chore: init commit
in worker.../repo/GITFOLDER.zip is the .git folder.
This commit is contained in:
@@ -0,0 +1,57 @@
|
||||
---
|
||||
description: Capture a snapshot of the current conversation and repo state.
|
||||
---
|
||||
|
||||
# Create Snapshot
|
||||
|
||||
You are capturing a snapshot of the current conversation and repo state so it can be replayed as an RL training task.
|
||||
|
||||
## Step 1: Ask annotation questions
|
||||
|
||||
**Important — tell the user this first, verbatim:**
|
||||
|
||||
> ⚠️ This snapshot captures your entire conversation history with me, not just the most
|
||||
> recent turn. If you told me the answer earlier in this conversation, or steered me
|
||||
> toward it, the agent will see that same context when the snapshot replays — and will
|
||||
> probably solve the task without making the mistake. Your task will be contaminated.
|
||||
>
|
||||
> If you've leaked the answer at any point in this conversation: if your agent can rewind
|
||||
> (Claude Code's `/rewind`), rewind to a point before the contamination and snapshot from
|
||||
> there. If it can't — codex has no rewind — this snapshot is not salvageable: start a
|
||||
> fresh session, reproduce the mistake without steering, and snapshot that instead.
|
||||
|
||||
Wait for the user to acknowledge before moving on.
|
||||
|
||||
Ask the user each of these questions **one at a time** as plain text, waiting for their response before proceeding to the next:
|
||||
|
||||
1. "What were you trying to do?"
|
||||
2. "What were you hoping was going to happen?"
|
||||
3. "What did the agent actually do instead?"
|
||||
|
||||
## Step 2: Propose a slug
|
||||
|
||||
Based on the user's answers, generate a **short kebab-case slug** (2-4 words) that captures the essence of the mistake. For example: `bad-refactor`, `wrong-test-strategy`, `missed-edge-case`.
|
||||
|
||||
Present your suggestion and ask the user to confirm or provide an alternative.
|
||||
|
||||
## Step 3: Write annotation file and run capture
|
||||
|
||||
Write the annotation to a temporary JSON file, then run the capture script.
|
||||
|
||||
Write this JSON to a temp file (use a path like `/tmp/snapshot-annotation-<timestamp>.json`):
|
||||
|
||||
```json
|
||||
{
|
||||
"what_trying": "<answer to question 1>",
|
||||
"what_hoping": "<answer to question 2>",
|
||||
"what_happened": "<answer to question 3>"
|
||||
}
|
||||
```
|
||||
|
||||
Then run:
|
||||
|
||||
```bash
|
||||
"${CLAUDE_PLUGIN_ROOT}/bin/capture-snapshot.mjs" --slug <slug> --annotation <temp-file-path> --output-dir "${CLAUDE_PLUGIN_ROOT}/../../snapshots" --plugin-data "${CLAUDE_PLUGIN_DATA:-${CLAUDE_PLUGIN_ROOT}/.data}"
|
||||
```
|
||||
|
||||
Report the script's stdout output verbatim to the user. Do not paraphrase or shorten paths.
|
||||
Reference in New Issue
Block a user