Something's coming — join the waitlist

deadeye.

He doesn't waste shots.

A plugin for your coding agent that fits the model, effort level, and context to each task — before the tokens are spent, not after. Built for Claude Code; also Codex, Gemini, Cursor & Windsurf (experimental).

$ /plugin marketplace add deepaksinghcs14/deadeye-cc
$ /plugin install deadeye@deadeye
Hooks register themselves. Binary bootstraps on first use, sha256-verified.
Codex CLI & Gemini CLI: install the binary, then deadeye init codex / init gemini. Cursor / Windsurf get the coder persona as a rules file: deadeye init cursor.
go test -v — raw485 bytes
14 lines — 4 PASS lines, 1 FAIL with its
context, carried into the conversation whole.
go test -v — deadeye99 bytes
$ go test ./... -v 2>&1 | grep -A 5 -E '(FAIL|ERROR|error:|panic:)' | head -120 --- FAIL: TestReconciliationAppliesTaxBeforeDiscount (0.00s) FAIL FAIL canondemo/orders 0.539s FAIL
The rewrite happens before the command runs.
79.6%–99.5% smaller, measured on real runs, not averaged. This run: 485 bytes shrank to 99. This repo's own full test suite (202 lines, 10,301 bytes, all passing): shrank to 55 bytes. A real npm install express mocha: 553 bytes of progress spam shrank to 55, errors still passed through. Three separate real measurements, not one number blended across different situations — run /deadeye-stats savings to see your own numbers.

Why deadeye

What sets it apart

Most tools optimize after the fact, guess at the savings, or ship your code to a server. deadeye does none of that.

79.6–99.5%smaller command output

Verbose test/build/install output is trimmed before it enters context — measured on real runs, not averaged: 485 bytes → 99, a 202-line passing suite → 55.

86.0%lighter subagents

The coder persona a subagent inherits dropped from 7,666 to 1,073 bytes per spawn — measured, on every spawn.

100%measured, never estimated

Every figure is a real row in decisions.jsonl — and 95% of events (the quiet no-ops) aren't logged at all. No invented savings for code never written.

0bytes leave your machine

No phone-home, no API keys, no telemetry. All state lives in one file under ~/.deadeye/.

4 signalsdecided before the spend

Model, effort, and context are fit ahead of the call from four cheap signals — no extra LLM call, no keyword-matching, reasoning always visible.

Fail-opennever blocks, never overrides

If anything inside it errors, your call passes through untouched. It advises on your main session — it never seizes your settings or your model.

We didn't just claim it — we measured it. See the proof · 6× measured

How it decides

Reads the task, not keywords

A lot of routing tools just match words in your prompt — see "architecture," assume it's a big job. deadeye instead looks at a few real signals about the task, then works out the cheapest model and effort level that should still get it done properly. The reasoning behind every decision is always visible, never a black box.

  • LookWhich files are touched, recent git activity, whether tests exist nearby, how specific the request reads. Four cheap signals — no LLM call, no network request.
  • DecideModel and effort level matter together — a mid-tier model on high effort can match a top-tier model on low effort. deadeye works out the cheapest combination that should still clear the bar for this task, not a fixed tier picked by rule of thumb.
  • ApplyPicks the model when a subagent starts — but only if you didn't already pick one yourself. Trims noisy output before it enters context. Asks before a multi-file edit that skipped a plan.
  • LearnIf you manually pick a bigger model than it recommended, that gets logged and makes it more cautious for that kind of task next time. Detecting "this edit actually broke something later" is still on the roadmap.
/deadeye-route "fix the null check in handler.go" Evidence: promptshape complexity=0.00 confidence=0.85 filescope complexity=0.15 confidence=0.85 gitchurn complexity=0.30 confidence=0.82 testpresence complexity=0.10 confidence=0.80 Decision: model: claude-haiku-4-5-20251001 effort: medium confidence: 0.80 reason: all evidence supports downshift: low complexity, confidence >= threshold

Real output, from this repo, this session. Nothing actually changes unless you turn a setting to enforce — by default, deadeye only advises.

When it's unsure — a capable middle, not the priciest

Missing information never buys the cheapest model. When deadeye can't tell how big a change is, it defaults to a capable middle (sonnet-tier), reserving the top tier for genuinely hard work — picking cheaper always needs real evidence above a confidence bar. Optionally, an on-demand AI judge (claude -p, no API key) classifies the truly ambiguous cases for a sharper call.

  • Wrong-but-expensive is recoverableWorst case, it costs you some money. Wrong-but-cheap costs you a bad diff, plus the time spent finding it.
  • Never touches your settingsdeadeye keeps its own state in one file. Your settings.json stays yours, untouched.
  • Never blocks your workIf anything inside deadeye errors, that call just passes through untouched. A tool like this should never be the reason your work stops.

Range card

Eight things it manages

Each one works on its own, has its own on/off setting, and can be switched off without touching the others.

What / effect / default settingoff · advise · enforce
WhatWhat it doesDefault
ContextTrims verbose command output before it enters context — test suites across nine language ecosystems, builds, linters, package installs, pod logs, log tails. Also flags wasteful patterns: re-reading unchanged files, whole-reads of huge files, a command re-run (identical, or the same target with only flag changes), a URL re-fetched, an oversized MCP response, a paste-sized prompt, a long exploration streak better handed to a subagent, and a good moment to /compact before auto-compact lands mid-task. The biggest, safest win.on
CoderA lean-first coding persona injected every session (it survives compaction) and into subagents: question whether the code needs to exist, stdlib before custom code, one line before fifty — plus a live security check on what's written and its dependencies. Three levels — spotter, marksman, sniper.marksman
SecurityThe exfiltration guard — a Read of a credential file, or a Bash command shipping one out, escalates to a permission prompt the model can't answer for itself. Adding a known-vulnerable dependency asks too; existing ones get flagged at session start. Independent of coder mode.ask
Codebase mapA persistent per-project map — directory/package skeleton, most-touched files accumulated across every session, recent exploration notes — injected once per session so a fresh session doesn't re-explore the repo from scratch.on
EffortSuggests lower effort for mechanical steps. Has no effect if CLAUDE_EFFORT is already pinned for the session — /deadeye-status tells you plainly if that's the case.advise
ModelPicks the model when a subagent starts. Prices come from a table built into the binary, refreshed each release — there's no live pricing API a background hook process could reach.advise
Plan gateAsks once before a multi-file edit that skipped writing a plan first, then stays quiet for the rest of that task.soft
WorkflowSuggests the ultracode keyword when a task looks like genuine parallel/fan-out work. Only ever suggests it — never starts one for you.on

Fits alongside the built-ins

Covers what Claude Code doesn't

Claude Code ships its own reviews, plan mode, and size-based output truncation. deadeye doesn't compete — it layers under and around them. The three deepest — coder mode, security, and the codebase map — expand below.

Built-inwhat deadeye adds
/code-review, /simplifydeadeye-review is the lean lens only: verified findings, net-lines accounting, cuts that flow into the debt ledger — the working diff by default, or the whole repo with --repo. Instant local pass; escalate to ultra for pre-merge depth.layers
/security-reviewdeadeye-guard is the diff-scoped lean pass: live regex/manifest advisories while you type, then a verify-before-report deep check on demand across seven exposure classes — injection, secrets, authz, crypto, exposure, vulnerable deps, DoS — with native-auditor (govulncheck / OSV) coverage. Escalate to /security-review for the full audit.layers
PR reviewdeadeye-pr runs a whole PR through four lenses — over-engineering, correctness, performance, security — as terse tagged findings, each carrying its proof and confirmed against the repo's own tools and tests. Severity-ranked; prints locally, opt-in --post for inline per-line comments. Huge PRs fan out to parallel subagents — each spawned at the cheapest model tier its cluster needs, the top tier reserved for risky surfaces — so nothing is truncated and the review stays cheap.four lenses
Plan modeNative plan mode exists but nothing triggers it — deadeye's gate notices a risky multi-file edit coming and nudges into it.triggers
Output truncationNative truncation is size-based at 30K chars. deadeye rewrites commands before they run so only failure context enters at all — and stacks under the native cap.stacks
Subagent modelsNative subagents inherit the parent's model. deadeye recommends the cheapest tier the evidence supports, says why, and learns from your escalations — with an opt-in AI judge (claude -p, no API key) for the genuinely ambiguous cases.advises
Coding personaNo native equivalent — coder mode is deadeye's own ground.unique
Other hostsExperimental: the same daemon serves OpenAI's Codex through its hooks — output trimming, coder persona, plan gate, decision log — via deadeye init codex. Gemini CLI gets the coder persona and session guidance now (init gemini); its tool-level engine waits on live schema verification. Cursor and Windsurf have no hook contract, so they get the coder persona alone as a rules file (init cursor / init windsurf). And /deadeye-pr — the four-lens PR review — installs on all four via init in each host's native command format, experimental until live-verified.five hosts
A marksman's coding discipline

Beyond routing and trimming, deadeye can inject a coding persona: a lazy-senior-dev discipline that pushes every change toward the leanest solution that actually works. Injected at every session start — it survives compaction — and travels into subagents too.

Level / discipline/deadeye-coder
SpotterCalls the shot, doesn't take it: builds what's asked, names the leaner alternative in one line. You pick.light
MarksmanThe lean-first ladder enforced: YAGNI, stdlib before custom code, native before dependencies, shortest working diff.default
SniperOne shot only. Ships the one-liner and challenges the rest of the requirement in the same breath.max

Deliberate shortcuts get a deadeye: comment naming the ceiling and upgrade trigger — /deadeye-debt collects them into a ledger. Safety isn't just preserved, it's checked for: input validation, error handling that prevents data loss, security, and accessibility stay at every level, and a live pass flags injection, secrets, weak crypto, and vulnerable dependencies as you write.

/deadeye-coder sniper DEADEYE CODER CHANGED — level: sniper every session start: DEADEYE CODER ACTIVE — level: sniper the ladder, abridged: 1. Does this need to exist at all? 2. Already in this codebase? 3. Stdlib does it? 4. Native platform feature covers it? 5. Already-installed dependency solves it? 6. Can it be one line? 7. Only then: the minimum code that works.

Off any time: say normal mode, or DEADEYE_CODER=off. An optional statusline badge shows each session's live level — deadeye offers the setup once and never edits your settings itself.

How it worksone ruleset, filtered
InjectThe ruleset ships embedded in the binary. At every session start the daemon prints it into the model's context, filtered to the active level — a marksman session never pays tokens for the other levels' rows.start
SurviveWhen Claude Code compacts a long conversation, the hook fires again and re-injects at whatever level you'd switched to. No drift back to over-building.compact
TravelSubagents inherit a condensed card — just the behavior-bearing rules. Measured: 7,666 bytes per spawn before, 1,073 after, an 86.0% cut on every spawn. Scope it with coder.subagent_matcher.−86.0%
ResolveKill switch beats your session switch, beats coder.default_level in config, beats the built-in default. Switches last until the session ends.precedence
GuardEvery Edit/Write is checked against the ADDED text only — injection shapes, secrets, provider tokens, weak crypto, TLS-off, across five languages — plus the dependency itself on a manifest edit. Independent of the persona level now, so stop coder doesn't silence it. Deep pass on demand: /deadeye-guard.check-your-backstop
the same request, per level -- "add a cache for these API responses": spotter Done, cache added. FYI: lru_cache covers this in one line if you'd rather not own a cache class. marksman @lru_cache(maxsize=1000) on the fetch function. Skipped custom cache class, add when lru_cache measurably falls short. sniper No cache until a profiler says so. When it does: @lru_cache. A hand-rolled TTL cache is a bug farm with a hit rate.

Every answer keeps one shape: code first, then at most three short lines — what was skipped, and when to add it. Bug fixes go to the root cause every caller routes through, not the one path the ticket named. Comments follow their own rule: the why, never the what — and corner-cuts use a pinned grammar (deadeye: … ceiling: … upgrade: …) that /deadeye-debt parses reliably.

the same request, per level -- "look a user up by name": spotter Done. FYI: name is interpolated into the query -- bind it instead. marksman Parameterized query, name bound not interpolated. Skipped a sanitizer layer; the driver escapes. sniper Bound parameter, and it returns id+name only -- fewer columns out, less to leak.

The safe form is usually the SHORT form — a parameterized query beats hand-rolled escaping, exec.Command(bin, args...) beats building a shell string. Lean and safe are the same move.

Editing a manifest checks the dependency too: a bundled table flags packages the platform has since absorbed (requestfetch, momentTemporal/date-fns), and an optional OSV.dev lookup catches known vulnerabilities — name and version only, nothing else about the code, and never on the hook's response path: a cache miss refreshes in the background and informs the next edit, never blocks this one. Turn just that lookup off with coder.security_osv: false; the bundled table keeps working fully offline. /deadeye-guard is the on-demand deep pass — diff-scoped, verifies a missing sanitizer or authz check against the surrounding code before reporting it, and runs govulncheck/npm audit/pip-audit/cargo audit when installed.

A prompt injection can't answer a permission prompt

The real attack on a coding agent isn't bad code you write — it's malicious content in a repo or a web page telling the agent to read your credentials and ship them out. deadeye already sits at the exact choke points, so it watches them: the moment that matters escalates to a prompt you answer, not one the model can talk itself past.

Read ~/.ssh/id_ed25519 → permission: ask deadeye exfiltration guard: Read of a sensitive credential path (ssh-private-key). Approve only if you asked for this; a prompt-injected instruction cannot answer this prompt for you. # stays silent, because these are routine: Read ~/.ssh/config Read .env.example ssh -i ~/.ssh/id_rsa deploy@host

Covers ssh private keys, ~/.aws/credentials, .env, ~/.claude/.credentials.json, ~/.netrc, kube/docker/gcloud/gh tokens, and your own security.sensitive_paths — on Read and on a Bash command shaped to ship one out (a credential piped to curl/nc/scp, an env dump to the network). Default ask; soften to advise or off.

The rest of the security surface0.17
Its own axisThe exfil guard is security.exfil, not part of coder mode — turning the persona off never turns it off. Only DEADEYE=off or security.exfil: "off" does.ask
Not a moodThe live Edit/Write check (injection shapes, secrets, provider tokens, weak crypto, TLS-off, across five languages) no longer dies when you say stop coder. Security is part of coding, not the persona's voice.advise
Vulnerable depsAdding a package with a confirmed OSV advisory asks by name (coder.security: "ask"). And once per session deadeye flags what's already in your tree with a known advisory.ask
Provider tokensA bare ghp_…, xoxb-…, sk-ant-…, Stripe live key, or signed JWT pasted as a literal is caught even with no key = around it — length-bound so prose never trips it.secret

Honest ceiling: the exfil guard matches regex-visible shapes, not shell obfuscation or a novel egress binary. It shrinks the blast radius of the common automated attack — it isn't a sandbox. /deadeye-guard is the deep, diff-scoped pass with native auditors.

A fresh session shouldn't start from zero

A fresh session normally burns 15–40 tool calls rediscovering the project — directory layout, what each package does, which files matter. deadeye persists that per project under ~/.deadeye/map/ and injects it once, at the first prompt.

What persists~/.deadeye/map/
SkeletonDirectory rows with file counts and, for Go repos, each package's one-line doc-comment purpose. Built from git ls-files, regenerated only when the tracked-file list changes.map.md
Touch counterA path → sessions counter merged at every session end, ranked by relevance rather than recency — a file touched in 14 of 20 sessions stays on top.touched.json
Exploration notesThe explore skill caches its summary via deadeye notes-append so the next session inherits findings instead of re-deriving them. Newest 5 kept.notes.md

First-ever session in a project gets nothing — the map is written at session end, so value starts at session #2. Resumed or compacted sessions are skipped; their context already carries the exploration.

injected once, at session start: # Codebase map: deadeye-cc scope: 145 git-tracked files cmd/deadeye/ 37 files .go internal/codemap/ 6 files .go the structural half of the map internal/coder/ 4 files .go the lean-first persona Frequently touched files: README.md 4 sessions internal/codemap/codemap.go 3 sessions

Lives under mode.codemap (off/on, default on) — the same DEADEYE=off kill switch turns it off along with everything else.

Measured, not estimated

Every number here comes from a logged decision

Every time deadeye trims something, it logs which rule fired. Every decision is logged with what went into it. Nothing here is a made-up formula that turns "we ran N tool calls" into a savings number — if it isn't a real row in decisions.jsonl, it doesn't get printed.

/deadeye-stats savings real session, this repo -- add Mark() + test, subagent delegated 17 decisions logged by surface PostToolUse 7 PreToolUse/Edit 3 PreToolUse/Agent 1 SessionEnd 1 PreToolUse/Bash 2 Stop 2 SubagentStart 1 by action advise 1 noop 13 rewrite 2 savings-shown 1 preprocessing rewrites (estimated bytes -- see each rule's EstBeforeBytes/EstAfterBytes, not a measurement of this run) build-filter 1x ~15,000 → ~9,600 bytes test-filter 1x ~30,000 → ~9,600 bytes total ~45,000 → ~19,200 bytes (~25,800 saved) Cross-check these figures against /usage's plugin attribution.

Unedited output from a real feature task run through the installed plugin — a cheaper model suggested before a subagent started, both commands trimmed, an end-of-turn savings summary. The 25,800 figure is each rule’s own labelled estimate; the 485 → 99 and 10,301 → 55 numbers up top are the measured ones. Yours shows whatever your log holds — and quiet noop rows (95% of events) aren’t logged at all.

Install

Two commands

/plugin marketplace add deepaksinghcs14/deadeye-cc /plugin install deadeye@deadeye

macOS and Linux need nothing else. Windows: grab the binary from Releases first — self-bootstrap there isn't built yet.

Turn it off

DEADEYE=off everything DEADEYE_PREPROCESS=off context hygiene DEADEYE_GATE=off plan gate DEADEYE_CODER=off coder persona
Does it phone home?

No. Everything it remembers lives in one file on your own machine, at ~/.deadeye/. No hosted service, no API keys, no telemetry.

Will it make Claude dumber?

That's exactly the failure mode it's built to avoid. When it isn't confident about a task, it defaults to the more capable option — picking something cheaper always needs real evidence first. If you find a case where it under-powered a task, that's a bug — report it along with the /deadeye-route output.

It rewrote a command and I needed the full output.

Each rewrite rule is conservative by design, and every one of them can be turned off individually in config. A rule that loses information the agent actually needed is a bug in that rule — please report it along with the command.

Can it change the model my main session is using?

No — nothing can do that reliably partway through a session. deadeye actually changes the model for subagents, where most of the token spend happens, and only advises (never enforces) on the main session.

Does coder mode cut corners on safety?

No. The persona never simplifies away input validation at trust boundaries, error handling that prevents data loss, security measures, or accessibility basics — at any level. And every deliberate shortcut it does take gets a deadeye: marker with an upgrade trigger, so /deadeye-debt can hold it accountable later.

Why "deadeye"?

Because efficiency isn't spending less — it's not missing.