Make your AI coder work like a senior developer. One Go binary gives
the agent a commit gate it cannot talk its way past, quality controllers
that refuse to call unfinished work done, and a self-learning loop that
turns every escaped bug into a permanently closed class. The agent stays
in control — nothing ever touches your code behind its back.
Quick start
/plugin marketplace add azrtydxb/procoder
/plugin install procoder
/procoder:init # installs the tools this repository needs
That's Claude Code; Procoder also ships adapters for every agent —
Cursor, Windsurf, Cline, Kilo Code, Roo, Kiro, Codex CLI, Copilot CLI,
Gemini, OpenCode, and anything that reads AGENTS.md. See
Every agent.
Before / after
Without Procoder, "done" is whatever the agent last said. With it, done
has to survive the gate — this is a real run, verbatim, on a demo repo
with an unformatted file, a conflict marker, and a staged junk file:
$ procoder check
unformatted main.go (run `procoder format "main.go"` for the result)
BLOCKING notes.md:1 merge conflict marker left in the file
BLOCKING notes.md:5 merge conflict marker left in the file
BLOCKING .DS_Store junk file staged — caches and garbage never belong in a commit
...
procoder gate: 0 clean, 2 unformatted, 0 unchecked, 1 out of scope, 8 hygiene finding(s) (3 blocking)
The agent gets the findings and the fixed content in the same turn; the
binary never edits a file itself. And a tool that failed is never
reported as clean — "unchecked" counts as failing, said out loud.
How it works
agent writes code ──► hook fires ──► binary computes findings
▲ │
└──── agent reviews and implements ◄─┘
Two modes, one principle:
- Self-serve — tools the agent runs itself: check, format, lint,
scan, query the index, audit the tree.
- Forced — hooks at fixed lifecycle points (every write, every
session start) the agent cannot skip.
- P-CONTROL — the binary computes and reports; the agent acts.
Nothing modifies code, files, or state behind the agent's back. The one
exception is
procoder prune --apply, which removes superseded copies of
the plugin from your cache — not repository content, and never without
being asked twice.
What's inside
The commit gate (procoder check) — formatting across the popular
languages (Go, Python, JS/TS/HTML/CSS, PHP, Rust, C/C++, Java, Kotlin,
Swift, Ruby, Dart, C#, shell — one canonical formatter each, the
project's config always wins), git hygiene (conflict markers, junk, oversized
files, AI-attribution lines), secrets, lint, ci and infra hygiene, and
documentation health, all through one code path so check, git, and
CI can never disagree.
That full set is what a repository that has adopted procoder gets — one
with a .procoder/ directory, or an AGENTS.md naming procoder. In
somebody else's repository only the checks that are true anywhere run
(secrets, oversized files, conflict markers, junk, AI-attribution lines),
and the ones that read file content see only the lines your commit wrote.
procoder does not apply its conventions to a project that never asked for
them, and every run says which mode it was in.
The quality chain — thinking before code, with a refusing controller
at every link: a spec interview closes design gaps before anything is
built (spec check blocks while sections are missing or questions
open), a plan turns the spec into tasks an engineer with zero
context could execute (plan check blocks placeholders), a backlog
holds larger projects as milestones → epics → user stories seeded from
specs and worked in scope-boxed sprints (one active sprint, explicit
carry-over, closes that refuse), and a todo list tracks standalone
work (todo close refuses until every acceptance criterion is checked,
evidence is recorded, and the gate is clean — story closes carry the
same rigor).
The self-learning loop — a pre-PR self-review catches
reviewer-class findings before a PR exists; anything that still escapes
becomes a lessons ledger entry whose adaptation (a linter rule, a
rubric line, a pinning test) must land before the work counts as done.
Downstream bot reviewers are the fallback net, not the net — and what
that net catches is not lost either: procoder copilot-leak collects
GitHub Copilot's auto-review findings, strips every trace of your code
from them, and — only after you say yes on a terminal — files them as
issues and records them as unlearned until someone writes the
adaptation that closes the class.
The test domain — procoder test runs the repository's real suite
with each ecosystem's canonical runner (go test, cargo test, the
package.json test script, pytest, gradle/maven): PASS with counts, FAIL
with the failing tests named, and NOT run, which is never the same
as green. Coverage is reported where the runner
measures it natively and never enforced. Set [test] policy = "block"
and a green suite becomes part of "done" — the closes refuse while it is
red or unverifiable.
The release controller — procoder release is the last refusal
before a tag: the version in sync across every file you list, the
changelog entry present, the tree clean, the gate clean, the suite
green. Every failure arrives in one list, and on success the git tag
command is printed for you to run. Procoder tags nothing itself.
The ten domains — security (gitleaks, semgrep, osv-scanner), best
practices (lint with curated baselines), maintainability
(dead code, complexity, plus deps freshness per ecosystem),
performance (bench against a saved baseline — Go only, said out
loud), documentation (broken refs, drift, diagrams, badges, adr
decision records, this very README's completeness), clean code
(formatting), testing, ci (pinned actions, timeouts), infra
(Docker, Terraform, Kubernetes, Helm), and GitOps discipline.
The code index — ctags + SCIP, the agent's fast map: find, refs,
callers, impact, unused, entrypoints.
Senior habits, encoded — engineering principles injected at
session start (build ladder: reuse → stdlib → platform → minimum code;
delegation: parallel subagents under a contract, watched and judged;
ADHD/ASD-friendly formatting for complex answers — problem cards,
decisions surfaced, noise filtered; all repo-overridable), deliberate
corner-cuts marked and harvested as
debt with revisit triggers, and an audit command that onboards
any existing codebase with a triaged scorecard.
Replaces the plugins you are already running
Procoder absorbed three earlier tools, and running them alongside it
means two sets of instructions competing for the same agent:
- superpowers —
plans for an engineer with zero context, spike/bounded/architectural
classification, evidence before "done", systematic debugging, TDD with
the mutation check. All of it here, with controllers that refuse where
the originals advised.
- ponytail — the build
ladder, the
debt: marker convention, the five-tag over-engineering
review, one instruction file serving every agent.
- serena — symbol-level
navigation, cross-file rename, interface implementations, project
memory. Now
procoder index and .procoder/, with no MCP server to
keep running. Serena's symbol-level write tools are deliberately
not adopted: the binary computes the rename and hands you the diff.
Full provenance map, including where the serena replacement stops:
Influences.
Configuration
Everything Procoder owns lives in .procoder/ — plain files, made to be
edited, and the repo's version always wins over the built-in default:
config.toml (policies, thresholds), PRINCIPLES.md, the github
templates, the docs/security rules, the review rubric, the lessons
ledger. Full reference:
Configuration.
The docs
The full story lives on the site, organised the way the
Divio documentation system
splits it: the tutorial
(Getting started),
the how-to guides
(Ship a change),
the reference
(every command), and
the explanation
(the quality chain,
how it's built).
What the reports mean
A file that could not be checked is never called clean. A task without
fresh verification evidence cannot close. This README is held to a
completeness check — a feature family it stops mentioning blocks the
gate. No benchmark numbers appear here because none have been run; any
future number will carry its method alongside it.
Where the rigor costs more than it returns, and where a check is narrower
than its name, is written down rather than left to be discovered:
Honest limits. What layer procoder occupies, and
what it does not try to be: Where procoder sits.
The external evidence behind each premise, and the premises that have none,
are separated and labelled in Research.
Implementation
One Go binary, no runtime dependencies, cross-compiled per platform into
dist/ and committed with the plugin — no npm, no network at hook time,
air-gapped included. go test ./... to develop; the design contract
lives in the docs and supersedes anything here that drifts from it.
给 AI 编程智能体加的一层纪律：一个 Go 二进制文件，为智能体提供一道它无法靠花言巧语绕过的提交闸门，配上会拒绝把未完成工作称为完成的质量控制器，还有一个针对漏网 bug 的自我学习闭环。