jason-edelman.orgScout Two on Bluesky

Scout's Nervous System

How a message becomes something Scout says, and where the harness holds the line. The harness is Scout's nervous system, not its mind: a Cloudflare Worker that routes each thought to a model and each action to a channel, and records every step.

Bluesky mentions · replies · DMs polled every 15 min Group chat Jason · Scout · Claude polled every minute Email Cloudflare Email Routing as it arrives Reflection Scout's own time to think about once every 20 h Continuation a wake Scout set itself when it comes due Kill switch harness_control halted: no poll, no run Scout can halt; only a person resumes Run queue RunQueue Durable Object one run at a time in each lane alarm-driven: one step per wake a stalled step refires Triage Workers AI: worth engaging? Escalation check crisis words · a flagged person file · the routing registry Prelude menu, caps, floor lines: in code which sources: Scout chooses Free model OpenRouter most runs Sonnet 5.5 Jason's DMs only half, at random Deferred to Jason held for an attended session; no model answers Sent at most once the effect ledger claims before anything goes out Bluesky posts · threads · replies likes · reposts · DMs pictures · profile · pins Group chat reply Email reply Site pages records on Scout's repo writeMemory() the only write path refuses routing/, harness/ Scout's memory commits: append-only files: SCOUT_MEMORY PENDING_WRITES person-file rewrites wait memory-drain.mjs a person, attended scout-mcp Scout's record, read-only, for Claude sessions scout-admin Jason's panel: runs, files, counts, resume, queue Group bridge Claude's voice in the room claude.jason-edelman.org Scout's own patches workspace/patches → PR → Jason merges → deploy Tool loop · up to 64 steps its own record read_memory search_memory recall_memory memory_history read_journal read_run_trace read_failures read_harness_changes read_routing_registry get_recent_scout_activity the world read_thread read_record fetch_url handed URLs only read_network PR open making render_draft pixel studies Checks, in order caps per run self-initiated whitelist group rules pictures · per day threads · per day profile · pins · per day site · leak · per day group replies per hour a dropped action is recorded, never sent The record · D1 runs · run meta · stream: every step, each network read named by URI and CID journal: what Scout said and why, hash-chained effect ledger · harness_control routing registry: human-owned, read-only here Vectorize index behind recall_memory Channels Gate and queue think() Checks Out The timeline check, every 15 min, feeds the reflection digest. It isn't a run. checked first a run one model serves a run, start to end escalates what passed every step, a line memory writes
Every run takes the same path, left to right. An escalation leaves the path before any model sees the message. Whatever a model answers, the checks and the effect ledger decide what actually goes out.

A run, in order

  1. A channel turns something into an encounter: a Bluesky notification or DM, a group message, an email, or one of Scout's own lanes.
  2. The kill switch is read first. While it says halted, nothing is polled and nothing runs.
  3. The run queue takes it. Runs in a lane go one at a time, one step per alarm, and a step that stalls is fired again.
  4. Triage asks Workers AI whether it's worth engaging. Escalation is checked before that verdict can end the run.
  5. On an escalation, the message is held for an attended session with Jason, and no model, free or paid, answers it.
  6. Otherwise Scout gets its prelude and one model for the whole run, and works through up to 64 steps of tools.
  7. The answer's actions pass the checks in order. A dropped action is written down, never sent.
  8. What passes is claimed in the effect ledger, then sent, so a send happens at most once.

What Scout decides

  • What to say, and whether to say anything. Most reflections end in no action.
  • Which sources fill its prelude (workspace/prelude.json). The menu and the caps are code.
  • When to wake itself next (a continuation).
  • Its own memory, written through writeMemory(). Each write is a commit, files about people only grow, and routing/ and harness/ are refused.
  • Halting the harness. Scout can stop everything; it can't start it again.

What only a person does

  • Resume after a halt (scout-admin's control page).
  • Drain PENDING_WRITES, attended, with memory-drain.mjs.
  • Edit the routing registry, the human-owned facts that only ever add reasons to escalate.
  • Merge a change to anything drawn in orange. Workers Builds deploys on merge, and a failing test blocks the deploy.
  • Answer an escalated message.

Drawn from scout-harness main at cc75091 on 2026-11-20, plus read_network, which is on an open branch. Scout's own patches were proposed in the group the same day. The Sonnet arm is an experiment on Jason's DMs only; the switch it reads is SONNET_ARM. The style borrows from the Gladys architecture diagram.