Ensure /new and chat close stop native foreground and background work before changing registry state. Preserve transcripts while surfacing release and recovery cleanup failures instead of reporting false success, with focused adapter and server regressions.
A model stage with harness=claude, provider=anthropic and a Claude model id
runs one claude -p process per visit with the server user's Claude sign-in.
Only the declared tools exist (--tools) and nothing else is approved
(--permission-mode dontAsk); edits are approved only below the run directory
and the stage's artifacts directory. A permissions block, run write paths,
delivery=true and reasoning_effort none/minimal are refused, not widened.
Usage comes from the CLI's final result and counts toward budgets. Failures
are fixed text; billing, quota and rate-limit refusals map to
ErrProviderQuota. The redacted stream is the session export and evidence.
provider-check --harness claude and model overrides are covered.
Ten runs each burned five visits in seconds on "OpenCode agent reported an
execution error": the provider answered HTTP 402 provider.quota and the
classifier read neither field. That refusal is now its own error with fixed
text, is not retried inside the stage, and ends the run instead of
following the error edge.
A failed visit was recorded on its step only, so a landing gate that sent
the run back to implement gave the implementer nothing about which checks
failed. The diagnostic (bounded) is now the next stage's previous output
and a memory entry.
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.
A new elevate request verifies the web sign-in password (scrypt hash, read-only,
rate-limited, constant time) and elevates that connection for 15 minutes.
chat_shell_run/input, rename/move/delete, terminal_open/input and POST /files
(via a per-connection upload token) are refused with code elevation_required
until then.
The server answered any Host, so a DNS-rebinding page reached /connect (101)
and could drive terminals. Every route now refuses a non-loopback Host on a
loopback connection (421), /connect refuses any Origin (403), and a server
started with -token-file requires X-Pipelines-Token (401; /mcp and /healthz
exempt). The CLI/TUI send the token only to loopback URLs.
A new terminal runs su -l for the server's own account in its PTY, so the
operator types the account password before getting a shell. Opening is
refused when su is missing, the server runs as root, or su opens a shell
without asking. The Pipelines-owned tmux backing, its reattach and its
persisted session names are removed: terminals end with the server and
are listed as ended. A starting server ends leftover pipelines-term_*
sessions on tmux -L pipelines only.
A dismissed run keeps its command, status, times and output within the
usual history bounds. chat_shell_runs lists dismissed runs apart
(chat_shell_dismissed_runs), chat_shell_run_output still reads them, and
the purge timer and startup purge are gone. Rows an older server purged
are listed with purged set and cannot be restored.
chat_shell_dismiss hides finished runs from every readback and pushes a
dismissal frame to every watcher; chat_shell_restore undoes it within a
30 s window, after which the command, output and error are purged and only
a tombstone remains, reported by chat_shell_runs for clients that missed
the event.
Finished terminals are sized synchronously from their rows at the view's
width (the post-frame fit made the transcript re-anchor forever); blank rows
xterm.dart keeps when rewrapping are dropped; desktop typing uses key events;
the composer leaves Cmd/Ctrl-V to a focused terminal; Ctrl-A reaches the
process. The probe fixture gains coloured terminal runs and an interactive
login. Notes record behaviour, protocol, security and verification.
Codex: model/list supportedReasoningEfforts per model, turn/start effort,
thread reasoningEffort as the current level. Claude Code: initialize
supportedEffortLevels, apply_flag_settings effortLevel at the next turn
start. OpenCode v2: model variants via the session model endpoint.
chat_effort sets a level; rows and chat_models carry the levels; Codex and
Claude levels persist in operator_sessions.chat_effort and are re-applied
before every turn.
Commands start in a PTY (creack/pty; macOS and Linux) of the client's size
with TERM=xterm-256color in their own session. chat_shell_input writes typed
UTF-8 text in order through a bounded queue; chat_shell_resize sets the size.
Output stays the raw terminal stream with byte offsets as sequence numbers,
chat_shell_run_output resumes from an offset, and a plain-text rendering is
persisted for Copy output and Send to agent. Input is never stored or logged.
Snapshot the chat directory with internal/filediff when a turn starts and
compare it when the turn ends; store changed turns (newest 100 per chat),
push chat_turn_diffs summaries, attach them to chat_messages replies and
serve one turn's patches through chat_turn_diff.
chat_shell_run starts $SHELL -lc in the chat session's directory in its own
process group, streams merged output to every watcher of the session, keeps a
1 MiB tail and persists each run; chat_shell_stop interrupts the group and
kills it after a grace period. chat_shell_share queues an ordinary prompt
with the command and a bounded output tail.
The chat Files panel needs to rename, move and delete entries. Three
websocket requests, dispatched beside list_directory, operate on one
entry inside the operator's home and reply with path_mutated {path} or
an error frame (does not exist, already exists, is not empty, outside
the home directory, cannot be moved into itself).
The entry's parent folder and the move destination are resolved through
symlinks and must stay inside home; the entry's own symlink is never
followed, so a link is renamed, moved or removed as a link. The home
directory itself, anything outside it and pseudo filesystems are
refused. A folder is deleted only when empty unless recursive is set;
os.RemoveAll unlinks symlinks rather than descending into them.
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.
The handler reuses the home-path resolution of GET /files, refuses special
filesystems and non-directories, writes with O_EXCL so an existing name is
answered 409 instead of overwritten, enforces the chat upload size limit and
replies with the stored name, path and size.