A fresh run started from a repository's main checkout gets its own
worktree and branch next to the repository, so runs never share a tree
or see the operator's uncommitted work, and several can run at once.
Retries and runs started in a linked worktree keep their directory. A
clean, landed run tree is removed after success; anything else is kept.
The delivery gate commits what the land stage left uncommitted in a run
worktree, and no longer requires the operator's main checkout to be
clean: Git's fast-forward refuses safely when a local change conflicts.
The proxy forwards an HMAC-derived session id (X-Pipelines-Session) on
/connect and POST /files; the server trusts it only on token-authenticated
requests and keys elevation by it, so reconnects stay elevated up to a 24 h
cap. Terminals rely on their su -l login. The app no longer forgets the
elevation on disconnect and asks again only on a fresh refusal.
Each installation generates a private Ed25519 key once (0600, under
~/.config/pipelines), signs version.json with a detached version.json.sig
when publishing APK and desktop builds, and embeds its own and listed
trusted public keys in app builds. The app verifies the feed signature
before trusting the artifact sha256, refuses unsigned or foreign feeds when
it has keys, and requires the sha256 and Android signer in signed feeds.
The mDNS record now carries ts= (tailnet IPv4 addresses) and public= (the
installer's HTTPS origin from PIPELINES_WEB_PUBLIC_URL) in TXT, and browse,
the discover_servers RPC and /discovery return every route to a machine as
addresses with kinds lan, tailscale and public. Endpoint stays the LAN route
for older clients. Unsafe ts/public values are dropped.
Every stage used to be told to read memory.md, the full output of every
earlier turn (19 KB average, 92 KB max), and paid for it on each model call.
memory.md is now essential memory, hard-capped at 4 KB and inlined in the
prompt: the task, the previous stage's # handoff section (or its truncated
output head), a one-line-per-stage outcome ledger and the server-captured
changed files. history.md keeps the old full memory on disk; read-capable
stages are pointed at it and input.md for on-demand detail, and it is never
inlined. The prompt asks each stage for a short # handoff section.
Claude Code records subagent/peer reports, task notifications, reminders and
skill bodies as user-role entries. Classify them server-side (history and live
stream) by origin/isMeta/queued_command mode: reports become kind agent_report,
task notifications kind notice, harness meta is hidden, and the mid-turn
wrapper is unwrapped to the operator's text. The app draws a collapsed Agent
report card and a one-line system notice on the agent's side.
When canonical main is not an ancestor of the item HEAD, the gate now
merges main into the item workspace under the delivery lock, re-runs the
declared checks on the merged commit, and fast-forwards main to it. A
conflicting merge is aborted with the workspace verified unchanged and a
diagnostic naming the files. Successful catch-ups are logged.
Later visits' input.md held only the previous stage's output, so the
operator's task reached only the first stage; when a Prepare stage
summarised (paths, HEAD), Code never saw the task or its acceptance
checklist and either blocked (run #202) or did part of it (#195, #197,
#198). Later inputs now carry the run prompt under '# Run task' (headings
demoted) before '# Previous turn output'.