DRC audit trail silently drops rows: Store to TIM Postgres builds SQL by concatenation and masks its own failure

[labgadget015-dotcom]

Found by a synthetic end-to-end run, not by a report.

What happened

Opening this issue routed correctly and ran the full council. n8n execution 7832 finished green in ~171s. No row reached tim.agent_runs.

Why the execution looked green

Store to TIM Postgres has onError: continueRegularOutput. The node failed, returned its error as ordinary output, and the workflow carried on to Log DRC to Drive and reported success. This is the same masking that hid three serialisation bugs for months.

The actual failure

Syntax error at line 3 near ":"

Line 3 is inside the interpolated issue body, at Context: the Event Router's ....

The node builds its INSERT by string concatenation, escaping only single quotes with .replace(/'/g, "''"). Postgres accepts the resulting SQL. I verified this directly against the same Neon database with a body containing a newline followed by a bare word and a colon, and it executed fine. So the parse failure happens inside the n8n node, before the statement reaches Postgres.

Impact

Any issue whose body trips this loses its audit row while the run reports success. tim.agent_runs is therefore not a reliable record of what the council processed. It currently holds 16 rows.

Fix

Convert the node to the parameterised queryReplacement form rather than concatenating values into the statement. That form is already proven in the Video Intelligence P0 workflow. Removing onError: continueRegularOutput afterwards is what makes a future regression visible.

Related

Same root cause class as the Normalise Issue Payload quote bug: hand-building a serialised format instead of using the safe constructor.

[labgadget015-dotcom]

@copilot Please analyze this bug report and suggest potential root causes and debugging steps.

[labgadget015-dotcom]

🤖 DRC Agent Analysis

Recommendation: 🔴 P0 CRITICAL

Summary: Solution 1 (Surgical Filter Enforcement with Idempotency Hash Guard) as recommended by REALIST, with elements of Solution 3 (Event Fingerprint Ledger) considered for future iteration

Next steps:

1. Step 1: Verify commit e0229648 is merged and export n8n Event Router workflow JSON to confirm 'edited' is absent from the Actionable Action? node filter array — this is the critical baseline check before adding the hash guard
2. Step 2: Run a single Supabase migration to create table issue_body_hashes (issue_number INT PRIMARY KEY, body_hash TEXT, updated_at TIMESTAMPTZ DEFAULT now()) — keep it minimal and separate from agent_runs schema
3. Step 3: Insert Code node 'Compute Body Hash' in n8n Event Router immediately after the sender [bot] filter using Node.js crypto: const h = require('crypto').createHash('sha256').update($json.body ?? '').digest('hex'); return [{json:{...$json, body_hash: h}}]
4. Step 4: Add HTTP Request node 'Check Existing Hash' performing Supabase REST GET to issue_body_hashes?issue_number=eq.{{$json.issue.number}} with Supabase service key auth
5. Step 5: Add IF node 'Hash Changed?' comparing computed body_hash to stored hash — equal branch routes to 'Log SKIP_DUPLICATE' Supabase insert into agent_runs (status=SKIP_DUPLICATE, no DRC call); unequal/absent branch upserts new hash and continues to DRC

Strategic fit: Consulting: high · Product: high · Tech debt: reduces

---

Analysed by GadgetLab DRC Agent (Dreamer → Realist → Critic) · Run run_1788509028963

[labgadget015-dotcom]

@copilot Please analyze this bug report and suggest potential root causes and debugging steps.

[labgadget015-dotcom]

Fixed and verified live.

The node now uses $1..$11 placeholders with the array-form queryReplacement already proven in the Video Intelligence P0 upsert. Active version a3aa3726, rollback 84bff55d. The version diff touches only this node.

Verified twice. First the statement was validated against the live Neon database using the exact body from this issue, inside a transaction that was rolled back. Then #285 exercised the real path with a deliberately hostile body.

Check	Result
Body stored	526 chars, SHA-256 identical to GitHub
Survived	8 colons, 3 quotes, 6 backslashes, a literal tab
Unicode	café, ✅ and 日本語 all preserved

Also added retryOnFail with three tries for transient database blips.

One thing is deliberately unchanged. The node still carries onError: continueRegularOutput, because it feeds the GitHub comment and the auto-PR dispatch, and a database outage should not block those. That means a genuine outage would still lose a row quietly. PR #286 covers that gap from the outside by asserting weekly that this node returned a row rather than a masked error.

[labgadget015-dotcom]

@copilot Please analyze this bug report and suggest potential root causes and debugging steps.