umbra AGENTS.md

A project instruction file for Umbra, an auditing tool that finds behaviors a test suite cannot observe and demonstrates each gap with a failing test. It defines the project's build, code, testing, and submission rules.

In plain words
What is it for?
Use it when working on Umbra to follow its required documents, protect the submission scope, apply its coding principles, and test its coverage-auditing behavior.
Why use it?
It gives coding agents one source of guidance for what Umbra should build and how it should be developed. It also keeps the implementation aligned with the project's documented claims and demo.

Instructions file for CodexOpenCode

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add instructions/openhonest/umbra/agents-md
Clone the repo
git clone --depth 1 https://github.com/openhonest/umbra

Made for: Codex, OpenCode.

Per session 2,438 This file is loaded in full into every session.
When invoked 2,438 The same file — it is already loaded in full.
Security scan C 1 finding. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.02438 $0.02438
Opus 5 $0.01219 $0.01219
Sonnet 5 $0.00488 $0.00488
Haiku 4.5 $0.00244 $0.00244

Measured 3d ago against content hash f8e8fe452759, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade C, and why

umbra AGENTS.md scanned grade C with 1 finding against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 3d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Recursive force deletehighDestructive command

rm -rf with a variable or a broad path is one typo away from removing the wrong tree.

rm -rf directory # NOT: rm -r directory
AGENTS.md · 171 lines

How it starts

The opening of the file, as written. The whole thing — 171 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Umbra: agent instructions

Umbra is an apophatic code auditor: it reports what a test suite is structurally unable to see, and proves each gap with a real failing test it writes. Built for the OpenAI Build Week Challenge (submit by 2026-07-21).

Read these first, in order:

  • docs/SPEC.md is the authoritative build target. Build to it. If something here and SPEC.md ever seem to disagree, SPEC.md wins on what to build, this file wins on how to work.
  • SUBMISSION.md is the problem/solution/approach narrative. Keep the code honest to what it claims.
  • docs/DEMO.md is the 60-second demo the whole thing serves. The failing-case reveal is the point; protect it.
  • docs/HONEST-CODE-PRINCIPLES.md is the canonical standard for how Umbra itself must be written. Not a suggestion.
  • docs/HONEST-TEST.md is the coverage theory Umbra operationalizes. Read it before you design the silence model; the Silence Score measures the sampling-versus-total gap it defines, and its honesty properties (purity, mutation, idempotency) are silence categories Umbra should detect.

The one rule that defines this project: thoughtful use of GPT-5.6

Umbra is judged partly on thoughtful model use. That means using GPT-5.6 only where it belongs, and never anywhere else.

  1. Structure is deterministic. No model calls. Reachability, branch analysis, what the suite exercises, what the type hints encode. This is ast and coverage.py, not a prompt. A model here is a bug.
  2. GPT-5.6 runs only on an already-located gap. It receives one isolated silence site plus the function source and type contract, and returns structured JSON. It never sees the whole codebase and never chooses what to test.
  3. Never trust, always run. Every case the model proposes is executed. Only cases that genuinely FAIL or ERROR survive into the report. A silence Umbra cannot demonstrate, Umbra does not claim. This execution gate is not optional; it is the honesty of the tool.

If you are about to add a model call, first ask whether structure can answer it. If it can, do that instead.

Read the full file on GitHub · 171 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 3d ago First seen · 171 lines · 2,438 tokens per session scan C f8e8fe452759

Subscribe to this mod's changes

umbra AGENTS.md is an instructions file published in the GitHub repository openhonest/umbra (3 stars, last pushed 21d ago), licensed Apache-2.0. It adds 2,438 tokens to every session, about $0.0122 per session on Opus 5. A static security scan graded it C with 1 finding (recursive force delete). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.

Related

Other instructions, from other repositories

drift copilot-instructions.md

Instructions for mick-gsk/drift, covering drift — verbindliche arbeitsgrundlage für alle agenten, primärmodus bei prompt-engineering, zuerst das richtige primitive wählen, nicht verhandelbare schärferegeln and mindestvertrag für jeden guten prompt.

mick-gsk/drift · 3,275 tokens

drift drift-push-gates.instructions.md

Nutze diese Instruction, wenn ein Commit, ein Push oder eine Release-Vorbereitung im Drift-Repo ansteht. Sie ist die autoritative Gate-Quelle fuer Pre-Push-Anforderungen, Bypaesse und lokale CI-Pflichten.

mick-gsk/drift · 2,980 tokens

drift drift-quality-workflow.instructions.md

Nutze diese Instruction bei nicht-trivialen Aenderungen, Reviews oder Fix-Loops im Drift-Repo. Fokus: adversarialer Gegenreview, iterativer Fix-Loop, expliziter Freigabestatus und Human-in-the-Loop vor Merge.

mick-gsk/drift · 1,887 tokens

drift drift-agent-quickref.instructions.md

Kompakte Task-zu-Gate Kurzanleitung fuer Drift-Agenten. Verwenden fuer schnellen Einstieg bei feat, fix, signal, chore, prompt und review Aufgaben.

mick-gsk/drift · 1,332 tokens

drift drift-rag-grounding.instructions.md

Nutze diese Instruction, wenn ein Coding-Agent Behauptungen ueber drifts eigene Policy, Signale, ADR-Entscheidungen, Audit-Ergebnisse oder Benchmark-Evidence aufstellt. Sie verpflichtet zur Zitation verifizierter Fact-IDs via driftretrieve und driftcite (ADR-091).

mick-gsk/drift · 999 tokens

drift drift-policy.instructions.md

Nutze diese Instruction immer als verbindlichen Gate-Wrapper fuer Drift-Arbeit. Sie enthaelt das Pflicht-Gate, das Kompaktformat fuer Trivialaufgaben und die Verweise auf die autoritative POLICY.md.

mick-gsk/drift · 683 tokens