@nanobpm/urban-pr-review-codefirst 1.0.0 → 1.1.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +12 -0
- package/README.md +153 -5
- package/components/review-round.json +2 -2
- package/deno.json +1 -1
- package/deno.lock +1603 -12
- package/main.ts +4 -3
- package/package.json +2 -2
- package/workflows/convergence-loop.ts +7 -4
package/CHANGELOG.md
CHANGED
|
@@ -1,3 +1,15 @@
|
|
|
1
|
+
# [1.1.0](https://github.com/jwulf/urban-pr-review-codefirst/compare/v1.0.0...v1.1.0) (2026-07-31)
|
|
2
|
+
|
|
3
|
+
|
|
4
|
+
### Bug Fixes
|
|
5
|
+
|
|
6
|
+
* **ci:** sync package-lock to @nanobpm/workflow@0.4.0 ([57aa919](https://github.com/jwulf/urban-pr-review-codefirst/commit/57aa919e4f5f1dbc2e9022cda19b2aa1594dff50)), closes [#1](https://github.com/jwulf/urban-pr-review-codefirst/issues/1)
|
|
7
|
+
|
|
8
|
+
|
|
9
|
+
### Features
|
|
10
|
+
|
|
11
|
+
* emit senior:pr-review so one reviewer serves both apps ([#1](https://github.com/jwulf/urban-pr-review-codefirst/issues/1)) ([0d3e39b](https://github.com/jwulf/urban-pr-review-codefirst/commit/0d3e39b112ce6cfb9d56874247f7d34d85e3e4fd))
|
|
12
|
+
|
|
1
13
|
# 1.0.0 (2026-07-31)
|
|
2
14
|
|
|
3
15
|
|
package/README.md
CHANGED
|
@@ -17,6 +17,152 @@ them (`convergence-loop.bpmn` + `workers/*/worker.ts`), code-first writes them
|
|
|
17
17
|
(`workflows/convergence-loop.ts`, with the record steps hosted in-process). See
|
|
18
18
|
[Persistence](#persistence) and [Frontend](#frontend) below.
|
|
19
19
|
|
|
20
|
+
## HOWTO: run agents that converge PRs for you, all day
|
|
21
|
+
|
|
22
|
+
### The problem this solves
|
|
23
|
+
|
|
24
|
+
Driving a PR to convergence is mostly *waiting*: open a review, wait 5–15 min for
|
|
25
|
+
the reviewer, address the comments, re-request, wait again — often ten rounds
|
|
26
|
+
deep, escalating to a human only when the reviewer is stuck. If a single agent
|
|
27
|
+
babysits one PR, it spends almost all of its time **idle-polling** a review that
|
|
28
|
+
hasn't landed yet. That is wasted wall-clock and a wasted worker slot.
|
|
29
|
+
|
|
30
|
+
This app inverts that. **The BPMN process the SDK derives owns the durable wait**
|
|
31
|
+
between rounds (a real `w.signal` message-catch event), so no agent is ever parked
|
|
32
|
+
holding a job slot while a review is pending. Instead you **hire a couple of agent
|
|
33
|
+
workers once**, and they pull `senior:pr-review` jobs from *whatever PR is ready
|
|
34
|
+
right now* — alternating across all in-flight reviews. Two agents can keep a dozen
|
|
35
|
+
PRs converging in parallel, and they only run when there is actual work to do.
|
|
36
|
+
|
|
37
|
+
Because the `review-round` task overrides its job type to **`senior:pr-review`**
|
|
38
|
+
(via `w.task("review-round", { jobType: "senior:pr-review" })`), the *same* hired
|
|
39
|
+
reviewer services both this code-first app and the model-first
|
|
40
|
+
[urban-pr-review](../urban-pr-review). Set it up by: **(1)** running this app,
|
|
41
|
+
**(2)** submitting PRs, and **(3)** hiring agent workers with
|
|
42
|
+
[`c8ctl nano`](https://github.com/jwulf/c8ctl-plugin-nano).
|
|
43
|
+
|
|
44
|
+
### 0. Prerequisites
|
|
45
|
+
|
|
46
|
+
- A running **Nano gateway/engine** (default `http://localhost:8080`). This is
|
|
47
|
+
what the app deploys to and what agents pull jobs from.
|
|
48
|
+
- **[Deno](https://deno.land/)** (to run this app) and the **c8ctl CLI with the
|
|
49
|
+
`nano` plugin** installed (to hire/run agents).
|
|
50
|
+
- On each machine that will *host an agent*: the **GitHub CLI** logged in
|
|
51
|
+
(`gh auth login`) or a `GITHUB_TOKEN`/`GH_TOKEN` in the environment, and the
|
|
52
|
+
agent harness itself — e.g. the **[Copilot CLI](https://github.com/github/copilot-cli)**
|
|
53
|
+
(`copilot`).
|
|
54
|
+
|
|
55
|
+
### 1. Install the app into your Nano IDE
|
|
56
|
+
|
|
57
|
+
Open the Nano console and **Projects → Import by reference**, pointing at this
|
|
58
|
+
app's folder (ADR 0041), or drop a `*.project-ref.json` next to your other
|
|
59
|
+
projects. `nano-ide.ext.json` marks it as an example. You can also just run it
|
|
60
|
+
standalone (next step) — the IDE import is only needed to launch/manage it from
|
|
61
|
+
the console.
|
|
62
|
+
|
|
63
|
+
### 2. Start the app
|
|
64
|
+
|
|
65
|
+
```sh
|
|
66
|
+
deno task start # → http://localhost:3000
|
|
67
|
+
```
|
|
68
|
+
|
|
69
|
+
That applies the DB migrations, deploys the derived flow, hosts the in-process
|
|
70
|
+
`w.run` record steps, serves the web UI at **<http://localhost:3000>**, and runs
|
|
71
|
+
the review-ready poller (which needs `GITHUB_TOKEN` to watch GitHub — without it,
|
|
72
|
+
re-reviews come only from the UI/CLI). Point it at a non-default gateway with
|
|
73
|
+
`NANOBPMN_BASE_URL`; change the port with `PR_REVIEW_PORT`.
|
|
74
|
+
|
|
75
|
+
### 3. Submit a PR
|
|
76
|
+
|
|
77
|
+
From the web UI, or from the CLI:
|
|
78
|
+
|
|
79
|
+
```sh
|
|
80
|
+
deno task submit https://github.com/owner/repo/pull/42 5 # 5 = maxRounds
|
|
81
|
+
```
|
|
82
|
+
|
|
83
|
+
Each submitted PR starts one durable `convergence-loop` instance that parks until
|
|
84
|
+
an agent services its `senior:pr-review` round.
|
|
85
|
+
|
|
86
|
+
### 4. Hire an agent worker
|
|
87
|
+
|
|
88
|
+
An agent is a CLI harness (Copilot CLI here) turned into a Nano job worker. Hire a
|
|
89
|
+
profile whose **rank + capability** produce the `senior:pr-review` token this
|
|
90
|
+
flow's task emits:
|
|
91
|
+
|
|
92
|
+
```sh
|
|
93
|
+
c8ctl nano hire \
|
|
94
|
+
--name reviewer \
|
|
95
|
+
--rank senior \
|
|
96
|
+
--capabilities pr-review \
|
|
97
|
+
--command 'copilot -p - --allow-all-tools' \
|
|
98
|
+
--model <your-model>
|
|
99
|
+
```
|
|
100
|
+
|
|
101
|
+
- `--rank senior` + `--capabilities pr-review` makes the worker subscribe to the
|
|
102
|
+
`senior:pr-review` job type — exactly the token this flow's `review-round` task
|
|
103
|
+
emits (thanks to the SDK job-type override), so **one profile serves both apps.**
|
|
104
|
+
- `--command 'copilot -p - --allow-all-tools'` starts the Copilot CLI reading its
|
|
105
|
+
prompt from **stdin** (`-p -`). The harness pipes the whole job JSON (prompt +
|
|
106
|
+
`job.variables`: `prUrl`, `repo`, `prNumber`, `round`, `answer?`) to stdin; the
|
|
107
|
+
review-round instructions tell the agent how to read it and where to write its
|
|
108
|
+
result.
|
|
109
|
+
- **`--allow-all-tools` is the crucial flag.** Without it, Copilot pauses to ask
|
|
110
|
+
permission before each tool call — and an unattended worker has no human to
|
|
111
|
+
answer, so the job stalls. `--allow-all-tools` lets it run the whole round
|
|
112
|
+
non-interactively. (Pair with `--deny-tool` to blocklist specific tools.)
|
|
113
|
+
|
|
114
|
+
> ⚠️ **Only enable `--allow-all-tools` for code and hosts you trust.** It grants
|
|
115
|
+
> the agent unattended, broad permissions (shell, file writes, network). Each
|
|
116
|
+
> job runs in a throwaway per-job workspace (see below), but the worker still
|
|
117
|
+
> runs as your user on the host — don't point it at untrusted PRs on a shared
|
|
118
|
+
> machine. Use `--deny-tool` to narrow it, or a container sandbox for stronger
|
|
119
|
+
> isolation.
|
|
120
|
+
|
|
121
|
+
> If you'd rather not override the job type in the flow, you can instead leave the
|
|
122
|
+
> derived `convergence-loop:review-round` type and point a worker at it explicitly
|
|
123
|
+
> with `c8ctl nano work reviewer --job-type convergence-loop:review-round`. This
|
|
124
|
+
> app ships with the `senior:pr-review` override so it shares a reviewer with the
|
|
125
|
+
> model-first app out of the box.
|
|
126
|
+
|
|
127
|
+
### 5. Put the agent to work
|
|
128
|
+
|
|
129
|
+
```sh
|
|
130
|
+
c8ctl nano work reviewer # polls for senior:pr-review jobs until Ctrl-C
|
|
131
|
+
```
|
|
132
|
+
|
|
133
|
+
Now every PR you submit gets picked up automatically. Start a **second** worker
|
|
134
|
+
(same command, another terminal or another machine) and the two alternate across
|
|
135
|
+
whichever PRs are ready — that is the idle-time you reclaim. Run more than one job
|
|
136
|
+
at once per worker with `--max-parallel 2`.
|
|
137
|
+
|
|
138
|
+
### Isolation — each job gets its own clean workspace
|
|
139
|
+
|
|
140
|
+
In the default **host mode** (`--sandbox none`), the worker provisions a
|
|
141
|
+
**throwaway, per-job workspace**: a fresh clone under `<state>/agent-runs/run-*`
|
|
142
|
+
checked out on the PR's head branch, exposed to the agent as `AGENT_WORKSPACE` /
|
|
143
|
+
`REPO_URL` / `REPO_BRANCH` / `REPO_REF`, and **reaped after the job**. So multiple
|
|
144
|
+
agents on one host don't step on each other. Host workers inherit *your*
|
|
145
|
+
`gh`/`GITHUB_TOKEN` login, so no extra auth is needed. Docker/podman sandboxes
|
|
146
|
+
exist (`--sandbox docker --image …`) for stronger isolation, but container-side
|
|
147
|
+
git provisioning is a later increment — container jobs don't clone yet, and don't
|
|
148
|
+
inherit your host login (pass credentials via `--secret-resolver host` /
|
|
149
|
+
`secretRefs`). For the review loop, **host mode is the recommended setup.**
|
|
150
|
+
|
|
151
|
+
### Run it across spare hardware (incl. a Raspberry Pi)
|
|
152
|
+
|
|
153
|
+
Nano ships ARM binaries (arm64/armv7/armv6), so the whole thing scales down nicely:
|
|
154
|
+
|
|
155
|
+
- **All-in-one:** run the gateway, this app, and one or two workers on your laptop.
|
|
156
|
+
- **Distributed:** run the **Nano gateway on a Raspberry Pi** (always-on, low
|
|
157
|
+
power), run this app anywhere, and put **agent workers on spare machines** — each
|
|
158
|
+
worker just needs the c8ctl CLI, the Copilot CLI logged in, and
|
|
159
|
+
`NANOBPMN_BASE_URL` pointed at the Pi. Add or remove workers at will; the BPMN
|
|
160
|
+
process holds all durable state, so workers are stateless and disposable.
|
|
161
|
+
|
|
162
|
+
The payoff: instead of one agent burning wall-clock polling a single PR, a small
|
|
163
|
+
pool of always-available workers keeps every open review converging — and idles to
|
|
164
|
+
zero cost when there's nothing to do.
|
|
165
|
+
|
|
20
166
|
## The flow
|
|
21
167
|
|
|
22
168
|
`workflows/convergence-loop.ts`:
|
|
@@ -35,8 +181,10 @@ loop:
|
|
|
35
181
|
```
|
|
36
182
|
|
|
37
183
|
- **`review-round`** is a `w.task` — an **external** worker. Point a coding-agent
|
|
38
|
-
harness at the job type `
|
|
39
|
-
|
|
184
|
+
harness at the job type `senior:pr-review` (e.g. a `c8ctl nano hire`d rank
|
|
185
|
+
`senior` + capability `pr-review` profile — the same one that services the
|
|
186
|
+
model-first app). The app never hosts or names it; it only owns the durable
|
|
187
|
+
orchestration around it.
|
|
40
188
|
- **`persist-*`** and **`persist-converged`** are `w.run` — app-hosted handlers.
|
|
41
189
|
`persist-round` returns `{ round: round + 1 }`; that return value (not an
|
|
42
190
|
ioMapping) is the loop's only state mutation.
|
|
@@ -55,7 +203,7 @@ the generated BPMN as `nano:shape` + `io.nanobpm.dataEnvelope` properties.
|
|
|
55
203
|
|
|
56
204
|
| Step | Derived job type / message |
|
|
57
205
|
| --- | --- |
|
|
58
|
-
| `review-round` (external) | `
|
|
206
|
+
| `review-round` (external) | `senior:pr-review` (overridden via `w.task("review-round", { jobType: "senior:pr-review" })`) |
|
|
59
207
|
| `persist-round` | `convergence-loop:persist-round` |
|
|
60
208
|
| `persist-escalation` | `convergence-loop:persist-escalation` |
|
|
61
209
|
| `persist-escalation-maxrounds` | `convergence-loop:persist-escalation-maxrounds` |
|
|
@@ -95,8 +243,8 @@ deno task purge
|
|
|
95
243
|
```
|
|
96
244
|
|
|
97
245
|
The reviewer itself is not started by `deno task start` — it is the external
|
|
98
|
-
`
|
|
99
|
-
|
|
246
|
+
`senior:pr-review` job. Host it with a coding-agent harness so the automated
|
|
247
|
+
review is fully decoupled from the durable orchestration.
|
|
100
248
|
|
|
101
249
|
Environment overrides: `PR_REVIEW_PORT` (default `3000`), `NANOBPMN_BASE_URL`
|
|
102
250
|
(gateway), `NANO_APP_DB_URL` (default `file:./app.db`), `NANO_PR_POLL_MS`
|
|
@@ -2,13 +2,13 @@
|
|
|
2
2
|
"$schema": "https://unpkg.com/@camunda/zeebe-element-templates-json-schema@0.44.0/resources/schema.json",
|
|
3
3
|
"id": "io.nanobpm.urban.pr-review-round-codefirst",
|
|
4
4
|
"name": "PR Review Round (agent, code-first)",
|
|
5
|
-
"description": "One decoupled review round serviced by an external agent (e.g. a Copilot instance via `c8ctl nano work`). The agent reads the latest review, addresses it, pushes, re-requests review, and returns a structured status. Mirrors the model-first template, bound to the
|
|
5
|
+
"description": "One decoupled review round serviced by an external agent (e.g. a Copilot instance via `c8ctl nano work`). The agent reads the latest review, addresses it, pushes, re-requests review, and returns a structured status. Mirrors the model-first template, bound to the shared `senior:pr-review` reviewer job type the flow now emits.",
|
|
6
6
|
"appliesTo": ["bpmn:Task"],
|
|
7
7
|
"elementType": { "value": "bpmn:ServiceTask" },
|
|
8
8
|
"properties": [
|
|
9
9
|
{
|
|
10
10
|
"type": "Hidden",
|
|
11
|
-
"value": "
|
|
11
|
+
"value": "senior:pr-review",
|
|
12
12
|
"binding": { "type": "zeebe:taskDefinition:type" }
|
|
13
13
|
},
|
|
14
14
|
{
|
package/deno.json
CHANGED
|
@@ -6,7 +6,7 @@
|
|
|
6
6
|
"exclude": [".deno-cache/", "nano-generated/"]
|
|
7
7
|
},
|
|
8
8
|
"imports": {
|
|
9
|
-
"@nanobpm/workflow": "npm:@nanobpm/workflow@^0.
|
|
9
|
+
"@nanobpm/workflow": "npm:@nanobpm/workflow@^0.4.0",
|
|
10
10
|
"@nanobpm/domain": "./nano-generated/domain.ts",
|
|
11
11
|
"@nanobpm/data": "./nano-generated/data-sdk.ts",
|
|
12
12
|
"@nanobpm/app": "./nano-generated/app-pages.ts",
|