koment

Keep the why next to your code — checked, so it can't quietly rot.

The problem

Someone wrote retry 5 times instead of 3, and there was a good reason. Six months later nobody remembers it, so it gets "simplified" back to 3 and the bug returns.

A comment was supposed to prevent that. Comments rot: they sit next to the code without being attached to it, and when the code changes nothing tells you the comment just became a lie. Your AI agent then reads that lie and acts on it.

What koment does

You record the reason outside the source, anchored to the exact lines it explains. Then koment watches those lines. Change them, and it says so:

$ koment check
internal/auth/token.go
  drifted   gotcha        internal/auth/token.go  01KZN63VC5SZASDYJBMPDC03WB
    Without it, clients whose clock runs fast get logged out mid-request. Bit
    us in #412.
1 annotation across 1 file: 1 drifted
koment: 1 annotation no longer resolves; revisit it or update the anchor

Exit code 1. The build stops. Nobody silently inherits a reason that no longer describes the code.

The records are YAML in .koment/, in your git repository. No database, no service, no account. They review in the same pull request as the change that motivated them.

See it running → — koment's own annotations, published by a workflow you can copy.

For AI agents

This is where it earns its keep. Agents write explanatory comments constantly, and they edit code without reading why it was written that way. koment makes both of those fail immediately, while the agent is still working:

• It writes a comment → the edit is refused, with instructions to record the reasoning as an annotation instead.
• It edits annotated code without checking → the turn cannot finish until the annotation is revisited.

One MCP server serves Claude Code, Cursor, Codex, opencode, Hermes, Zed and the rest, so every agent reads the same reasoning through the same interface. Set yours up →

Humans get the same loop in the editor: a squiggle under the comment, and a quick fix offering convert to annotation or keep it, on the record.

Install

brew install koment-dev/tap/koment

mise use -g github:koment-dev/koment          # mise
go install github.com/koment-dev/koment/cmd/koment@latest
docker run --rm -v "$PWD:/repo" -w /repo ghcr.io/koment-dev/koment:3 check

Try it in a minute

From inside any git repository:

koment bootstrap

That sets up .koment/ and wires whichever agents you use. Now record a reason and watch it get checked:

koment add src/auth.go \
  --excerpt 'if token.Expiry.Before(now.Add(-clockSkew)) {' \
  --kind gotcha \
  --title 'Skew subtraction keeps fast clocks logged in' \
  --body 'Without it, clients whose clock runs fast get logged out
          mid-request. Bit us in #412.'

koment check     # ok

Now edit that line and run koment check again. It fails, and tells you which reasoning no longer matches its code. That is the entire product; the rest is delivering it to the people and agents who need it.

koment ui        # read them in a browser
koment show src/auth.go

How it works

1. You annotate a snippet. koment records the prose, exact excerpt, surrounding source context, the commit you were on, and who you are.

2. The record lands in .koment/annotations/<id>.yaml, one file per annotation. Concurrent agents create independent files instead of replacing one shared list.

3. Resolution searches the current file for that excerpt and produces exactly one status:

   	meaning	build
   ok	found where it was last seen	passes
   ambiguous	several contextual candidates remain	fails
   drifted	file exists, the annotated code is gone	fails
   orphaned	the file is gone	fails

4. koment check exits non-zero on ambiguous, drifted or orphaned. That is the whole mechanism: uncertain rationale is worse than no annotation, so it has to be impossible to ignore.

5. When it fails, koment reanchor <id> --excerpt '<new text>' repoints it — keeping its id and creation date, recapturing context and the line for you. Nothing re-attaches automatically; a person confirms the reasoning still holds.

Anchoring is by excerpt, never by line number — line numbers rot on the next edit above them. The commit hash is recorded, but only to reconstruct history; it never decides whether an annotation still applies. Those are two different questions with two different mechanisms.

Local writes change the checkout you are already reviewing. Served writes never touch a replica or push a default branch: they create an exact annotation on a deterministic branch and return only after its pull request exists. Static publications remain immutable and read-only.

Three ways to run it

Pick one. Each is a place to stop, not a step you have to take, and moving between them is not a migration — all three read the same .koment/ in git, so there is nothing to export, import or back up.

Several repositories

One deployment serves many. Identity is assigned independently of provider path, and every refresh resolves a branch to one immutable commit before replacing a repository's active snapshot:

# yaml-language-server: $schema=https://raw.githubusercontent.com/koment-dev/koment/main/schema/server.schema.json
repositories:
  - id: payments
    name: Payments API
    provider: github
    remote: you/payments
    default_branch: main
    default: true
  - id: web
    name: Customer Web
    provider: github
    remote: you/web
    default_branch: main

The service starts only when a non-loopback listener has either trusted-proxy identity or scoped bearer credentials. Private repositories and reviewed writes also require a GitHub token. The Helm chart documentation shows the secret formats and boundary.

Local commands need none of this configuration — koment finds the owning checkout by walking up from the working directory.

Kubernetes

koment publishes an OCI Helm chart to oci://ghcr.io/koment-dev/charts/koment:

helm install koment oci://ghcr.io/koment-dev/charts/koment \
  --set repositories[0].remote=you/your-repo \
  --set metrics.enabled=true \
  --set metrics.serviceMonitor.enabled=true \
  --set metrics.dashboard.enabled=true

The application port authenticates source, rationale, UI and MCP; only liveness and readiness are public. Metrics use a separate listener so an ingress cannot accidentally expose them with application authentication. The chart includes a Grafana dashboard, ServiceMonitor, hardened pod defaults, optional NetworkPolicy and disruption controls, and a digest-pinned helm test probe.

Configuration

Every flag can be set from the environment. --flag-name becomes KOMENT_FLAG_NAME, and an explicit flag always wins.

Git is the only authoritative record. Local reads resolve the YAML directly against the working tree; disposable read models cannot restore or overwrite Git.

koment <command> --help lists every flag alongside its variable.

Work where the code lives

The reference VS Code extension starts koment lsp, renders annotation bodies as virtual inline text, reports drift and prohibited comments as diagnostics, and adds native add, reanchor, convert and explicit acknowledgement actions. The prose is never inserted into the source buffer. Every release attaches a signed VSIX; the same VSIX is published to the VS Code Marketplace and Open VSX when the repository publisher tokens are configured.

Other editors can use the standard hover, diagnostics, code-lens, code-action and execute-command surface from koment lsp without reimplementing storage or anchoring.

Claude Code can install the repository's own project-scoped marketplace plugin:

/plugin marketplace add koment-dev/koment
/plugin install koment@koment-dev

It bundles writable MCP configuration, injects the strict procedure at session start and runs the policy gate before Claude can finish a turn. Install a released koment binary and run koment agents install in the repository first.

OpenCode ships a parallel plugin at plugins/koment/.opencode-plugin/. Add it to opencode.json by Git reference or path, and it installs the same hooks — deny ordinary Go comment intent and run the policy gate on session end. ADR 0126 records the decision to ship both a generated adapter and a plugin directory.

Give it to your agents

koment agents install
koment mcp --write

The generated repository adapters give agents a strict procedure and configure the writable stdio server. It has three read tools:

• koment_get(file, repository?) — annotations for the file an agent is about to edit
• koment_search(query, repository?) — find reasoning by topic; omitting repository searches all of them
• koment_repositories() — what this deployment serves, with counts

Write mode adds four tools:

• koment_add — create agent-attributed rationale
• koment_reanchor — explicitly move an existing anchor
• koment_convert_comment — record a comment as rationale, then remove it
• koment_acknowledge_comment — retain an exceptional comment only after an explicit acknowledgement

Those four mutation tools operate on a local checkout. The authenticated served MCP surface exposes koment_add and returns the deterministic branch, commit and pull-request URL; source-mutating conversion and reanchor stay local so the service cannot overwrite an agent's separate worktree.

Every annotation arrives with its resolution status and its repository, so a stale one is never presented as current and a result is never detached from its scope. When a path exists in several repositories koment_get refuses and names the candidates rather than guessing.

Publish it

Your reviewers are not going to install anything. Give them a URL: one workflow file renders every annotation to static HTML plus annotations.json and search.json, then puts the atomic snapshot on GitHub Pages with no server, database or authentication to design.

- uses: actions/checkout@v7
- uses: koment-dev/koment@v3
- run: koment check
- run: koment site --out dist

Every page names the commit it was rendered from, so a snapshot can never pass for the current tree. koment check in the same workflow means a build that would publish drift fails first.

The whole workflow, ready to copy →

A site renders your source as well as your annotations, so publishing one from a private repository publishes that source. Grouped publications render one snapshot per repository and connect them through the ordinary repository switcher; there is no selector landing page.

Documentation

• Publishing — the copy-paste workflow, and moving to a served instance later
• Bootstrap — what it is, the data model, running it, releases
• Getting started · Writing good annotations · CLI reference · CI

What koment is not

• Not general project documentation. koment holds knowledge bound to a place in the code; broad project guidance belongs in the documentation system your team already uses.
• Not a memory system. A consolidating store paraphrases, merges and eventually forgets. koment is a record, not a belief.
• Not line-precise. Anchors are snippets.

Prior art

Codetations / Magic Markup is the closest existing work — document-external annotations with LLM re-anchoring. The anchoring problem dates to Microsoft Research, 2001 and is still unsolved in the general case. The UI owes its shape to konflate, which makes an invisible layer visible for Flux the way koment tries to for rationale.

License

AGPL-3.0-or-later for the open-source grant. LICENSE is the verbatim licence text and nothing else, so that GitHub and every automated licence scanner identify it correctly; the terms specific to this project are here rather than appended to it.

The complete corresponding source is this repository, https://github.com/koment-dev/koment, which satisfies the AGPL §13 obligation for anyone interacting with a koment server over a network.

Commercial licences are available on request for organisations whose policy excludes AGPL, or that want warranty or indemnification — write to licensing@koment.dev.

Releases ≤ v0.6.0 were distributed under the MIT licence and remain MIT in perpetuity. The last MIT-tagged source may be forked from v0.6.0 under those terms indefinitely. The decision is recorded in ADR 0117.

Does using koment make my code AGPL? No. koment is a tool; annotations it writes are data, not derivative works of the tool. The same legal class applies as compiling a program with GCC: the tool's licence governs the tool, not what you build with it. This is the intent of every OSI-approved AGPL deployment that processes external input; if your organisation's counsel disagrees, the commercial licence resolves it.