning to planning.
• No general task or file routing changed. The rule is consulted as an edit only while the conversation owns an active code draft.
• HARNESS_VERSION is 2026-09-13-short-form-draft-implementation-9.

Method

1. Replay the reported sentence through classifyIntent with a complete HTML draft open.
2. Use MockLlm to return one valid exact replacement and run the real code conversation path.
3. Assert that the draft changes, the bounded edit prompt is used, and both host checklist items finish.
4. Classify the same sentence in an empty conversation and assert that it stays ordinary chat with no mutation route.
5. Run the full suite, typecheck and build.

Results

• H1 supported (C): the exact reported sentence routes to code editing when the draft is active.
• H2 supported (C): one exact edit changed the draft; $draft and $check both finished as done; no tool call occurred.
• H3 supported (C): without a draft, “yes implement a more robust navigation system” remains chat.
• Targeted code-conversation suite: 17/17 passed.
• Full suite: 708 tests, 0 failures. Typecheck and build pass.

Threats to validity

• The observed live model was not rerun after the change. The tests prove route, checklist and draft-state behaviour, not DeepSeek's edit quality.
• Agreement prefixes are a small explicit set. Unseen conversational leads may still miss the edit route and should be added only after observation.
• Browser verification quality is unchanged; a completed checklist means the existing supported checks passed, not that every navigation requirement did.

Conclusion

H1–H3 are supported by deterministic replay and harness tests. The reported requests, including the final “yes, implement this” form, no longer ask for permission twice and now reach implementation when a draft is active.

Follow-ups

1. Repeat the navigation request in the desktop app and inspect the real model's exact edits and browser checks.
2. Keep the model-proposed prose checklist as conversation context only; the host checklist remains the authority for work actually performed.

Reproduce

node --import tsx --test tests/codeConversation.test.ts
npm test
npm run typecheck
npm run build