You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Copy file name to clipboardExpand all lines: plugins/autoresearch/skills/autoresearch/SKILL.md
+87-1Lines changed: 87 additions & 1 deletion
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -310,6 +310,92 @@ Droid sessions have finite context. To handle this gracefully:
310
310
-**Think longer when stuck.** Re-read source files, study the data, reason about what's actually happening. The best ideas come from deep understanding.
When the experiment loop ends (termination condition met, user interrupts, or context exhausted), finalize the results into clean, reviewable branches. This is the last phase of an autoresearch session.
git log --oneline --stat $(git merge-base HEAD main)..HEAD
327
+
```
328
+
329
+
### Step 2: Group Changes
330
+
331
+
Group kept experiments into **logical changesets**. Each group should:
332
+
- Represent a single coherent optimization or change
333
+
- Not share modified files with other groups (so branches can merge independently)
334
+
- Have a clear description of what it achieves and the metric improvement
335
+
336
+
Present the proposed grouping to the user for approval:
337
+
338
+
```
339
+
Group 1: "Reduce model depth from 8 to 6"
340
+
Files: train.py (DEPTH, HEAD_DIM, N_EMBED)
341
+
Metric improvement: val_bpb 1.15 -> 1.08 (-6.1%)
342
+
Experiments: #3, #7, #12
343
+
344
+
Group 2: "Switch to cosine LR schedule"
345
+
Files: train.py (lr_schedule, warmup_steps)
346
+
Metric improvement: val_bpb 1.08 -> 1.05 (-2.8%)
347
+
Experiments: #15, #18
348
+
```
349
+
350
+
Wait for user confirmation before proceeding.
351
+
352
+
### Step 3: Resolve File Conflicts
353
+
354
+
If groups share files, resolve before creating branches:
355
+
- Merge the groups into one (if changes are related)
356
+
- Split the file changes more carefully (if they're truly independent modifications to different parts)
357
+
- Ask the user which group gets priority
358
+
359
+
Groups **must not share files** — each branch must be independently mergeable. If all changes touch the same file and can't be separated, create a single finalized branch with all improvements combined.
- Metric: <name> improved from <baseline> to <best> (<delta>%)
373
+
- Confidence: <score>x noise floor
374
+
- Experiments: <count> total, <kept> kept"
375
+
```
376
+
377
+
### Step 5: Verify and Report
378
+
379
+
For each finalized branch, run the benchmark to confirm the improvement holds, run any checks if applicable, and verify it merges cleanly with main.
380
+
381
+
Present a summary to the user:
382
+
383
+
```
384
+
Created 2 clean branches from 20 experiments:
385
+
386
+
autoresearch/finalize/reduce-depth
387
+
val_bpb: 1.15 -> 1.08 (-6.1%)
388
+
Ready for review
389
+
390
+
autoresearch/finalize/cosine-schedule
391
+
val_bpb: 1.08 -> 1.05 (-2.8%)
392
+
Ready for review
393
+
394
+
Original experiment branch preserved: autoresearch/<session-branch>
395
+
```
396
+
397
+
The original experiment branch is always preserved — finalization creates new branches.
398
+
313
399
## Mission Worker Mode
314
400
315
-
When running as a mission worker, the feature description specifies the optimization goal, termination condition, files in scope, and constraints. Read it carefully, follow the same loop procedure above, and respect the termination condition. When the condition is met, stop and report results in the handoff.
401
+
When running as a mission worker, the feature description specifies the optimization goal, termination condition, files in scope, and constraints. Read it carefully, follow the same loop procedure above, and respect the termination condition. When the condition is met, run finalization and report results in the handoff.
0 commit comments