Rewrite format.md around turn files, the output.md order (# outcome, edge word, # handoff, detail), checks, configs, human stages and the land stage delivery gate (sh -c in a fresh detached worktree, clean checkouts incl. untracked, 64 KB output cap, opencode only). Align review guidance across practices, verification, orchestration and the feedback-loop example. Fix loop-limit semantics in control.md against runner.go, budget_tokens sizing, read for large inputs, blocked-else-error and one stage naming set. Move dated verification logs to docs/verification-history.md and brokered chat, TUI keys and harness internals to docs/operator-reference.md; label operator-only references in SKILL.md.
On a large Git workspace the 32 MiB retained-text budget was spent
alphabetically on files that never change, so small late edits were
recorded with no_patch. The stage-start snapshot now keeps text only for
files dirty or untracked at start; a clean tracked file keeps its hash and
stage-0 index blob id. When it changes or is deleted, the after capture
reads that blob in one `git cat-file --batch` and uses it only if its
SHA-256 matches the start read, so stage commits/resets do not move the
baseline and filters or assume-unchanged never yield a wrong patch. The
after capture shares unchanged files' representation instead of spending
budget on them again.
FileChange gains no_patch_reason (binary, file_size, budget, lines,
evidence_size, git_status, baseline). Non-Git workspaces keep their
behaviour.
OpenCode and Codex chats implement Detach (called by chatManager.Close)
instead of a parallel DetachSessions interface. Docs follow the Claude Code
host and the systemd KillMode=process unit. The opt-in live restart check
copies only the stage provider's login row into its scratch OpenCode
database, builds the CLI for skill install and removes its scratch tree.
A stopping server no longer interrupts harness sessions. Runs are canceled
with a shutdown cause, so stage clients leave admitted sessions running and
owned; start-up recovery re-adopts live OpenCode V2 stage sessions (wait,
read back the result, continue the graph), restarts unadmitted visits,
continues runs between stages, and falls back to interrupted with a reason
when the harness is gone. Chats with live work are recorded on shutdown and
reattached on start; turns that ended during the downtime are folded in
from history. The server-spawned V2 chat serve is now a detached owned
service. PIPELINES_SHUTDOWN_SESSIONS=stop restores the old behaviour.
A pipelines-claude-host process owns each claude -p CLI and its pipes,
spools its stream-json output with offsets and serves a private Unix
socket. The server detaches on shutdown and reattaches on start,
replaying the spool to rebuild turns, approvals, questions, background
tasks and the live model. Explicit close, idle release and CLI exit stop
the host; a crashed host falls back to --resume with a notice.
Stages now run on a detached Pipelines-owned opencode serve started with
OPENCODE_CONFIG_CONTENT carrying stage tool_output (16384 bytes / 400 lines)
and compaction.auto defaults, configurable through PIPELINES_OPENCODE_STAGE_*.
Chats keep their existing service and configuration; a server-spawned chat
serve can take PIPELINES_OPENCODE_CHAT_CONFIG. The service shares the user's
OpenCode config, credentials and session database, keeps a stable private
registration so retained references (recovery, interrupt, export) stay valid,
revives on the same port after a crash, replaces itself when idle after a
settings change, and is stopped on shutdown only when idle.
PIPELINES_OPENCODE_STAGE_SERVICE=shared restores the previous behaviour.
run.sh now records flutter's JSON results beside the log and prints a
summary from tool/memwatch/summary.dart: one verdict line (pass/fail/skip
counts, peak MB, duration, exit status), only failing tests with their
error and app stack frames (framework frames dropped, per-test and total
caps, tail of printed output), the tests still running when killed, and
the log and JSON paths. Quiet is the default when stdout is not a
terminal; MEMWATCH_OUTPUT=full also prints the whole log. The watchdog
loop and exit status are unchanged.
tool/analyze.sh bounds flutter analyze to a verdict line and the first
40 issues. Tests: test/tool/memwatch/summary_test.dart and the
end-to-end tool/memwatch/quiet_test.sh over fixture suites.
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.
The server skip fallback for stages with a skipped edge now applies only
when every earlier visit recorded a complete file-change capture with no
changed files and no image artifact. Any change (no repo-specific UI path
or extension list), any partial/unavailable/missing capture, a human
decision or a checkpoint-retry prefix runs the model. auto_skip=false opts
a stage out. The fallback verdict names the server fallback.
Capture compares over-budget text by hash and, in Git, lists paths it
could not compare when their git status changes, marked no_patch, so a
large workspace no longer records files: [] for a real edit. Run detail
gains a warnings list for partial/unavailable captures, also printed by
inspect on stderr.
System default plus Inter, IBM Plex Sans, Source Sans 3 and Atkinson
Hyperlegible Next (OFL 1.1, bundled as on-demand assets and loaded with
FontLoader, so the web fetches only the previewed/chosen files). Stored in
shared_preferences, applied live to Material and Cupertino text themes;
RobotoMono styles unchanged. Licences registered with LicenseRegistry.
The list applies chat_activity pushes and each row's activity by seq, so
every listed chat shows a running turn, not only the chat open here. An
idle chat with running background tasks wears a quiet badge with tooltip and
semantics "Background tasks running (N)". Older servers keep the old
behaviour. List rows that never changed carry the server's current seq so
they order after snapshots kept from an earlier server process.
chat_sessions rows carry activity (turn_running, background_tasks, seq) from
the hub's own state, and chat_activity is pushed to every connection when a
session's turn starts or ends or its running background-task count changes.
A room whose last viewer left stays observed while background tasks run,
bounded by the unattended limits; a Claude Code process exit ends its tasks.
The template catalog is now Example 1-4: one stage, two stages (handoff),
feedback loop (runner-owned limits, retries, timeout) and human decision
(type=human approve/revise/abort with notify hooks). All read-only or leave
changes uncommitted; none commits, pushes or deploys.
A brand-new database (schema created by this Open) gets the examples once at
server start, bound to the configured default model and harness. A
server_settings marker records the decision, so removed examples never come
back and existing installs are marked without any pipeline being added or
rewritten.
Full screen button in the viewer's bottom bar opens a root full-screen route
with the same bar and top bar; exit by the bar, Close, Escape or back. It
renders the already fetched bytes, never the system browser. The external
Open action is now labelled Open in browser.
Native web views claim their gestures (vertical drags inline, every pointer
in full screen) so the page swipe no longer starves them. On the web, full
screen adds zoom buttons that scale the live sandboxed frame. Links: in-page
anchors scroll, http(s)/mailto open outside, everything else is inert.
Capture kit: --touch, swipe=, pinch=, tabs=; html=1 fixture artifact.
The selection area claimed a primary mouse press on a link as a selection
drag once the pointer moved 2 px, so in the browser ordinary clicks with a
little hand or trackpad movement opened nothing. A link's recognizer now
accepts a primary mouse press on pointer down, as an anchor does; the click
still cancels past the tap slop, drags from plain text still select across
links, and touch long press keeps Copy link.
Only visible assistant reply text, a pending form or an unresolved
permission request is unread. Tool results, user prompts, harness
notices/agent reports and blank deltas only drive the working state.