# pragent

An AI pull-request reviewer for Gitea that posts **inline comments with suggested
fixes**, not a wall of prose — and reports what each review cost.

A webhook wakes for any PR on a repo whose default branch carries a
`.pr-review.json` with `"enabled": true`. The service checks the repo out at the
PR's head commit, reads the changed files *and the code around them*, runs the
repo's own linters, and posts a review anchored to real lines.

**See it work:** [`pragent-demo` PR #1](../pragent-demo/pulls/1) — a PR with
planted defects, and the review it drew: 9 findings, 3 critical, all anchored
inline.

```
**[CRITICAL]** search_notes builds its SQL by string concatenation: owner and
term come from request['query'] and are spliced directly, so a term like
`' OR 1=1 --` reads every row in the table.

Fix: Parameterise owner and term with placeholders and a real LIKE pattern.

🪙 ~362 tok (11% · attributed output)
```

## Why this exists

CodeRabbit, Greptile and Qodo are good products with fixed review dimensions and
per-seat pricing. `pragent` targets the case where a platform team needs to **add
its own dimensions** — an internal compliance rule, a service-catalog ownership
check, a house performance idiom — without forking a vendor's reviewer. The
analyzers, the data, and the analytics are yours.

It is also self-hosted end to end: the model endpoint is a config value, so the
code never has to leave your network.

## Status

A **pilot** is live and reviewing real PRs. The current runtime architecture is
documented in [`docs/architecture.md`](docs/architecture.md); older framework
plans remain in [`docs/plans/`](docs/plans/) as historical design material.

What works today:

- a central webhook service, so onboarding a repo is *add the bot + commit
  `.pr-review.json:enabled = true`*
- whole-repo context: the reviewer reads callers and types, not just the hunk
- inline comments with language-highlighted suggested fixes, anchored to
  post-change lines and validated in Python before posting
- per-commit dedupe, and prior reviews fed back so a re-push synthesises rather
  than repeats
- `.pr-review.json` for per-repo focus and house rules (also the opt-in flag)
- token-usage reporting on every review, measured from opencode `step_finish`
  events
- Langfuse traces, equivalent-cost reporting, evaluation scores, and feedback
  harvesting
- containment against hostile PR content (see [Security](#security))

Not yet: status checks, fail-close, attention tiering enforced in code (it is
currently a skill the agent follows), multi-model routing, and a CLI framework.

## How a review runs

```
PR opened on repo with `.pr-review.json:enabled = true`
   │  Gitea webhook (HMAC-verified, body-capped, concurrency-bounded)
   ▼
review_pr()
   1. opt-in    .pr-review.json:enabled=true on base? if not, skip.
   2. dedupe    already reviewed this exact sha? stop.
   3. fetch     diff + .pr-review.json from the BASE branch
   4. checkout  repo archive at head sha → temp workdir
   5. sanitize  delete author-controlled agent-instruction files
   6. brief     .pragent/brief.md, untrusted parts explicitly fenced
   7. review    opencode agent: read code, run linters, emit findings JSON
   8. anchor    validate every line against the diff's post-change lines
   9. post      inline comments + summary, as pragent-bot
```

Steps 1, 3, 8 and 9 are deterministic Python. The model's only job is step 7 —
producing correct findings. It never talks to Gitea, and a finding whose line
does not validate becomes a summary bullet rather than a misplaced comment.

## Setup

Onboarding a repo, once the service is running for that owner:

1. add `pragent-bot` as a **Write** collaborator
2. commit `.pr-review.json: {"enabled": true}` to the repo's default branch
3. open a PR

Standing up the service itself — the webhook, the image, the Gitea SSRF
allow-list, the per-owner webhook registration — is in
[`pilot/README-webhook.md`](pilot/README-webhook.md). A legacy per-repo CI-step
path is in [`pilot/README.md`](pilot/README.md).

The model endpoint is supplied at runtime via `PRAGENT_MODEL_BASE_URL`; the
committed `opencode.json` carries a placeholder.

Per-review token spend, latency, equivalent cost, and evaluation scores are
shipped to a self-hosted Langfuse: [`pilot/README-langfuse.md`](pilot/README-langfuse.md).
Emission is a silent no-op unless `LANGFUSE_HOST` and the key pair are set.

## Extending it

The review "factory" is [`.opencode/`](.opencode/README.md) — agent definitions
and skills as plain Markdown. Adding a review dimension is dropping a file in,
not writing code:

| Add | How |
|---|---|
| A review lens | `.opencode/agents/<name>.md` + one allow-list line in `pragent.md` |
| Domain knowledge | `.opencode/skills/<name>/SKILL.md`, referenced from the load table |
| Per-repo rules | `.pr-review.json` in the repo being reviewed |

Shipped skills: `attention-tiering` (the cost governor), `review-methodology`,
`findings-schema`, `linter-playbook`, `security-lens`, `malicious-change`,
`comment-craft`.

## Security

The reviewer runs an autonomous agent with shell access over a checkout of **the
PR author's branch**, and its bot account holds a Write credential. Anyone who
can open a PR can therefore put arbitrary text in front of the model and
arbitrary files on its disk — the setup exploited in the [April 2026 disclosures
against Claude Code Security Review, Gemini CLI Action and Copilot Agent][csa].

Four controls, none of which rely on the model behaving:

1. **No credentials in the agent's environment.** The subprocess environment is
   built from an allow-list, not inherited. There is nothing to exfiltrate.
2. **No author-controlled instruction files on disk.** Nested `AGENTS.md`,
   `CLAUDE.md`, `.cursorrules`, a repo `opencode.json` — all deleted before the
   agent starts, so a PR cannot ship its own system prompt. They are still
   *reviewed*, as data.
3. **Untrusted-data framing.** PR text and diffs are fenced; the agent reports
   injection attempts as `critical` findings instead of following them.
4. **Reviewer config comes from the base branch**, so a PR cannot rewrite the
   rules it is judged by.

Plus: tar-slip guards on the archive, a non-root container, and bounded
concurrency. Full threat model and residual risks: `pilot/README-webhook.md`.

[csa]: https://labs.cloudsecurityalliance.org/research/csa-research-note-comment-control-github-prompt-injection-20/

## What it costs

The pilot runs against a self-hosted model and bills nothing per token, but the
token *work* is real. `pilot/cost_model.py` prices it against published API
rates, calibrated against runs measured through the usage telemetry
(`OBSERVED_RUNS` in that file — append to it, don't guess). Tokens are summed
from opencode `step_finish` events per review.

Each review is governed by hard limits: 20 completed steps, 120,000 total
tokens, 20,000 output tokens, and 480 seconds by default. Limits can be
overridden deployment-wide with `PRAGENT_MAX_REVIEW_STEPS`,
`PRAGENT_MAX_REVIEW_TOKENS`, `PRAGENT_MAX_REVIEW_OUTPUT_TOKENS`, and
`PRAGENT_REVIEW_TIMEOUT`, or per repository in the trusted base-branch
`.pr-review.json`; repository values win.

Two measured reviews of a ~1100-line PR in this repo: 28 and 31 agent steps,
~2.1M input tokens each, **zero cache reads or writes**. The demo repo's PR, same
tier: 126K tokens.

The measurements above are historical uncapped runs. A capped run preserves
completed output, reports the cap reason in the review, and records it in
Langfuse.

| Model | this repo, ~1100-line PR | demo repo PR |
|---|---:|---:|
| Claude Opus 5 | ~$10.79 | ~$0.71 |
| Claude Sonnet 5 | ~$4.32 | ~$0.28 |
| Claude Haiku 4.5 | ~$2.16 | ~$0.14 |

Three things that estimate wrong if you skip them:

1. **The loop resends its context every step.** Cost is roughly quadratic in step
   count, not linear in diff size. This is what `attention-tiering` exists to cap.
2. **Repository size dominates diff size.** The 16x gap above is the same
   reviewer on the same tier — the difference is how much repo there was to read.
3. **Prompt caching is worth about a third of the bill** and is not currently
   happening on this stack. Check `cache_read` before budgeting.

```bash
python3 pilot/cost_model.py --help      # other mixes, volumes, models
```

## Development

```bash
python3 -m pytest tests -q    # stdlib-only tests, no network
```

The pilot is stdlib-only Python by design — it runs from a bare `python:slim`
image with the scripts mounted, and has no dependency resolution to go wrong at
review time.

## License

Not yet chosen. Until one is added, no reuse rights are granted.

_pilot eval judges test 1788201461_