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.
npx agentmods add instructions/muhdur/oraclemcp/agents-mdgit clone --depth 1 https://github.com/MuhDur/oraclemcpWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/instructions/muhdur/oraclemcp/agents-md)<a href="https://agentmods.dev/instructions/muhdur/oraclemcp/agents-md"><img src="https://agentmods.dev/badge/instructions/muhdur/oraclemcp/agents-md.svg" alt="Measured on agentmods" height="20"></a>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.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.04176 | $0.04176 |
| Opus 5 | $0.02088 | $0.02088 |
| Sonnet 5 | $0.00835 | $0.00835 |
| Haiku 4.5 | $0.00418 | $0.00418 |
Grade C, and why
oraclemcp 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 4d 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.
deletion, or `rm -rf` on tracked paths without explicit in-session approval. How it starts
The opening of the file, as written. The whole thing — 314 lines — stays where its author put it; the contents beside it link to each section on GitHub.
AGENTS.md - oraclemcp
Operating rules for agents working in this repository.
oraclemcp is an unofficial, engine-free, safe-by-default Oracle Database
MCP server in pure Rust: a small Cargo
workspace of 9 oraclemcp-* library crates plus the oraclemcp binary. Its
optional plsql-intelligence feature embeds the PL/SQL engine for offline
tools; the separate plsql-mcp server is deprecated. Independent open-source
project; not affiliated with Oracle.
RULE 1 - ABSOLUTE
Do not delete any file or directory unless the operator gives the exact command in-session. This includes files you just created. If something should go, stop and ask first.
Irreversible / outward-facing actions
Never run git reset --hard, git clean -fd, git push --force, branch
deletion, or rm -rf on tracked paths without explicit in-session approval.
Never force-push main. Do not commit on the operator's behalf without a clear
in-session go-ahead. crates.io publishes and registry listings are permanent
(versions immutable, names claimed forever); treat publishing as a gated,
deliberate step.
Swarm operating constitution
Eighteen rules. Rules 1-12 were mined from the 2026-07 multi-repo swarm
retrospective (docs/plan/RETRO_SWARM_CAMPAIGN_2026-07.md §3G,
docs/plan/PLAN_ENGINEERING_PROGRAM.md §27.3); rules 13-17 were mined from the
2026-07-21 five-agent session, one per incident it produced (rule 18 from an
incident during the session that encoded the others). Binding on every
agent in this repo, solo or swarmed — most are new; a few name-and-link existing
rules above so the constitution stays the one place to check:
- Never defer planned work on your own initiative — deferral is the operator's call, not an agent's judgment call.
- Green means honestly green; surface red before the operator finds it.
- Claims must be evidence-backed — never assert what you haven't just run and checked.
- Reread this file (and
README.md) until understood, every session, before acting. - Think before acting ("ultrathink"): verify, then execute — don't patch on a hunch.
- Be resource-disciplined: don't trash the host, the disk, or the
token/session budget (
CARGO_BUILD_JOBScaps, scoped-pbuilds over full-workspace ones, no unbounded concurrent compiles). - Keep driving autonomously, but follow explicit operator choices — model, agent freshness, scope — exactly; deviation is the fastest path to anger.
- The fail-closed guard is sacred and tighten-only — see "The safety invariant" below; this rule doesn't restate it, it just makes the constitution complete.
- Confidentiality is absolute: field-test/live-customer identifiers never
leave quarantine (
todelete/, gitignored) or enter a committed artifact. - No surprise costs — cloud resources (OCI, etc.) stay free-tier; a hard rule, not a target.
- Land complete, not sliced across version bumps or half-shipped across sessions.
- Escalate blockers to the operator; delegate unforeseen work to the
tracker (
br create), don't quietly derail the authoritative prompt's scope. - A modified file that is not yours is another agent mid-edit, not a
defect. Check
git statuson a file before declaring it broken, and judge the committed truth (git show HEAD:<path>) before filing a build blocker or going idle on one. - Close evidence comes from a tree verified clean of other agents' work.
Derive the evidence
sourceblock from git rather than asserting it; commit your in-scope work first, and generate a whole-tree reproducibility proof from a dedicated clean worktree at HEAD. - Read the gate verdict yourself; never infer a pass from a successful
push.
git pushreports what the remote accepted, not what the gate decided. A gate that printed a failure is a failure no matter how the push went. - Never block a turn on an unbounded wait. Check once, report, move on. Every wait carries a deadline, and reaching the deadline is a result to report — not a reason to wait again. A blocked turn queues every dispatch behind it.
- A struct field and its initializers are ONE logical change, landed in ONE
commit by ONE agent. The same holds for any edit whose halves do not
compile apart: a
git mvand its references, a trait method and its impls, an enum variant and its exhaustive matches. Split it across panes and you break the build for everyone in the shared checkout. - Commit explicit paths, then verify what landed.
git commit -- <path>...andgit show --stat HEAD; never-a/git add -Ain a shared checkout. A deletion of a path that still exists in the worktree is a stale index snapshot committed over someone else's landed work, not a delete.
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.
- 4d ago First seen · 314 lines · 4,176 tokens per session scan C daf4367461d4
oraclemcp AGENTS.md is an instructions file published in the GitHub repository MuhDur/oraclemcp (3 stars, last pushed 11d ago), licensed Apache-2.0. It adds 4,176 tokens to every session, about $0.0209 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.
Other instructions, from other repositories
seekstone CLAUDE.md
Instructions for shaqmughal/seekstone, covering claude.md, what this repo is, commands, the harness itself (run after npm install) and architecture.
Synapsor-Runner copilot-instructions.md
Copilot instructions for Synapsor/Synapsor-Runner, a project described as: Let AI agents query and update Postgres/MySQL without raw SQL, unrestricted schema access, or database credentials in the model; writes stay reviewed.
mcp-database CLAUDE.md
Claude Code instructions for DiegoBulhoes/mcp-database, covering claude.md, commands, architecture, invariants (do not weaken) and adding or changing a tool.
cortex AGENTS.md
AGENTS.md instructions for dinglebear-ai/cortex, a project described as: Rust MCP/HTTP log aggregation platform for homelab syslog, Docker logs, OTLP ingest, SQLite/FTS search, and AI transcript correlation.
cortex GEMINI.md
Gemini CLI instructions for dinglebear-ai/cortex, a project described as: Rust MCP/HTTP log aggregation platform for homelab syslog, Docker logs, OTLP ingest, SQLite/FTS search, and AI transcript correlation.
mcp-toolbox GEMINI.md
Gemini CLI instructions for googleapis/mcp-toolbox, covering mcp toolbox context & style guide, project overview, tech stack, key directories and development workflow.