457-silent-failure-hunt-ecc

457 — Silent Failure Hunt: ZAO Agent System

Status: Research complete Topic: agents Type: Audit & Debugging Date: 2026-04-20 Last-Validated: 2026-05-21 Original-Query: What silent failures exist in the ZAO agent system? Find and prioritize fixes. (reconstructed) Tier: v2-ready Tool: general-purpose subagent running silent-failure-hunter persona (ECC plugin agent not in registered subagent export — workaround used) Scope: src/lib/agents/ + 3 admin API routes

---

Key Decisions / Recommendations

Finding	Severity	Fix Priority
runner.ts:128 — Burn failure ignored after swap	CRITICAL	FIX THIS WEEK. Tokenomics broken (deflationary mechanism bypassed).
runner.ts:93 — Auto-stake failure unguarded, trade continues	CRITICAL	FIX THIS WEEK. Agent buys ZABAL unstaked.
autostake.ts:43-75 — Auto-stake catch swallows failure, no DB record	HIGH	Fix same PR as runner.ts:93.
events.ts:19-23 — logAgentEvent silently drops DB insert errors	HIGH	Fix this week. Audit trail loss.
burn.ts:31-35 — Burn error returns null, no DB log	MEDIUM	Fix same PR as runner.ts:128.
runner.ts:143 — Farcaster post failure doesn't fail trade	MEDIUM	Accept. Log only (trades > casts).
broadcast/route.ts:65-72 — audit logged after response	MEDIUM	Wrap in .catch() safe fire-and-forget.
agents/route.ts:30-42 — Promise.allSettled falls back to empty arrays on DB failure	MEDIUM	Throw on rejection instead of silent fallback.

Totals: 2 critical, 2 high, 4 medium, 0 low = 8 findings.

---

Comparison of Options (fix strategy)

Option	Scope	Pro	Con
A. Fix all 8 in one PR	big	fastest to clean state	larger review surface
B. Criticals only (runner.ts + autostake.ts + burn.ts + events.ts) in PR 1, rest in PR 2	focused	easier review	2 PRs
C. One PR per fix	atomic	tiniest reviews	8 PRs, overhead

Recommendation: B — criticals first for merge-to-prod speed, medium fixes as follow-up.

---

Detailed Findings

CRITICAL 1 — runner.ts:128 — Burn failure ignored

File: src/lib/agents/runner.ts:128

const hash = await executeSwap(agentName, quote);
await burnZabal(agentName, BigInt(quote.buyAmount));

await logAgentEvent({
  tx_hash: hash,
  status: 'success',  // <-- marked success even if burn failed
});

What breaks: burnZabal() returns null on failure (burn.ts:34). Trade logs as success regardless. ZABAL meant to burn accumulates in agent wallet.

Who won't notice: No one — no event logged, no alert. Tokenomics deflationary mechanism broken silently. Days or weeks of unexecuted burns.

Fix:

const hash = await executeSwap(agentName, quote);
const burnHash = await burnZabal(agentName, BigInt(quote.buyAmount));
if (!burnHash) {
  throw new Error('Burn failed after successful swap — trade marked failed');
}

---

CRITICAL 2 — runner.ts:93 — Auto-stake failure unguarded

File: src/lib/agents/runner.ts:93

try {
  // Auto-stake check (14-day cycle)
  await maybeAutoStake(agentName);

What breaks: maybeAutoStake() catches internally (autostake.ts:72-75), logs, returns. If Privy wallet rejects the stake tx, function exits normally. Runner proceeds to buy ZABAL. ZABAL accumulates unstaked.

Who won't notice: Admin dashboard shows no stake event (correct), no failure event either (missing). Agent keeps buying.

Fix: Change autostake.ts to throw on failure (see HIGH 1 below). Runner will then catch + halt cycle.

---

HIGH 1 — autostake.ts:43-75 — Auto-stake catch swallows failure

File: src/lib/agents/autostake.ts:43-75

try {
  await executeSwap(...approve data...);
  await executeSwap(...stake data...);
  await logAgentEvent({ status: 'success' });
  logger.info(`Auto-staked 100M ZABAL...`);
} catch (err) {
  const msg = err instanceof Error ? err.message : String(err);
  logger.error(`[${agentName}] Auto-stake failed: ${msg}`);
}

What breaks: Catch block only logs. No logAgentEvent({ status: 'failed' }). No rethrow. Caller (runner.ts:93) sees success.

Fix:

} catch (err) {
  const msg = err instanceof Error ? err.message : String(err);
  await logAgentEvent({
    agent_name: agentName,
    action: 'add_lp',
    status: 'failed',
    error_message: msg,
  });
  logger.error(`[${agentName}] Auto-stake failed: ${msg}`);
  throw err;  // <-- CRITICAL: stop caller from assuming success
}

---

HIGH 2 — events.ts:19-23 — logAgentEvent swallows DB errors

File: src/lib/agents/events.ts:19-23

export async function logAgentEvent(params: {...}) {
  const db = getSupabaseAdmin();
  const { error } = await db.from('agent_events').insert(params);
  if (error) {
    logger.error(`[${params.agent_name}] Failed to log event: ${error.message}`);
  }
}

What breaks: Insert fails (DB down, RLS, constraint), function logs and returns void. All callers await logAgentEvent(...) expecting durable log, get nothing. Audit trail silently incomplete.

Fix:

if (error) {
  throw new Error(`Failed to log agent event: ${error.message}`);
}

Risk: throwing might break some callers. Triage per call site — most callers should propagate (this is load-bearing audit).

---

MEDIUM 1 — burn.ts:31-35 — Burn error no DB record

File: src/lib/agents/burn.ts:31-35

} catch (err) {
  const msg = err instanceof Error ? err.message : String(err);
  logger.error(`[${agentName}] Burn failed: ${msg}`);
  return null;
}

What breaks: Null return, no agent_events row. Caller must check for null (currently doesn't — see CRITICAL 1).

Fix: Pair with CRITICAL 1 fix. Add logAgentEvent({ status: 'failed', ... }) before return null, OR change to throw so caller must handle.

---

MEDIUM 2 — runner.ts:143 — Farcaster post failure doesn't fail trade

Accept. Trades shouldn't depend on cast success. Log only.

---

MEDIUM 3 — broadcast/route.ts:65-72 — audit logged post-response

File: src/app/api/admin/broadcast/route.ts:65-72

    const data = await res.json();
    await logAuditEvent({...});
    return NextResponse.json({ success: true, hash: data.cast?.hash });

What breaks: audit await'd but after response commits. If logAuditEvent throws, response already sent successful, audit missing.

Fix:

logAuditEvent({...}).catch(err => logger.error('Audit log failed:', err));
return NextResponse.json({ success: true, ... });

---

MEDIUM 4 — agents/route.ts:30-42 — Promise.allSettled empty fallback

File: src/app/api/admin/agents/route.ts:30-42

const [configResult, eventsResult] = await Promise.allSettled([...]);
const configs = configResult.status === 'fulfilled' ? (configResult.value.data ?? []) : [];
const events = eventsResult.status === 'fulfilled' ? (eventsResult.value.data ?? []) : [];

What breaks: DB down → admin sees empty list, no error. Can't tell "no agents" from "DB failure".

Fix:

if (configResult.status === 'rejected') throw configResult.reason;
if (eventsResult.status === 'rejected') throw new Error('Failed to fetch agent events');

---

Process Notes

Tool used: Agent tool with subagent_type: general-purpose, prompt embedded silent-failure-hunter persona. ECC plugin's silent-failure-hunter agent exists as a file but isn't registered as a subagent by the plugin (only ~40 of 48 agents registered). Mid-session copy to ~/.claude/agents/silent-failure-hunter.md doesn't register until next Claude Code restart.

Time taken: ~50s for full sweep of 11 files.

---

ZAO Ecosystem Integration

Fixes touch:

• src/lib/agents/runner.ts — 2 critical fixes
• src/lib/agents/autostake.ts — 1 high fix
• src/lib/agents/burn.ts — 1 medium fix
• src/lib/agents/events.ts — 1 high fix (load-bearing)
• src/app/api/admin/broadcast/route.ts — 1 medium fix
• src/app/api/admin/agents/route.ts — 1 medium fix

Pair with:

• Vitest tests under src/lib/agents/__tests__/ and src/app/api/admin/__tests__/
• evals/ fixtures (doc 441) once populated with real scenarios

---

Sources

• ECC silent-failure-hunter agent definition: ~/.claude/plugins/cache/everything-claude-code/everything-claude-code/1.10.0/agents/silent-failure-hunter.md
• ZAO OS docs 441/442/448 (ECC integration)
• Silent failure anti-patterns reference