amyrmahdy/decima-agent-decisions

Decima Agent Decisions

About 400k labelled decisions that a coding agent's hooks make on every step, plus a hand-written test set of 130. The questions are the small, frequent ones:

• Is a real secret about to be written to disk?
• Should this shell command run, ask the user, or be denied?
• Is this command read-only?
• Which tool, and which exact command, fits this step?
• What should happen next: run, look first, test first, hand to a small model, or escalate?
• Which model tier does this task need?

This is the training data of decima-agent, a 321M local decision model for Claude Code-style hooks. It is released so others can train, audit or beat it.

Built by A. M. Madani (@amyrmahdy). Code: github.com/amyrmahdy/decima.

Configs

config	train	test	what the state is → what is decided
secret	126,111	2,111	a Write/Edit (file, new text, often a diff in a long agent context) → is a live credential written?
bash	77,867	2,513	a proposed shell command, with a policy → allow / ask / deny, and "does it exfiltrate / delete / need root?"
readonly	14,351	600	a command → does it leave every file, setting and remote unchanged?
tool	26,247	1,503	a goal → Read, Glob, Grep, Edit, Write, Bash, WebFetch, WebSearch, Task, TodoWrite, AskUserQuestion or an MCP tool
command	29,467	1,637	a goal → the exact command among near misses (git reset --soft vs --hard)
next_step	74,581	2,258	a trajectory or a proposed step → the next action
router	11,442	768	a coding task → haiku / sonnet / opus-class tier, or a non-coding model
outcome	39,035	1,205	test output or an ops alert → did the tests pass? does a human need to step in?
agentbench	—	130	hand-written, held out from all of the above (see below)

About a third of the states are wrapped in a long agent context (recent turns, a file list, the last diff), so the decisive line has to be found in 0.5–2k tokens.

Columns

column
id	stable row id
task	finer task name inside the config (secret, secret-llm, bash-llm, tool-para, …)
origin	procedural (exact label from code), llm-file+code-slot, llm-paraphrase, llm+blind-check (see below)
kind	choose (one of N), verify (yes/no) or score (ordered levels)
state	the situation, usually the JSON a PreToolUse hook receives
question, choices	the decision, as asked
gold	index of the right choice
probs	soft target (most mass on gold), the distribution the model was trained on

agentbench uses the jabr/classifier-benchmark / TypeSafe shape instead: type (noul or choice), instructions, criteria (JSON), state, choices, expected, gold.

from datasets import load_dataset
ds = load_dataset("amyrmahdy/decima-agent-decisions", "bash")
print(ds["test"][0])

How the labels were made

Every agent label is exact or agreed by two different models:

• procedural: exact, from code.
  • Real-format credentials (AWS, GitHub, Stripe, Slack, GitLab, npm, Hugging Face, JWT, connection strings, …) vs placeholders, env lookups, docs example keys, hashes, UUIDs, public keys, test fixtures and base64 images.
  • A policy-labelled command catalogue: read-only, build/test, recoverable change, destructive, exfiltration, root, pipe-to-shell.
  • Tool and command choices with near-miss distractors; next actions from rules over the situation; test-run and ops-alert outcomes.
• llm-file+code-slot. Qwen3-Coder-Next writes a realistic file with a <<VALUE>> slot; code fills it with a real or fake credential, so the label stays exact.
• llm+blind-check. Qwen3-Coder-Next writes realistic commands, tasks and steps with an intended label; Gemma 4 (26B-A4B) labels each one blind, and only rows where both agree are kept.
• llm-paraphrase. Gemma 4 rewrites a procedural goal in natural words; rows that lose a required entity (a file name, a flag) are dropped.

All fake credentials are random strings in real formats; none is a working key. Invented e-mail addresses at real providers were rewritten to reserved .example domains.

Test splits are harder than train on purpose

The test splits hold out whole families rather than random rows: credential formats (Google, SendGrid, Twilio, basic-auth URLs), fake-value types (certificates), file types (Terraform, INI), command families, catalogue entries, tools and phrasings, and LLM domains (terraform, helm, gcloud, rsync, openssl). Train rows whose state also appears in test were removed. Test accuracy therefore measures generalisation to things never seen in training.

agentbench: 130 hand-written cases

Written separately from every generator, in the states a real hook sends: secret gate 20, bash gate 22, read-only 8, next action 14, tier router 15, tests passed 8, ops intervention 10, tool choice 12, command choice 15, next step 6. Credential-shaped values in it are fake. decima-agent scores 0.90 on it, decima-base (no agent training) 0.58.

Limitations

• English only. The states are English code, commands and prose.
• Synthetic. These are not real agent logs. Real commands are messier (cd <dir> && python -c "…"); logs from your own hooks are the best data for tuning a gate to your project.
• The policy is ours. "Ask" covers network calls, installs and anything outside the workspace; your project may want local services allowed. Labels follow the policy text in each question.
• Labels from llm+blind-check rows are only as good as two models agreeing; the procedural rows are exact.

License

CC BY 4.0. The writing models are Apache-2.0 (Qwen3-Coder-Next; Gemma 4 26B-A4B). Please cite:

@software{madani2026decima,
  author = {Madani, Amir Mahdi},
  title  = {Decima: open, calibrated decision models},
  year   = {2026},
  url    = {https://github.com/amyrmahdy/decima}
}