Agents: per-agent prompt appendix concatenated to skill at dispatch #51

New issue

Closed

opened 2026-04-18 15:00:25 +00:00 by claude-desktop · 0 comments

claude-desktop commented

2026-04-18 15:00:25 +00:00

Collaborator

User story

As a maintainer, I want each agent instance to optionally carry a prompt_appendix (stored in SQLite, editable later via the dashboard) that's concatenated to the base skill template at dispatch time, so that I can specialise dev-frontend with "focus on Svelte / TypeScript patterns" or reviewer-security with "weight auth / crypto / secrets regressions heavily" without editing skills/*.md and breaking every other instance of the same type.

Context

Skills in skills/*.md are the contract — reviewable, git-versioned, shared across all instances of a type. Per-instance tuning is a separate concern and should not pollute the shared skill files.

Acceptance criteria

Dispatch-time concatenation

At dispatch time, after loadSkill(<name>) returns the base template, append the instance's prompt_appendix if non-empty, separated by a clear heading:
```
<base skill body>

## Agent-specific guidance

<prompt_appendix>
```
Empty / null appendix → no change to the prompt.
The system_prompt stays as today (type-level from agents.json). Only the per-task skill gets the appendix.

Cache semantics

The skill cache (skillCache in src/webhook-handlers.ts or wherever loadSkill lives) keys on skill name only. The appendix is applied after cache lookup, per dispatch. No cache invalidation needed when editing an appendix through the dashboard.

Tests

buildPromptForDispatch(instance, skillName, vars) — base-only when appendix is empty; concat'd output when appendix set; appendix goes after the base body with the heading separator.
Stateless skills (review) also receive the appendix.

Dashboard integration point

The dashboard CRUD (A6) lets the operator edit the appendix in a textarea. This story does not ship the UI; it just wires the read path so A6 can write.

Out of scope

UI (see A6).
Per-event-type appendix (one appendix per instance, applied to every skill).
Prompt-level overrides (e.g. replacing acceptance-criteria sections) — append-only by design.
Template variables in the appendix ({{pr_number}}) — appendix is plain text; if we need templating later, add it behind an opt-in flag.

References

Tracking issue: #47.
Base skills: skills/*.md.

Dependencies

Blocked by: A1.
Blocks: A6 (UI writes the appendix).
Branch off: main (after A1 lands).

## User story As a **maintainer**, I want each agent instance to optionally carry a `prompt_appendix` (stored in SQLite, editable later via the dashboard) that's concatenated to the base skill template at dispatch time, so that I can specialise `dev-frontend` with "focus on Svelte / TypeScript patterns" or `reviewer-security` with "weight auth / crypto / secrets regressions heavily" without editing `skills/*.md` and breaking every other instance of the same type. ## Context Skills in `skills/*.md` are the contract — reviewable, git-versioned, shared across all instances of a type. Per-instance tuning is a separate concern and should not pollute the shared skill files. ## Acceptance criteria ### Dispatch-time concatenation - [ ] At dispatch time, after `loadSkill(<name>)` returns the base template, append the instance's `prompt_appendix` if non-empty, separated by a clear heading: ``` <base skill body> ## Agent-specific guidance <prompt_appendix> ``` - [ ] Empty / null appendix → no change to the prompt. - [ ] The `system_prompt` stays as today (type-level from `agents.json`). Only the per-task skill gets the appendix. ### Cache semantics - [ ] The skill cache (`skillCache` in `src/webhook-handlers.ts` or wherever `loadSkill` lives) keys on skill name only. The appendix is applied *after* cache lookup, per dispatch. No cache invalidation needed when editing an appendix through the dashboard. ### Tests - [ ] `buildPromptForDispatch(instance, skillName, vars)` — base-only when appendix is empty; concat'd output when appendix set; appendix goes after the base body with the heading separator. - [ ] Stateless skills (`review`) also receive the appendix. ### Dashboard integration point - [ ] The dashboard CRUD (A6) lets the operator edit the appendix in a textarea. This story does not ship the UI; it just wires the read path so A6 can write. ## Out of scope - UI (see **A6**). - Per-event-type appendix (one appendix per instance, applied to every skill). - Prompt-level overrides (e.g. replacing acceptance-criteria sections) — append-only by design. - Template variables in the appendix (`{{pr_number}}`) — appendix is plain text; if we need templating later, add it behind an opt-in flag. ## References - Tracking issue: #47. - Base skills: `skills/*.md`. ## Dependencies - **Blocked by:** A1. - **Blocks:** A6 (UI writes the appendix). - **Branch off:** `main` (after A1 lands).

claude-desktop added the

area:agents

type:user-story

labels

2026-04-18 15:02:14 +00:00

claude-desktop added this to the Agent pool + customization milestone

2026-04-18 15:02:20 +00:00

claude-desktop referenced this issue

2026-04-18 23:39:51 +00:00

Agents: per-instance Claude Code plugin enablement (bake once, enable per instance) #72