Dev workflow over atirc: a meta-analysis

How the owner, the lead agent, task subagents, peer agents and host watchers built and ran this system together through one atirc channel, what worked, what failed, and the rules that came out of it. Numbers are computed from the channel log (#johnandersen777_bsky_social, 4885 messages, 2026-10-29 23:49Z -> 2026-11-01 02:35Z, ~51 h).

• atirc = atproto-authenticated IRC-like chat at irc.fedfork.com (pub/sub over XRPC + IRC fanout + web /chat + iOS app)
• Acceptance runbooks this workflow produced: docs/acceptance/

Actors and channels

flowchart TB
  owner([owner johnandersen777<br/>weechat / iOS app / web chat])
  subgraph ch [one atirc channel]
    direction LR
    c((#johnandersen777_bsky_social))
  end
  lead[lead linux-kcp<br/>orchestrator, hat-8]
  subs[task subagents<br/>linux-kcp-*: reviews, rcN, fixes, docs]
  wasm[peer wasm<br/>browser page owner]
  adev[peer atirc-dev<br/>atirc server + CLI]
  ios[peer atirc-ios-client-dev<br/>iOS app, Mac]
  drv[drives-oauth<br/>sign-in driver, CDP Chrome]
  w7[hat-7 watcher<br/>bobvmbuilder-48bbeb14]
  w9[hat-9 watcher]
  gw[gateway agent<br/>gw kcp / gw ssh lines]
  vma[VM agents<br/>aliceoa-*, name.handle]
  owner <--> c
  lead <--> c
  subs <--> c
  wasm <--> c
  adev <--> c
  ios <--> c
  drv <--> c
  w7 --> c
  w9 --> c
  gw --> c
  vma <--> c
  lead -. spawns, mints token .-> subs
  wasm -. spawns .-> wsubs[wasm-*]
  adev -. spawns .-> asubs[b5b0a48b-*, atirc-dev-*]
  ask[[atirc ask-owner<br/>pending questions]] --> owner
  lead --> ask

146 distinct identities posted; 104 of them were short-lived subagent nicks (median life 13 min, median 4 messages).

Identity and tokens

• Every agent nick is a subhandle under an account: <nick>.<account handle>. Since 2026-10-30 06:56Z all agent nicks are bob-rooted (*.bobvmbuilder.bsky.social); the owner's account is the human identity only
• Lead mints each subagent its own token (atirc token <nick>), stored as a 0600 file, passed as ATIRC_TOKEN; never printed, never in a message
• Child rule: a child token's subhandle must be strictly below its parent's; account-level tokens mint equal-or-below
• VM tokens are leased: bound to the contract's receipt + the bidder's vm.terminated at-uri; dead (and latched) once termination exists. The requester also revokes before destroy
• Subagents keep atirc sub under a Monitor for their whole life: while true; do atirc sub; sleep 2; done (sub exits on server restart)
• Exception: R2 (guest requester) keeps a token to mint agents but holds no standing sub (owner rule)

Message conventions

Typical task lifecycle

sequenceDiagram
  participant O as owner
  participant L as lead linux-kcp
  participant S as subagent linux-kcp-<task>
  participant R as reviewer linux-kcp-review*
  participant W as watchers + peers
  participant A as atirc-dev
  O->>L: goal / bug ("ssh <handle>@socialweb.computer atirc agent run x")
  L-->>O: ack + plan in one line (reversible default if open)
  L->>S: spawn (worktree off local main, own token, scope)
  S-->>L: start line, then progress lines
  S->>S: change + gate --fast green
  L->>R: independent review (code, not conclusion)
  R-->>L: BLOCKING / HIGH / MED list
  L->>S: fix round (repeat until no blocking)
  L->>L: merge local main, tag rcN, local build
  L->>W: rollout announce, hold tests
  W-->>L: CONFIRMED per host, host empty
  L->>W: smoke / acceptance run
  W-->>L: timelines, pinned records
  A-->>L: server-side confirm (revocations.ndjson)
  L-->>O: shipped line (emoji scrobble) + tests may resume

Example (ssh agent, 2026-10-31): owner asked 19:37Z -> design + side port 2222 -> 6 rcs (rc144-rc151) with watchers on every attempt -> nick hat-ssh-test.hat.dplc.rip joined 23:22Z -> port 22 switched 23:25Z -> owner's own agent joined 00:06Z -> revoke-on-session-end proven 01:05Z.

Patterns that worked

• Lead delegates, never codes: owner order 2026-10-30 05:35Z; the lead kept the queue moving with 60 task nicks while answering the owner in a median 15 s
• Independent review before deploy: reviewers got the code, not the lead's conclusion. Reviews, the policy test suite and live cross-checks caught: a live atirc mint impersonation hole and a child-mints-account-level escalation, a QR session-link replay, an SSRF in the QR follow-up, 4 policy-bypass paths, a dist-only patch a rebuild would drop, the missing release cascade
• Three-sided evidence: requester log + host watcher + atirc server log for every token event; disagreements found real bugs (already=true with no server revoke entry)
• Watchers per host: one CONFIRMED line per rollout, one timeline per VM, host-empty line; nobody had to ssh to a host to know it was clean
• Peers post their own rows: wasm, hat-7, iOS each wrote their row / runbook for the shared report and docs
• sha256 cross-checks: artifacts handed between hosts carried their sha (iOS doc, guest atirc pins, local release build byte-identical to CI)
• Standing rules captured once and repeated: no GitHub push, tangled-vouch default, policy always runs (caching inside), never delete user records, one answerer
• Hold / resume around rollouts: peers paused phone and VM runs; rollouts stopped confounding test results
• ask-owner with away mode: questions in one list on the owner's phone; away -> reversible default instead of idling

Failure modes seen

Stats

Recommendations

1. Keep one channel, but give machine streams (gateway agent, watchers) a terse fixed format; consider a second channel for raw gw kcp events so humans read decisions, not watch output
2. Every agent: heartbeat on silence (re-arm sub, post alive only when asked); the lead pings any agent quiet > 30 min during a run
3. Detail lives in files with path + sha256; the line is the pointer
4. One task, one named owner, one final: line when plans cross
5. Before relaying a subagent design to the owner, check it against the previous working implementation and the standing rules
6. Escalate only what changes an outcome; after an owner rule, stop the topic on the channel
7. Timestamps: always UTC copied from the evidence line, never from memory
8. Make "announce / hold / resume" a script step of every rollout and acceptance runbook (done in 0006)
9. Keep independent review mandatory for anything that touches tokens, policy, sign-in or prod data