Multiplayer agent harness for work.
Make suggested prompts collaborative and contextual (#1246)
Regan Bell34c30aa2a64d
diff --git a/.github/workflows/pgbouncer.yml b/.github/workflows/pgbouncer.ymlindex cad73c149..4d3835aca 100644--- a/.github/workflows/pgbouncer.yml+++ b/.github/workflows/pgbouncer.yml@@ -10,8 +10,8 @@ jobs:runs-on: ubuntu-latesttimeout-minutes: 15steps:- - uses: actions/checkout@v4- - uses: actions/setup-node@v4+ - uses: actions/checkout@93cb6efe18208431cddfb8368fd83d5badbf9bfd+ - uses: actions/setup-node@a0853c24544627f65ddf259abe73b1d18a591444with:node-version: "24.18.1"cache: npmdiff --git a/plugins/web-ui/README.md b/plugins/web-ui/README.mdindex 8814b750f..01a438678 100644--- a/plugins/web-ui/README.md+++ b/plugins/web-ui/README.md@@ -56,13 +56,13 @@ is bundled into the public application.{"id": "weekly-brief","title": "Wake up to a fresh briefing",- "prompt": "Set up a recurring briefing. Ask me which topics, cadence, timezone, and delivery destination to use.",+ "prompt": "Let's set up a recurring briefing on the topics I follow.","icon": "schedule"},{"id": "project-app","title": "Build a home for my projects",- "prompt": "Build a private project tracker app. Start by understanding my workflow and existing data.",+ "prompt": "Let's build a private app to keep track of my projects.","icon": "app"}]@@ -93,7 +93,11 @@ staggered by user. The cron runner uses the normal owner-scoped session, runtimeselection, memory, history, tools, and authorized data access. It has no deliverydestination. The standing task in `src/suggestions/activities.ts` asks it to researchrelevant context, avoid mutations or notifications, and return three validated-activity objects. It does not create a separate reduced-context model call.+activity objects. Draft prompts use a natural, collaborative voice, such as "Let's...",+with relevant facts and uncertainties as neutral context, without attributing knowledge,+feelings, or beliefs to the user. They preserve explicit user+preferences and scope while leaving the approach to the responding agent. It does not+create a separate reduced-context model call.The UI reads the latest valid result from the cron's completed personal session,including responses larger than the truncated fire-log preview. It displays thediff --git a/plugins/web-ui/test/suggested-activities.test.ts b/plugins/web-ui/test/suggested-activities.test.tsindex 24a1cea49..e637d30a5 100644--- a/plugins/web-ui/test/suggested-activities.test.ts+++ b/plugins/web-ui/test/suggested-activities.test.ts@@ -5,7 +5,7 @@ import { parseSuggestedActivities } from "../../chassis/src/suggested-activitiesconst activity = {id: "weekly-brief",title: "Wake up to a fresh briefing",- prompt: "Set up a weekly briefing. Ask me which topics and schedule to use.",+ prompt: "Let's set up a weekly briefing on the topics I follow.",icon: "schedule",};diff --git a/src/suggestions/activities.ts b/src/suggestions/activities.tsindex 8c30226b3..50d951d62 100644--- a/src/suggestions/activities.ts+++ b/src/suggestions/activities.ts@@ -14,7 +14,8 @@ Prefer concrete next steps tailored to current work. Avoid completed tasks and dReturn only a JSON array of exactly three objects with id, title, prompt, and icon. No markdown or surrounding prose.id: a unique lowercase slug, at most 64 characters.title: a polished invitation of 4–9 words, at most 65 characters. Use sentence case, lead with a verb and an outcome, and keep it understated.-prompt: an actionable first-person request, at most 1200 characters, that the person can edit before sending. Include enough grounded context to make it useful. Ask for missing inputs, and never invent people, metrics, deadlines, source access, or completed actions.+prompt: a natural, collaborative request, at most 1200 characters. Prefer phrasing such as "Let's..." to introduce the desired outcome, then briefly share the relevant grounded context in neutral language, usually in 2–4 sentences. State uncertainties about the situation directly, without attributing knowledge, feelings, or beliefs to the user. Leave the approach and any necessary follow-up questions to the responding agent. Preserve explicit user preferences and scope, but do not add procedural checklists, precautionary prohibitions, approval requirements, or instructions about how the agent should think or use tools. Never invent people, metrics, deadlines, source access, or completed actions.+Example prompt: "Let's get the staging playground working again for testing. The latest checkpoint is in the playground conversation and HANDOFF.md. Last time, staging returned a 503 and AWS access was blocked. Either may have changed in the meantime."icon: one relevant emoji, or "yc" for YC-specific work only when deployment guidance establishes a YC context. Never use image URLs or markup.`;export interface SuggestedActivityProfile {