feat(base2): add JavaScript Control-Flow Verification Rule to system prompt...

[chicagoist]

prompt...

Pull Request: Add JavaScript Control-Flow Verification Rule to Buffy Agent

---

Repository: CodebuffAI/codebuff
Target file: agents/base2/base2.ts
Target section: systemPrompt → # General guidelines
Branch: main

---

Title

feat(base2): add JavaScript Control-Flow Verification Rule to system prompt

---

Problem

When debugging JS in HTML files, the agent falls into a "syntax-blind loop":
it sees node --check fail, fixes brace imbalances mechanically, and declares
success when the parser passes — without verifying the code actually executes.

Real-world case study (30-round debugging cycle)

1. Agent ran node --check → SyntaxError (missing braces)
2. 10+ rounds of string-replacement fixes
3. node --check passed → agent declared "fixed"
4. All interactive features were still broken

Root cause: quiz logic was wrapped in if (boxes.length > 0) where boxes = querySelectorAll('.answerbox[contenteditable]'). Those elements were removed
during a UI migration to checkboxes. boxes.length was always 0. Code was
syntactically valid but logically dead. The agent had grep, file reading,
and DOM analysis — but no mandatory checklist to use them.

This is a recurring class of bugs. The fix is not a new tool — it's a
mandatory methodology encoded in the system prompt.

---

Proposed Change

One new bullet in # General guidelines, placed after the closing backtick of
the Composio conditional block (${ENABLE_COMPOSIO_TOOLS ? ... : ''}) and
before the literal line # Spawning agents guidelines.

Unified diff

--- a/agents/base2/base2.ts
+++ b/agents/base2/base2.ts
@@ systemPrompt template literal, General guidelines section @@
 - **Discover and install skills:** ...
 ${ENABLE_COMPOSIO_TOOLS
   ? `...`
   : ''
 }
+- **JavaScript Debugging:** After fixing JS syntax errors, do NOT stop at
+  \`node --check\` passing — the code may be syntactically valid but logically
+  dead. Verify: (1) every \`querySelector(All)\` target exists in the HTML body,
+  (2) execution reaches every \`addEventListener\` (trace from IIFE entry through
+  all wrapping \`if\` conditions), (3) no \`if (elements.length > 0)\` guard wraps
+  live logic around removed DOM elements. Only after all three checks pass should
+  the fix be considered complete.
 # Spawning agents guidelines

Rationale for single bullet

Existing guidelines use compact - **Topic:** Actionable instruction format.
This follows the same style. The three sub-checks are inline for brevity while
remaining specific enough to execute.

---

The Three Rules (reference)

#	Rule	Check
1	DOM cross-reference	grep '<selector>' *.html → at least 1 match must exist
2	Execution-path trace	From IIFE entry to each addEventListener — every wrapping if must be reachable
3	boxes.length > 0 pattern	Any guard wrapping >10 lines of logic is a red flag after UI migration

---

Contribution Checklist

• Change scoped to agents/ (per CONTRIBUTING.md)
• No backend, database, billing, deployment, or secret-management code
• Follows existing code style (template literal, bullet format)
• Single bullet, minimal change — no structural refactoring
• Applies to all modes (free/default/max/fast) — rule is in shared systemPrompt

---

Alternatives Considered

Approach	Verdict
New instructionsPrompt section	❌ Too verbose, breaks existing flow
Separate skill file	❌ Must be loaded manually each session
Single bullet in Guidelines	✅ Always active, minimal, consistent

---

License

By submitting this pull request, I confirm that my contribution is made under
the terms of the project's license.

[codebuff-team]

Thanks for the detailed writeup and the root-cause case study — that's a real class of bug (syntactically valid but logically dead code due to stale selectors) and worth thinking about.

That said, I'd hold off on porting this as-is:

1. Scope mismatch: base2.ts's systemPrompt is the general-purpose guideline set used across all agent tasks, not just JS/HTML debugging. Adding a JS-specific, DOM-specific checklist here bakes a narrow heuristic into every prompt, for every language and task. This bloats the prompt for the ~95% of sessions that never touch a querySelector.

2. Overfit to one anecdote: the rule is derived from a single 30-round debugging session. It's plausible this generalizes, but there's no evidence here that this wording measurably improves agent behavior versus just adding tokens. Prompt changes like this really want before/after eval results (e.g., against a small suite of "looks fixed but isn't" repro cases), not just a plausible-sounding fix.

3. Placement/mechanics: the diff correctly targets the general guidelines section and matches the existing bullet style, so mechanically it's fine — no issue there.

4. Better fix location: if this genuinely recurs, a more targeted approach would be a specialized JS-debugging skill/agent invoked when doing HTML/JS work, rather than a permanent guideline in the base prompt.

If you have eval numbers or several repro cases showing this rule reduces false 'fixed' declarations, that would make a much stronger case for landing something like this — possibly as a narrower, generalized rule (e.g., 'don't declare a fix complete based on static checks alone; verify the code path executes') rather than DOM/JS-specific mechanics.