fix: make NO_REPLY suppression and sanitization explicit

[lilyshen0722]

Outcome

Agent messages now use two explicit sentinel contracts:

1. suppression is total-match — only sentinel-only replies stay silent;
2. sanitization removes bare mid-content producer leakage while preserving deliberate code-formatted mentions.

Closes #783.

Changes

• keep whole-reply suppression for bare, duplicated, whitespace-separated, and fenced sentinel-only gateway output
• strip bare mid-content tokens without suppressing substantive replies
• preserve sentinel mentions inside variable-length backtick spans and fenced code
• preserve the historical outer-fence unwrap for substantive fenced replies
• align the system-exchange fallback suppression rule
• update the operator doc
• add unit and runtime-route coverage for bare leakage, code-formatted mentions, and fenced-silence precedence

Proof

Node 20:

• focused unit + runtime-route suites: 34/34 green
• tsc --noEmit -p tsconfig.typescheck.json: clean
• mutation removing bare-token stripping: exactly the paired bare-leakage unit test turns red
• mutation stripping inside code spans: exactly the paired code-formatting unit test turns red
• mutation allowing fenced sentinel-only output through: exactly the paired precedence test turns red

The first code-span implementation triggered a high-severity CodeQL polynomial-regex alert on attacker-controlled message text. The shipped implementation uses delimiter pairing plus a linear scan and cleared CodeQL at the prior head; fresh checks are running for this comment/test-only follow-up.

Focused ESLint is not a useful signal for these legacy test files: it reports their pre-existing TS-module resolution and iteration-rule violations. No new lint finding was identified in the changed lines.

Scope boundary

• Apply total-match NO_REPLY semantics in legacy agent bridges #784 tracks two legacy bridge .includes("NO_REPLY") matchers outside this kernel storage path.
• _external/clawdbot is not initialized in this workspace, so its separate sentinel semantics remain explicitly unaudited.
• The CLAUDE.md/skill wording correction and complete cross-layer checklist remain in the gated sprint-close docs batch.

[lilyshen0722]

Review — @sprint-review · 47c94325

Verdict: approve with one edge case to either fix or document. The fix is better than the line-anchoring I proposed, and the proof holds up.

Fails-without-the-fix: confirmed, precisely

Ran the PR's suite against main's agentMessageService.ts with the new tests in place:

✕ preserves code-formatted sentinel mentions
Tests: 1 failed, 4 passed

Exactly one test, no collateral. The new contract is genuinely tested.

No assertion inversions — worth noting explicitly

Checklist rule 3 would have required a cited ruling for any flipped assertion on main. There are none: 'Reply with NO_REPLY when done.' → 'Reply with when done.' is preserved unchanged. The contract change is strictly additive — bare-token leakage behaves as before, backtick-quoted mentions are newly preserved. That's the cleanest possible shape for this change and it means the rule-3 question never arises.

The design is better than what I suggested

I proposed line-anchoring (strip only when the sentinel is alone on its line). This distinguishes bare token = producer leakage from backtick-quoted = deliberate mention, which is a sharper cut: it preserves the case that actually matters (evidence quotes, code samples, docs) without preserving genuine model leakage in prose. The variable-length backtick matcher handles inline spans and fenced blocks without touching whitespace-sensitive content around them, which keeps the 2026-07-16 indentation fix intact.

The systemExchangeTriggers fallback being aligned in the same PR is the right call — otherwise it becomes the fourth semantics.

---

The edge: a fenced sentinel-only message is suppressed, contradicting the code's own comment

The comment above the fence branch says:

A fully fenced reply is explicitly code-formatted … Preserve sentinel mentions inside it exactly.

That isn't true when the fence contains only the sentinel, because the total-match check runs before the fence check, against post-fence-strip content:

const stripped = outerFence ? outerFence[1] : raw;
const trimmed  = stripped.trim();
if (/^(?:NO_REPLY\s*)+$/.test(trimmed)) return '';   // ← fires first
if (outerFence) return trimmed;                       // ← never reached

Verified:

input	result
```text\nNO_REPLY\n```	'' — message vanishes
`NO_REPLY`	`NO_REPLY` — preserved ✅

So a doc or review comment that illustrates the sentinel as a standalone fenced block posts as nothing. That is precisely the documentation use case this PR exists to enable — the KB entry describing this rule, if it shows the token as a fenced example, disappears.

This is a real tradeoff, not an obvious bug, which is why I'm not blocking on it. Reversing the order means a model that wraps its silent reply in a code fence would post a visible NO_REPLY into a pod — the current order is defensively correct for that case. Two acceptable resolutions:

1. Keep the order, fix the comment, and add a test pinning fenced-sentinel-only → '' as intended behavior rather than an accident. (My preference — it's the safe direction, and an untested edge is how this whole class of bug starts.)
2. Exempt the outer-fence case from total-match, accepting the leakage risk.

Either way the current state has a comment asserting something the code doesn't do — which is the phantom-contract shape the sprint's own checklist rule 10 names.

Two consistency items outside this PR's diff

• CLAUDE.md:418 still says the sentinel "will be sent verbatim." It is not verbatim — bare tokens are stripped, before and after this PR. This was flagged as an open action two rounds ago and is still open; it isn't in this diff. Now that fix: make NO_REPLY suppression and sanitization explicit #785 changes the surrounding behavior, the line is wrong in a new way and should ride whichever PR lands next.
• The review-checklist draft's rule 7 records a superseded ruling. It states mutation is "stripped at the delivery layer … word-boundary" — the pre-fix: make NO_REPLY suppression and sanitization explicit #785 behavior. fix: make NO_REPLY suppression and sanitization explicit #785 changes that to bare-stripped/backtick-preserved. The checklist is landing in a different PR, so two in-flight artifacts now disagree about the same contract. Whichever lands second needs to reconcile.

What I did NOT verify

• I ran the unit suite and reproduced the main-comparison; I did not run the clawdbot-e2e service test in this diff, nor the full backend suite.
• CI status on 47c94325 not checked at review time.
• The edge-case results above are from direct calls to sanitizeAgentContent, not from a live pod post.
• #784's two legacy .includes(...) bridges are correctly out of scope here; I have not looked at them.