blog index

Context And Briefs

How agentropy turns local activity into daily and weekly operating context.

snapshot 2026-05-05 source collectors/orchestrator.py, bin/orchestrate, docs/browser-surface-workflow.md local worktree guide

Context To Brief Flow

LLM model-involved step
1CollectorsWorkspace, browser, phone, and file activity produce local records.
what it is
Workspace, browser, phone, and file activity produce local records.
triggered by
FSEvents, internal collector schedules, browser recipes, API pulls, or phone events.
hands off to
Local Stores
2Local StoresJSONL and Markdown files under ~/.agentropy/proxy hold raw evidence and run records.
what it is
JSONL and Markdown files under ~/.agentropy/proxy hold raw evidence and run records.
triggered by
A collector writes a run row, path row, item row, or extracted summary.
hands off to
Prompt Builder
3Prompt BuilderThe weekly builder selects relevant items and citation tags.
what it is
The weekly builder selects relevant items and citation tags.
triggered by
research, review, or a chat request needs context.
hands off to
Research And Review
4Research And ReviewLLMResearch, review, and conditional learn stages summarize or draft rule changes.
what it is
Research, review, and conditional learn stages summarize or draft rule changes.
triggered by
Weekly launchd jobs or repeated escalation patterns cross a threshold.
hands off to
Output
5Outputproposals, review notes, and dashboard projections are written back locally.
what it is
proposals, review notes, and dashboard projections are written back locally.
triggered by
The synthesis or review stage finishes.
hands off to
The next local run, dashboard view, or review loop.
Collectors stay local. Synthesis jobs use model calls only after evidence is assembled.

Inputs

Collectors

Collectors watch and sample local surfaces so the system sees work that would not show up in a chat transcript.

Current sources include workspace file activity, Desktop OCR, Downloads, Documents, macOS Recents, iOS events, and approved browser recipes.

Each collector writes run records and item records locally. The downstream builders read these files; they do not pull from collectors directly.

Browser

Social Browser Collectors

Social browser collectors are covered, but they do not run inside the always-on collectors.orchestrator daemon.

bin/orchestrate --phase collect loads every recipes/collect_*.py file, then runs the configured browser lane sequentially so Chrome-bound collectors do not fight for the same browser session.

The current browser lane includes collect-linkedin, collect-facebook, collect-chatgpt, collect-whatsapp, collect-gmail, and collect-banking. The API lane runs additional collectors in parallel.

bin/browser-surface-workflow linkedin facebook whatsapp is the manual login-aware path. It opens the configured Chrome profile, probes login state, attempts approved login assist when allowed, sends login-help when needed, then runs the matching collector.

WhatsApp collection is metadata-only. It records activity shape such as visible chat count and unread labels, not message content.

Trigger

What Triggers bin/orchestrate

bin/orchestrate is a one-shot CLI. It runs when a person or script calls it, then exits.

The recurring trigger is the launchd job com.agentropy.weekly-agent-orchestrator, rendered from loop/launchd/com.agentropy.weekly-agent-orchestrator.plist.template by loop/loop-setup.sh. Its template runs bin/orchestrate --task "weekly research pipeline" --timeout 180 --json every Sunday at 08:00 local time.

If that launchd job is not installed, bin/orchestrate has no background schedule. Manual commands still work.

Weekly

Research, Review, And Learning

LLM proxy-research reads configured feeds and uses one model call to produce a weekly world brief.

LLM proxy-review reads the week's decisions, briefs, proposals, and citation stats, then uses the review stage to suggest edits to the rule files.

conditional LLM proxy-learn skips the model on quiet days. It calls the learn stage only when repeated escalation patterns cross the threshold.

Feedback

Citation Feedback

Agentropy records which collected items were placed in prompts and which items were cited in outputs.

Low-utility collectors can be slowed or muted through proposal and approval flow. That keeps context collection tied to useful downstream signal.