Overview
Abide (github.com/coldteadotai/abide) from Cold Tea hooks into Claude Code, Codex, and OpenCode to enforce project instruction files. On each edit or turn it sends Jev one typed question per compiled rule with the rule text and diff snippet, never the full chat log. Probabilities above 0.8 trigger an in-session repair message naming the rule and source line; mid-band scores surface notes without blocking. README replay benchmark on 93 Claude Code sessions reports about one in thirteen turns breaking a rule, with Jev checks around 300 ms and roughly a tenth of a cent per turn on measured replays. npm package @coldtea/abide stores keys locally; abide audit judges existing trees for preflight reports.
Problem: AGENTS.md and CLAUDE.md rules are too semantic for linters, so coding agents break style constraints from the first edit.
Built for: Teams on Claude Code, Codex, or OpenCode who want every edit checked against a compiled rubric without paying chat-model prices per hunk.
First indexed on Jev Directory: 2026-09-23
Creator and team
- Name
- Cold Tea
- Handle
- @coldteadotai
How Jev is used
- Role in the product flow
- Per-rule guardrail Noul-style probability on each edit or end-of-turn diff against compiled rubric JSON
- Primitives
- NoulScore
- State in
- Single rule quote plus scoped file diff or aggregated turn diff; conversation history excluded by design.
- Decision out
- Calibrated violation probability per rule ID with banded actions (repair, note, or ignore).
- Compile AGENTS.md, CLAUDE.md, and related files into .abide/rubric.json
- On hook fire, batch one Jev question per applicable rule for the diff
- Threshold probabilities into repair, note, or silent paths
- Append repair instructions to tool results or turn follow-ups for the agent
Abide is the guardrail mirror to Hermes approvals or jev-review dashboards: instead of scoring whole PRs for humans, it sits on the agent hot path and asks cheap boolean-style questions per rule. That only works because Jev returns probabilities, not essays you must parse. Linters still own mechanically checkable rules; Abide skips duplicating ESLint. calibrate and tune commands close the loop when a rule never fires or fires everywhere. Keys stay in ~/.abide/.env or repo-local env files; README states Cold Tea does not receive your diffs on their servers.
Sourced performance claims
- Replay benchmark README cites about 300 ms per check and roughly a tenth of a cent per turn on 93 sessions.Source: github.com/coldteadotai/abide benchmarks/replay
- Measured replay found Jev flagged 39 edits and 15 turns with independent reviewer confirmation on subsets.Source: github.com/coldteadotai/abide README
- Public GitHub repo coldteadotai/abide had about two hundred eleven stars when this listing was drafted.Source: GitHub star count September 2026
Features and stack
Features
- Hooks for Claude Code, Codex apply_patch, and OpenCode plugin mode
- Committed rubric.json mapping rules to instruction line citations
- abide audit and check for batch or pre-commit sweeps
- Replay, calibrate, and tune tooling with JSON output flags
Stack
- TypeScript
- npm CLI
- TypeSafe or Vercel AI Gateway keys
- Agent hook APIs
Pricing: Open source MIT; Jev or gateway usage billed per your key with README-measured sub-cent per edit checks.
Links
FAQ
- Does Abide send my chat log to Jev?
- README states Jev sees the rule and diff only, not the conversation, so late edits get the same scrutiny as early ones.
- Which agents are supported?
- Claude Code settings, Codex hooks (accept the four entries once), and OpenCode plugin installs documented in README tables.
- What happens at 0.86 probability?
- Scores at or above 0.8 trigger a repair message with rule id and instruction quote; 0.5 to 0.8 is note-only per README bands.
Related learn guides
Original Jev guidance that pairs with this product pattern.
- Jev use cases
The patterns builders actually search for: moderation, routing, triage, RAG verify, and agent gates.
- Jev vs LLM classification
When to gate with System One probabilities instead of asking a chat model to label things.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.