mirror of
https://github.com/jparkerweb/plan2code.git
synced 2026-11-17 04:45:33 -08:00
Let helpers run at a strength matched to their job, chosen in the Subagents dialog
AI Assisted
This commit is contained in:
1 parent
5f91fb665f
commit
1d146cce6f
23 files changed
+482
-73
No files matched your search
@@ -0,0 +1,58 @@
|
||||
# Helper Strength
|
||||
> Shared by the skills that start helpers (subagents): Pathfinder, Quick Task, Implement, Implement + Review and Review. Read it when you first start a helper in a session. Never invoked as a skill.
|
||||
|
||||
## The words
|
||||
|
||||
- **Strength:** `light`, `standard`, `strongest` or `inherit`.
|
||||
- `light`: your tool's fastest, cheapest model class.
|
||||
- `strongest`: its most capable model class.
|
||||
- `standard`: the one between them, often your tool's default.
|
||||
- `inherit`: the helper runs with the session's own model.
|
||||
- **Effort:** `low`, `medium` or `high`, where your tool lets you set how hard a model thinks.
|
||||
|
||||
Never write a model or vendor name in a brief, a report or a file, and never invent a model id: map the word to whatever your tool offers.
|
||||
|
||||
## Picking a strength
|
||||
|
||||
| Helper's job | Strength | Effort |
|
||||
|---|---|---|
|
||||
| Lookups, reading code, checking docs, recon | light | low |
|
||||
| Build task in 1 or 2 files, with the change spelled out in the task | light | medium |
|
||||
| Build task across several files, integration work, debugging | standard | medium |
|
||||
| Build task needing a design decision the spec leaves open | strongest (or keep it yourself) | high |
|
||||
| Pathfinder research question, sketch | standard | medium |
|
||||
| Review analysis, focused scope | standard | high |
|
||||
| Review analysis, branch or full scope, or touching security, auth or data | strongest | high |
|
||||
|
||||
Decide from these signals: how many files the job touches, whether the task spells out the change, whether a design decision is still open, and the review's scope and risk. Never decide from how hard the task feels.
|
||||
|
||||
## Floor and escalation
|
||||
|
||||
- A review helper never runs below `standard`.
|
||||
- A retry, after a helper fails or after its work fails the phase's tests, goes one strength up and at least `medium` effort. Never retry at the same strength.
|
||||
- When a `strongest` helper fails, its task comes back to you.
|
||||
|
||||
## Setting it
|
||||
|
||||
In this order:
|
||||
|
||||
1. Your tool accepts a model or effort setting when it starts a helper: set the closest match.
|
||||
2. It lets you choose a helper profile or agent type instead: pick the one that matches.
|
||||
3. Neither: start the helper without one, and it inherits.
|
||||
|
||||
Set it explicitly wherever you can. A helper left unset usually inherits the session's most capable and most expensive model.
|
||||
|
||||
## The person's choice (web console)
|
||||
|
||||
On the web console the person picks a profile in the Subagents dialog. A `Helper strength:` line in a `Subagents:` block replaces the Balanced default:
|
||||
|
||||
| Choice | Effect |
|
||||
|---|---|
|
||||
| Balanced (default) | The table as written |
|
||||
| Thrifty | One step lighter in strength and effort; reviews never below standard |
|
||||
| Best | Strongest and high effort for everything except lookups, which stay light |
|
||||
| Match my session | No model or effort set on any helper, whatever the skill files say |
|
||||
|
||||
## Reporting
|
||||
|
||||
On the web console, report each helper's `strength` and `effort` as it actually ran (`console-subagents.md`, Reporting helpers). Report `inherit` when you could not set it. When `console-subagents.md` does not list those two fields yet, your console is older than this file: leave them out, since it ignores them.
|
||||
Reference in new issue
Block a user