Luigit
repositories / pi-ext

pi-ext

bugabingas pi extensions

owned by admin

extensions/good-job/README.md

Raw
Rendered preview
good-job

good-job

Good Job turns brief praise and complaints into reusable feedback about how Pi should behave.

Use it to capture a useful moment while it is still in the conversation, then review patterns later instead of relying on memory.

Start here

Type gj or good job after Pi does something worth repeating.

Type wtf when Pi does something you want to prevent.

Add details when useful.

gj Kept the change focused and explained the trade-off clearly

Your feedback still reaches the active agent normally.

Good Job separately analyzes the preceding conversation in the background and saves a structured learning when it succeeds.

Review feedback

Run /gj to browse saved learnings, with unaddressed problems first. Problems without applied mitigation are labeled and counted; praise remains available.

Run /gj status to see queued, completed, and failed feedback work.

Run /gj learn to ask your active agent for concise, cited diffs based on all saved learnings and remediation metadata. It reads the relevant files before proposing changes; it does not apply them.

Run /gj open to open the local feedback directory.

Feedback analysis sends relevant conversation content to the selected model and saves learnings locally. Do not use it for conversations whose contents you do not want processed that way.

The generated learning is designed not to reproduce credentials, secrets, long source excerpts, or private data, but review stored records before sharing them.

Track remediation

/gj learn asks the agent to propose linked remediation metadata alongside each change. After approval, the agent is instructed to apply the approved changes and record their status itself; you do not need to remember record IDs or run a separate bookkeeping command. Only relevant outcome evidence justifies verified; instruction edits or unrelated green checks do not.

For manual maintenance, select a problem in /gj, then press r in its detail view to remediate it. RPC clients receive the equivalent action dialog. You can also run /gj remediate to select a problem, or /gj remediate <problem-id> to edit it directly. Dialogs record a change reference, status, application date, and verification evidence when relevant. Choose proposed, applied, or verified; already-applied or verified changes need no intermediate saves. Verification requires a verification date and evidence describing the observed check, outcome, and scope. Cancel any dialog or decline the save confirmation to leave stored metadata unchanged. TUI and RPC dialogs are supported; noninteractive modes cannot record remediation through this command.

Remediation is stored separately; original feedback stays unchanged and praise has no mitigation status. Each problem has one current remediation entry; editing replaces its metadata, not an audit history. Saved status appears in /gj lists and details. /gj learn asks the agent to retain addressed records and compare later problems with application dates and verification evidence. Verified means the recorded check passed, not that recurrence is impossible.

Optional settings

By default, records use Pi's local app-data directory and analysis uses the current session model.

Set a directory or choose a separate analysis model only when you need that control.

{
  "good-job": {
    "directory": "~/shared/gj",
    "model": "provider/model-id",
    "thinkingLevel": "high"
  }
}

Settings may come from trusted project .pi/settings.json or user settings.

See Installation.

Reference

Commands

Command Purpose
/gj List, inspect, and learn from explicit feedback

Setting names

Setting
good-job.directory
good-job.model
good-job.thinkingLevel

Setting values

Setting Default Valid values Purpose
good-job.thinkingLevel Unset "minimal", "low", "medium", "high", "xhigh", "max" Job thinking level: Leave unset to retain the job model's default reasoning level.
<img src="assets/banner.svg" alt="good-job" height="70">

# good-job

Good Job turns brief praise and complaints into reusable feedback about how Pi should behave.

Use it to capture a useful moment while it is still in the conversation, then review patterns later instead of relying on memory.

## Start here

Type `gj` or `good job` after Pi does something worth repeating.

Type `wtf` when Pi does something you want to prevent.

Add details when useful.

```text
gj Kept the change focused and explained the trade-off clearly
```

Your feedback still reaches the active agent normally.

Good Job separately analyzes the preceding conversation in the background and saves a structured learning when it succeeds.

## Review feedback

Run `/gj` to browse saved learnings, with unaddressed problems first.
Problems without applied mitigation are labeled and counted; praise remains available.

Run `/gj status` to see queued, completed, and failed feedback work.

Run `/gj learn` to ask your active agent for concise, cited diffs based on all saved learnings and remediation metadata.
It reads the relevant files before proposing changes; it does not apply them.

Run `/gj open` to open the local feedback directory.

Feedback analysis sends relevant conversation content to the selected model and saves learnings locally.
Do not use it for conversations whose contents you do not want processed that way.

The generated learning is designed not to reproduce credentials, secrets, long source excerpts, or private data, but review stored records before sharing them.

## Track remediation

`/gj learn` asks the agent to propose linked remediation metadata alongside each change.
After approval, the agent is instructed to apply the approved changes and record their status itself; you do not need to remember record IDs or run a separate bookkeeping command.
Only relevant outcome evidence justifies `verified`; instruction edits or unrelated green checks do not.

For manual maintenance, select a problem in `/gj`, then press `r` in its detail view to remediate it.
RPC clients receive the equivalent action dialog.
You can also run `/gj remediate` to select a problem, or `/gj remediate <problem-id>` to edit it directly.
Dialogs record a change reference, status, application date, and verification evidence when relevant.
Choose `proposed`, `applied`, or `verified`; already-applied or verified changes need no intermediate saves.
Verification requires a verification date and evidence describing the observed check, outcome, and scope.
Cancel any dialog or decline the save confirmation to leave stored metadata unchanged.
TUI and RPC dialogs are supported; noninteractive modes cannot record remediation through this command.

Remediation is stored separately; original feedback stays unchanged and praise has no mitigation status.
Each problem has one current remediation entry; editing replaces its metadata, not an audit history.
Saved status appears in `/gj` lists and details.
`/gj learn` asks the agent to retain addressed records and compare later problems with application dates and verification evidence.
Verified means the recorded check passed, not that recurrence is impossible.

## Optional settings

By default, records use Pi's local app-data directory and analysis uses the current session model.

Set a directory or choose a separate analysis model only when you need that control.

```json
{
  "good-job": {
    "directory": "~/shared/gj",
    "model": "provider/model-id",
    "thinkingLevel": "high"
  }
}
```

Settings may come from trusted project `.pi/settings.json` or user settings.

See [Installation](../../README.md#install).

## Reference

<!-- config:start -->
### Commands

| Command | Purpose |
| --- | --- |
| `/gj` | List, inspect, and learn from explicit feedback |

### Setting names

| Setting |
| --- |
| `good-job.directory` |
| `good-job.model` |
| `good-job.thinkingLevel` |

#### Setting values

| Setting | Default | Valid values | Purpose |
| --- | --- | --- | --- |
| `good-job.thinkingLevel` | Unset | `"minimal"`, `"low"`, `"medium"`, `"high"`, `"xhigh"`, `"max"` | Job thinking level: Leave unset to retain the job model's default reasoning level. |
<!-- config:end -->