deadeye is a plugin for coding agents that selects a suitable model and effort level for each task, while trimming verbose command output before it enters context. It is built for Claude Code and also supports Codex CLI, Gemini CLI, Cursor, and Windsurf on an experimental basis.
deadeye is a plugin for coding agents that fits model choice, effort level, and context handling to the task before tokens are spent. It is built for Claude Code, with support for Codex CLI and experimental support for Gemini CLI, Cursor, and Windsurf.
The plugin analyzes six task signals, recommends a model-and-effort combination, and shows the reasoning behind the decision. It also reduces noisy command output before it enters the agent’s context, adds a lean-first coder persona to sessions and subagents, and keeps its state locally. By default, deadeye advises rather than taking control; it does not override a model selected by the user, and failures pass through without blocking the coding-agent call.
deadeye evaluates files touched, recent Git activity, nearby tests, task wording, task specificity, and read-only status. It uses these six signals to recommend a model and effort level together, with visible evidence and confidence rather than a keyword-only rule.
When the six local signals cannot place a task confidently, an AI judge can classify its description through the user’s own Claude login. The judge is enabled by default, can be turned off with a configuration flag, and does not receive the user’s code.
Verbose output from tests, builds, linters, package installs, pod logs, and log tails is rewritten before entering context. Errors and relevant failure context are retained, while repeated commands, unchanged-file rereads, oversized responses, and other wasteful patterns can be flagged.
A coder persona is injected into the main session and subagents and survives compaction. It promotes questioning whether code is necessary, using the standard library before custom code, and avoiding unnecessary implementation.
deadeye can advise on routing, output, and other managed behaviors, with separate on/off settings and advise/enforce modes. It only selects a subagent model when the user has not already selected one, and it fails open if an internal component errors.
Hooks register themselves and the binary bootstraps on first use with SHA-256 verification. State is kept in one file under `~/.deadeye/`; the site states that no code, telemetry, API keys, or third-party service are required.
Use deadeye when a test command produces lengthy output. It can reduce the amount entering the agent’s context while carrying failures and their surrounding context forward, helping the agent focus on the actionable result.
For a targeted fix such as a null check, deadeye can inspect repository and task signals and recommend a lower-cost model and suitable effort when the evidence supports a downshift. The recommendation remains advisory unless enforcement is enabled.
Use the injected coder persona for subagents that need a compact, lean-first instruction set. The site reports a measured reduction in persona size per spawn, and the persona is designed to persist through compaction.
When local evidence is insufficient to classify a task, the optional AI judge can use the task description to make a sharper routing decision. This is intended for genuinely ambiguous cases rather than every request.
It is built for Claude Code. The site also lists Codex CLI, Gemini CLI, Cursor, and Windsurf; Gemini, Cursor, and Windsurf are described as experimental.
By default, deadeye advises and shows its reasoning. It only picks a subagent model when one was not already selected by the user, and the site says it does not touch the user’s settings.json. Separate behaviors can be configured for advise or enforce modes.
The site states that no code is sent anywhere and that there is no telemetry, API key requirement, or third-party service. For the optional AI judge, an ambiguous task’s description is sent through the user’s own Claude login; the site says the code is not sent.
For Claude Code, the site shows `$ /plugin marketplace add deepaksinghcs14/deadeye-cc` followed by `$ /plugin install deadeye@deadeye`. Hooks register themselves and the binary bootstraps on first use with SHA-256 verification. Other listed agents use their respective `deadeye init` commands or a generated rules file.
deadeye is designed to fail open. If something inside the plugin errors, the coding-agent call passes through untouched rather than being blocked.
www.edgee.ai
Edgee is an agent gateway for coding teams that reduces token usage, routes requests across models, and provides usage visibility. Its Compression V2 works as a drop-in CLI layer for Claude Code, Codex, OpenCode, Cursor, and other supported agents.
harproject.dev
HAR HQ is a team governance and observability layer for organizations running coding agents with HAR. It standardizes verification, tracks AI-related work and spend, and helps engineering teams review and improve agent workflows.
gmicloud.ai
AI infrastructure for production inference, training, and fine-tuning on NVIDIA GPUs.
learn.microsoft.com
Semantic Kernel is Microsoft’s public GitHub software project for integrating large language model technology into applications. The repository includes implementation and documentation directories for .NET, Python, and Java development.
onecli.sh
OneCLI is an open-source platform for giving employees personal AI agents that can work across company tools. It provides isolated sandboxes, gateway-enforced policies, scoped credentials, approvals, and audit logs for team use.
fireworks.ai
Generative AI platform for serving, training, fine-tuning, and deploying open-source models
