A stage's final session write can land just after the step settles. Retry the
read-only native export once after 200 ms when its envelope is invalid or names
another session, so the stage records session.html instead of dropping evidence.
Fall back to Homebrew and local install locations when opencode is not on the
service's minimal PATH, so a launchd/systemd restart cannot fail a run with
exec: opencode not found.
Expose the OpenCode model list and session model switch over the chat protocol,
carry a per-session model provider on chat rows, and apply a preloaded behavior
brief as a durable session instruction entry when a chat is created.
Carry browser files as base64 prompt attachments, write them to private
temporary files and pass file:// references to OpenCode, and map the user
turn's top-level files back into history. Replay a running step on subscribe,
forward only a caller-supplied msg_ id, and persist cancel on the run and live
step before interrupting the runner.
Remove authoring from the skill description and move the authoring commands
and examples behind an explicit-request section in help. A skill test keeps the
default description from advertising authoring again.
Use existing pipelines for run requests and put task specifics in the run
prompt. Creation, edits and publication now require an explicit user request;
the authoring skill test enforces that guidance.
Add the Run pipeline dialog with prompt, server directory and a per-run
notification choice; add --notify auto/on/off to run and retry; move the live
tool call onto the tool-count line and tick the reasoning duration; stop chat
frames from repainting a different screen.
Store a chat's harness with its operator session, expose chat_backends and a
backend choice in New chat, and route each session through a manager. The Codex
adapter uses the App Server's Unix WebSocket for threads, turns, history,
questions and interrupt; OpenCode remains the default and Claude is not
advertised until its sidecar and login are wired.
Raise phone body/meta type, controls and dock targets to 44/56px, make the chat
back arrow a visible 44px control, and give the star/menu 44px hit areas without
covering the chat status. Keep the shipped skill free of personal device and
project capture paths.
Enter inserts a newline and Ctrl/Cmd+Enter sends; slash commands still run on
Enter. Chat renames are stored in the server registry so every device sees the
same title, with a one-time upload of old local titles. The star no longer
covers the row status, and shell frames patch the background bar in place so the
composer stays mounted while streaming.
- Persist chat stars in the operator-session registry (chat_star) so every
browser/device shows the same Starred group; migrate old local stars once.
- Add a compact pipeline graph shape to list responses and draw a lightweight
list miniature; replace the cryptic id/version line with the stage count.
- Surface OpenCode form questions as a chat card (chat_answer), replay a
pending form to late subscribers, and notify for background questions.
- Make the mobile dock's Search cell equal to every destination, keep the
tab row visible with 100svh, and let Source use the full pane height.
- Let a fresh sign-in auto-connect to its own origin; use a Lax session cookie
with Expires so Remember me survives app/link launches.
- Keep shipped starters and the skill agnostic: no repository paths, project
areas or personal device commands; add Diagnose/Fix and Research starters.
- Refresh the installed skill and document the new browser conventions.
The one-line verdict summariser expects the old field map; given a turn's single
markdown body it treated the result as untrusted reference material and produced
a rationale unrelated to the work ("The text was treated as untrusted..."). Turn
outputs now keep the honest absent-reason state: the card still shows the option,
and no model call is spent inventing a reason.
Turn bodies were embedded verbatim, so a body's own ## headings looked like
sibling turns and could even duplicate the system's "Files added to public/"
line. Body headings are now demoted two levels (capped at H6) outside fences;
fenced code is untouched.
A probe stage read "public/" in its task, created a workspace-relative public/
folder, and also copied to the run's shared path. The prompt now states that the
absolute path is the shared folder and a relative public/ is a different,
unshared directory.
ParseOutput returned only the text after the outcome value, so a stage that wrote
its result first and "# outcome" last handed the next stage "(no output)" — the
probe's Compose and Critique turns were both lost this way. The body is now the
whole document with just the outcome heading and value line removed, in order,
byte for byte. Tests cover a trailing and a mid-document outcome.
The successful retry in TestDefaultRetriesAndTimeoutResetEachAttempt was
racing its own 20ms deadline when the full suite ran in parallel. Keep the
two-attempt behavior and give the retry realistic headroom.
Each model visit now gets <run>/stages/<seq>-<stage>/ with input.md,
memory.md, output.md, private/ and a shared run-level public/. The model
reads the two input files, follows the stage's instructions verbatim and
returns complete markdown; only the # outcome section is parsed.
- add internal/turn layout, render and output parsing
- write input.md/memory.md before a visit and output.md from the final
message; inventory all turn files as stage artifacts
- accumulate completed turns and system-written public file additions in
memory, including across retries and resume
- remove the JSON envelope, strict field validation, line format and the
repair ladder for model stages
- republish pipeline authoring prompts and docs in the new style
Checks: go build ./..., go vet ./..., go test -count=1 ./..., gofmt -l .