renezander030

Dropped without a count

A context layer over chat, tickets, wiki, CRM, mailboxes and meeting notes has to answer as the person asking. The usual build gets this wrong in a way that passes every demo: one service account reads everything, an index is built from it, and a filter at query time trims the result. The filter runs after the read. It leaks in 3 places.

The 3 leaks

1. Counts. "How many pages in the sales space mention Acme" answered with a number, for a caller who cannot open the sales space.
2. Titles. A trimmed result that still shows the title of a meeting the caller was not in.
3. Provenance chains. A fact shared with the caller whose chain links to a write-up the caller cannot open, so the chain becomes a directory of things they are not allowed to see.

The pattern

• Every read runs delegated, with the caller's own token. The layer holds no permissions of its own, the sources enforce theirs, which they already do well.

renezander030

Never let an agent write to the knowledge graph: queued proposals, one writer, and why a refusal that exits 0 is not a refusal

Retrieval bugs are loud. Write bugs are silent — a concurrent add_episode() puts your facts in the wrong graph, a refused proposal that exits 0 tells the caller it succeeded, and both look exactly like success until someone reads the data months later.

Last tested: September 2026 — graphiti-core 0.29.3, FalkorDB graph module ver 42003 on Redis 8.6.3, Python 3.13. See Changelog at the bottom.

If this saves you a re-ingest, follow @renezander030 — production notes on agent pipelines, retrieval and approval gates.

Reference implementation: graphiti-local (kg propose / kg-workspace approve / drain).

renezander030

Vector dimension mismatch, expected 1024 but got 768 — the embedding width your graph stored is not the one your config claims (Graphiti + FalkorDB)

Nothing errors when the wrong-width vector is written. The write path truncates with a Python slice, and a slice past the end of a list is a no-op. You find out at query time, from the database, about rows you ingested weeks ago.

Last tested: September 2026 — graphiti-core 0.29.3, FalkorDB graph module ver 42003 on Redis 8.6.3, Ollama nomic-embed-text, Python 3.13. See Changelog at the bottom.

If this saves you a re-ingest, follow @renezander030 — production notes on agent pipelines, retrieval and approval gates.

The preflight checks described here ship in graphiti-local as kg doctor.

renezander030

Validate before you log: the RAG query guard that belongs above your first log line (TypeError: Cannot read properties of null (reading 'substring'))

The first thing most pipelines do with user input is interpolate it into a debug log. If validation sits one line below that, your "guard" runs after the crash — and everything that doesn't crash goes into the logs unvalidated.

Last tested: August 2026, Node 22 / TypeScript. See Changelog at the bottom.

If this saves you a confusing stack trace, follow @renezander030 — production notes on agent pipelines, retrieval and approval gates.

Real-world fix using this pattern: tetherto/qvac#3729, merged into Tether's QVAC RAG package.

renezander030

GPT-5 / o-series vision calls failing in Go: "this model is not supported MaxTokens, please use MaxCompletionTokens" — and the empty response with finish_reason: "length"

Renaming max_tokens is not enough. go-openai rejects reasoning-model requests client-side on seven independent grounds, and even a request that passes can return zero content with no error at all.

Last tested: August 2026 — go-openai master (reasoning_validator.go), GPT-5 / o-series vision endpoints. See Changelog at the bottom.

If this saves you a debugging afternoon, follow @renezander030 — production notes on agent pipelines, retrieval and approval gates.

The fix described here is merged in Tencent's WeKnora: Tencent/WeKnora#2614

renezander030

The Agent Review Cost Worksheet

What every AI-authored change costs you in senior review hours, and how to buy that cost down.

Most organisations book a coding agent as a productivity gain and leave review as a background activity nobody owns a number for. This worksheet produces that number and then attacks it. Forward it to engineering, finance, or procurement as-is.

---

Step 1: the figure nobody can produce

renezander030

Agent memory that degrades instead of failing: keyword fallback when the embedding provider dies ("Premature close", APIConnectionError)

Retrieval with a single vector path goes to zero recall the moment the embedder is unreachable. Two branches plus reciprocal rank fusion keeps it answering.

Last tested: August 2026. See Changelog at the bottom.

If this saves you an outage, follow @renezander030 — production notes on agent memory, retrieval and approval gates.

Reference implementation: github.com/renezander030/agentic-task-system

renezander030

The AI Sign-Off Checklist

Trust basis for AI-shipped work: eight questions before an agent's output lands.

Before you approve anything an AI agent produced, run these. If you cannot answer them, your sign-off documents a decision without securing it. Forward this to engineering, security, or procurement as-is.

---

Everything reduces to two questions

renezander030

Stop the "invisible malware" attack on your AI coding agent (30-line hook)

TL;DR — A malicious GitHub repo can own your machine the moment an AI coding agent touches it, without a single line of malware in the repo. The payload lives in a DNS TXT record and is pulled at runtime, so scanners, code review, and the agent itself never see it. Below is the attack in plain terms and a small, fail-safe PreToolUse hook that blocks the two actions the attack depends on — while leaving your own repos untouched.

Reported by 0DIN (Mozilla's GenAI bug bounty), write-up via

renezander030

One query across GitHub, Notion, and your task app — cross-source agent memory for Claude Code (RRF, no vector database)

Connectors give your agent access to one tool each. This fuses them: a single search across every backend, ranked by meaning, each result tagged with where it came from.

Last tested: 2026-06-19 — @reneza/ats-mcp + @reneza/ats-adapter-composite. See Changelog at the bottom.

If this saves you setup time, follow @renezander030 — practical AI / coding-agent infrastructure notes.

Full repo with the MCP server + adapter source: github.com/renezander030/agentic-task-system