fix(core): stop isUnusableStage1Reply deferring valid bare-code replies (#11504)

[lalalune]

Root cause

isUnusableStage1Reply (packages/core/src/services/message.ts) classified a Stage-1 simple-path reply as junk when any character repeats 5+ times anywhere in the text:

if (/(.)\1{4,}/u.test(trimmed)) return true;

Two-level code indentation is 8+ consecutive spaces, so a gemma-4-31b HumanEval-style bare python body

def has_close_elements(numbers: List[float], threshold: float) -> bool:
    for idx, elem in enumerate(numbers):
        for idx2, elem2 in enumerate(numbers):
            ...

matched the rule and got replaced with "I'm not sure how to answer that.". The starts-with-fence escape (/^```[a-z0-9_-]*\s+/iu) only rescued replies that begin with a code fence — bare code bodies, prose + fenced block replies, pretty-printed JSON, and markdown ----- dividers all died. This is the runtime-pipeline gap called out in .github/issue-evidence/10199-gemma-4-31b-cutover/review-package-multiharness/scorecard.md: eliza-harness HumanEval 0.40 vs 1.00 for the same model (gemma-4-31b on Cerebras) on hermes/openclaw raw harnesses (~60% of code turns deferred), and in production it discards legitimate code/structured replies.

Fix

Tighten the degenerate-run rule so it fires only when the whole reply is a single repeated code point (non-whitespace content, length ≥ 5, exactly one unique character):

const nonWhitespace = [...trimmed.replace(/\s+/gu, "")];
if (nonWhitespace.length >= 5 && new Set(nonWhitespace).size === 1) {
	return true;
}

No benchmark-mode special case — the unusable signal is now scoped to actual degenerate output, which fixes production code-reply UX and the benchmark cell at once.

Deliberately preserved protections (all still return unusable):

• empty / whitespace-only replies
• JSON-scaffold fragments ({}, ::::, ,,,)
• digit-only replies (still rescued downstream by isTerseReplyWorthKeeping when the answer is genuinely numeric, e.g. "4" for "What is 2+2?")
• non-allowlisted 2–8 char all-caps enum echoes (RPPY, RESPOND) with the Say PONG intentional-token rescue intact
• single-character spam, including across whitespace (aaaaa, !!!!!!!!, aaaaa aaaaa)

Behavior deliberately loosened: a 5+ repeated-char run inside a longer reply (code indentation, ----- dividers, ..... ellipses, "soooo") no longer marks the whole reply unusable — that was the bug.

Tests

New regression test delivers bare-code and structured Stage 1 replies verbatim drives the full runV5MessageRuntimeStage1 path with: (a) a gemma-style bare nested-indent python body, (b) a fenced code block, (c) prose + fenced code, (d) pretty-printed JSON, (e) a markdown divider reply — each ships verbatim with exactly one model call. Extended the known-junk test with " " (whitespace-only), "!!!!!!!!", and "aaaaa aaaaa".

bunx vitest run src/__tests__/message-runtime-stage1.test.ts
 Test Files  1 passed (1)
      Tests  74 passed (74)

bunx vitest run src/__tests__/message-action-dedupe.test.ts \
  src/__tests__/message-benchmark-integration.contract.test.ts \
  src/__tests__/message-routing-live-regression.test.ts \
  src/__tests__/messages-format.test.ts \
  src/__tests__/message-stable-prefix.test.ts \
  src/__tests__/message-stage1-context-catalog.test.ts
 Test Files  6 passed (6)
      Tests  54 passed (54)

bunx turbo typecheck lint "--filter=@elizaos/core"
 Tasks:    61 successful, 61 total

Refs #11504

🤖 Generated with Claude Code

[greptile-apps]

Your trial has ended. Reactivate Greptile to resume code reviews.

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: c7fbad06-f041-4923-adf8-0df44d103a15

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fix/11504-stage1-bare-code-reply

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[greptile-apps]

Your trial has ended. Reactivate Greptile to resume code reviews.

[claude]

Claude encountered an error —— View job

---

I'll analyze this and get back to you.