The netstat-based parser matched the literal English "LISTENING"
state string, which is translated on non-English Windows (e.g.
"ABHÖREN", "ÉCOUTE", "ESCUTANDO"). On those systems the regex matched
zero lines, so killProcessOnPort silently did nothing -- fail-open,
not a regression, but ineffective for the exact users the fix targets.
Replace it with `Get-NetTCPConnection -State Listen -LocalPort <port>`,
which reads the same underlying WinNT API netstat's display layer
translates, so it's unaffected by OS display language. Verified
against a real listening port on this machine (matched the actual
owning PID).
Also fixed a stale duplicate of the "killProcessOnPort is a no-op on
Windows" comment left behind in server/index.js.
killProcessOnPort() was a no-op on win32 (POSIX-only, via lsof/kill),
so a restart could leave the old OpenCode process holding the port
while a new instance spawned on a different one. That's a plausible
contributor to a chronic pattern seen in production logs: repeated
"OpenCode process exited, restarting" cycles and hundreds of
ECONNRESET/proxy errors over multiple days on Windows.
Give killProcessOnPort a real Windows branch: parse `netstat -ano`
for PIDs listening on the target port, filter out our own pid, and
force-kill each via `taskkill /PID <pid> /F` (no /T -- we don't own
that process, so only the listener itself is killed, not any
children it may have).
waitForPortRelease()'s existing soft-fail-and-warn behavior is left
untouched -- it's a deliberate safety net for any platform where the
port doesn't free up in time, not just Windows, and the restart
already rebinds event-stream readers to the actual resulting port via
onOpenCodeRestarted.
MCP authorization was broken in several stacked ways. The browser return
leg landed on the SPA behind the auth gate, so the system browser saw a
login page instead of finishing; the pending-context store silently
saved nothing because its route had no JSON body parser; and the
callback-URL config write started deferring behind Apply & Restart, so
authorization ran against a runtime without the URL and dead-ended on
OpenCode's loopback listener.
The return leg is now completed entirely server-side by an
unauthenticated GET /mcp/oauth/callback that only forwards a code whose
state matches a parked context. Desktop with the local server and VS
Code switch to OpenCode's native flow over its fixed loopback port —
no config writes or restarts at all, with a one-time cleanup of the
previously written callback URL — and its completion signal drives the
page instead of blind status polling. Remote, hosted-web, and mobile
keep the server-callback flow, applying a queued callback-URL write
immediately since authorization cannot wait for a manual restart.
Also: a server queued behind Apply & Restart now shows an Awaiting
restart badge and explanation instead of connect/reauthorize buttons
that can only fail, and Reauthorize is offered only while the server is
actually connected.
Adds a work-status panel beside the transcript. Context fill, model and
cost, todos, running subagents and the permission requests blocking
them, branch and working-tree state, MCP servers, pinned messages and
context sources were scattered across the header, the composer and the
context panel — a blocked subagent was reported nowhere at all. The
panel reads them from live channels rather than persisted history, and
becomes an overlay where the chat is too narrow to seat a column.
It is on by default, including for existing installs. Because it now
carries these readouts, the desktop header and composer drop the ones it
duplicates: todo and changed-files chips, usage and MCP tabs. VS Code
and mobile keep theirs — neither hosts the panel.
Fixes MCP authorization, which was broken from the panel, invalidated by
a directory switch through a redirect URI that encoded the working
directory, and left the desktop app in the background because browsers
will not follow a custom-protocol link without a user gesture. The
settings page no longer asks the user to understand the MCP spec before
adding a server: one field takes the command or the link, with the kind
inferred and a visible override, and client-registration fields appear
only when a server actually asks for its own credentials.
Also: skills load from the panel instead of only when the composer's
slash autocomplete opens; the header button names the current instance
rather than falling through to the word "Instance" for relay hosts.
Three new optional UI settings keys, all migrated. No change to stored
MCP server configuration.
Show the running instance's local server URL and tunnel URL (when a
tunnel is active) as labeled, click-to-open buttons on the About page.
/api/system/info now reports the instance port and tunnel URL, resolved
lazily from the tunnel runtime so each Git-worktree instance identifies
itself in the UI without parsing terminal output.
Refs OPE-194
When the managed OpenCode process exits but a server survives on the old
port (Windows: killProcessOnPort is a no-op, so the orphaned process tree
keeps the port), restartOpenCode() times out waiting for the port and
spawns a fresh server on a NEW port. HTTP/proxy traffic follows the new
port, but the global message-stream hub's upstream SSE reader stays pinned
to the old server's /global/event stream — that connection never closes —
so new events never reach the UI and the chat stops updating until the
app is restarted (#2638).
Lifecycle now fires an optional onOpenCodeRestarted hook after a
successful managed restart; index.js wires it to the new
messageStreamRuntime.rebindUpstream(), which restarts the shared hub
(its reader re-dials buildOpenCodeUrl → the current port) and closes
directory-scoped sockets so their per-connection readers rebuild against
the new port. External servers are untouched (their port cannot change).
Fixes#2638
parseMdFile now matches gray-matter (used by OpenCode) for file shapes
OpenChamber previously failed to parse: frontmatter whose closing '---'
sits at end-of-file without a trailing newline, a UTF-8 BOM prefix, and
YAML with unquoted colons in scalar values (via the same sanitizer
OpenCode applies). OpenCode parses these files, so OpenChamber must
too: otherwise the whole file was treated as the prompt body and a
save rewrote the existing YAML block into the body, prepending a
duplicate frontmatter block.
Refs OPE-178
The env var was already read and passed to the managed OpenCode server
spawn, but any non-empty string was accepted. Reject values that are
not a valid IP (IPv4/IPv6, brackets allowed) or DNS-style hostname with
a clear [config] error and fall back to the secure loopback default so
a typo can never silently bind a non-loopback address.
Refs OPE-231
OpenCode's authorize response reports how the client must finish: `code`
expects a pasted code, while `auto` requires the client to call
oauth/callback immediately and hold it open — upstream blocks in there
polling for the device code or waiting on its loopback redirect, and only
that call persists the credential. Every auth plugin OpenCode ships uses
`auto`; none use `code`.
The page implemented only `code`. It opened the browser, showed a paste
field no provider can fill, and never called back, so a successful sign-in
stored nothing and the app sat unchanged. Authorization now drives the UI:
`auto` chains straight into the callback behind a waiting state with a
cancel, and the paste field appears only when a provider actually asks
for a code.
Two smaller failures shared that surface. Prompts were never collected,
which put GitHub Copilot Enterprise out of reach entirely, so a method
that declares them now asks first and passes the answers to authorize.
Device codes are also recovered from the instructions text, where they
actually live — the old code read fields the API does not return, so the
copy button never appeared.
The callback is exempt from the ordinary proxy deadline and gets a
15-minute budget, bounded by the shortest upstream expiry we know of.
A human sign-in with 2FA does not fit in four minutes, and expiring it
turned a completed login into a 504.
Resolve ProvidersPage and lifecycle conflicts with custom providers and
AppImage ARGV0 stripping. Address review follow-ups: OAuth index helper +
tests, single auth-methods load trigger, shared Google env-alias module with
VS Code parity coverage, and deferred restart for custom provider upsert.
Co-authored-by: Serhii Dziupin <makeittech@users.noreply.github.com>
OpenCode can mark Google connected via GEMINI_API_KEY while the Generative AI
SDK only reads GOOGLE_GENERATIVE_AI_API_KEY, so chat asked for a key that was
already present. Alias unset sibling names on managed launch for web and VS Code.
Co-authored-by: Serhii Dziupin <makeittech@users.noreply.github.com>
Add config-mutation-response.js to owning docs, describe restartDeferred
and external requiresManualRestart on /api/config/reload, and remove the
unused buildConfigMutationResponse export.
Co-authored-by: Serhii Dziupin <makeittech@users.noreply.github.com>
Use getRequestDirectory and x-opencode-directory like the other skill
mutations, and pin renamable list/store mapping with focused tests.
Co-authored-by: Serhii Dziupin <makeittech@users.noreply.github.com>
Keep ARGV0/env-u assertions from this branch and the DA startup-reply
expectations from main's terminal PTY-before-viewport fix.
Co-authored-by: Serhii Dziupin <makeittech@users.noreply.github.com>
Resolve skills store conflicts by keeping repository-local directory
resolution from main and deferred-restart handling from this branch.
Co-authored-by: Serhii Dziupin <makeittech@users.noreply.github.com>
Expose authoritative renamable on skill list responses using the same
managed-root policy as renameSkill, drop the divergent UI path heuristic,
and remove an unused rejection-test fixture.
Co-authored-by: Serhii Dziupin <makeittech@users.noreply.github.com>
bun-pty merges the OS environ into PTY children, so deleting ARGV0 from the
JS env object alone left the AppImage path in the shell. Wrap Linux PTY
spawns with env -u ARGV0, clear native ARGV0 under Bun via libc unsetenv,
and always clear process.env even when no login-shell snapshot exists.
Co-authored-by: Serhii Dziupin <makeittech@users.noreply.github.com>
AppImage exports ARGV0 into the process environment. zsh treats that as
argv[0] for every external command, which broke Python venv detection in
the integrated terminal and managed OpenCode sessions.
Clear ARGV0 in Electron before login-shell probing, refuse to re-apply it
from shell snapshots, and strip it from terminal PTY and managed OpenCode
launch environments.
Co-authored-by: Serhii Dziupin <makeittech@users.noreply.github.com>
Restrict in-place skill rename to managed skill directories, require
frontmatter name to match before moving, roll back/reject with tests,
hide rename in the UI for unmanaged paths, and drop unused toast keys.
Co-authored-by: Serhii Dziupin <makeittech@users.noreply.github.com>
Derive the effective OpenCode config layer (custom > project > user) from
provider sources and send it on PUT /api/provider so project/custom edits
update that layer instead of creating a global user override. Resolve
OPENCODE_CONFIG at call time and add UI/web/VS Code coverage for scoped
upserts.
Co-authored-by: Serhii Dziupin <makeittech@users.noreply.github.com>
Rename skills by moving the skill directory and updating frontmatter
name instead of recreate-with-stub-description, which wiped the body
and supporting files.
Co-authored-by: Serhii Dziupin <makeittech@users.noreply.github.com>
Defer OpenCode reloads after settings mutations, track pending changes,
and expose a top-right Apply & Restart OpenCode action with a counter so
sessions stay available until the user explicitly applies.
Co-authored-by: Serhii Dziupin <makeittech@users.noreply.github.com>
A diff is ordered by file path, which is almost never the order in which a
change makes sense. This adds a Walkthrough surface that reorders it: the model
groups related hunks into stops, explains what each group changes about
behavior, and orders the stops so each builds on the last. It explains and
orders; judging code stays with the existing Review action.
Reviews uncommitted work (all, staged, unstaged), a branch against its base, or
a pull request. Generation is always user-initiated — nothing runs on a timer,
on a file change, or as a side effect of opening a panel.
Invariants worth preserving:
- Hunk identity is derived on the server and only there. Ids are content
hashes, so an anchor that no longer resolves is proof the code it described
changed, and staleness needs no heuristics. The client matches ids to ids and
never recomputes them; two implementations would have to agree forever.
- The digest is never truncated. A diff that does not fit the model's context
is refused with an actionable reason, because a walkthrough written against
half a diff reads as confident and is wrong.
- Nothing disappears. Lockfiles and other generated output are excluded from
the model's input by name — never by size — and everything no stop covers is
listed at the end, so "have I seen all of it" stays answerable.
- Cost is explicit. Results are content-addressed, so returning the working
tree to an earlier state costs nothing; generation outlives its request, so a
refresh detaches the client rather than discarding paid-for work, and only an
explicit cancel stops it.
Supporting changes to shared modules:
- git: expose the existing getRangeDiff as GET /api/git
listUntrackedPaths and getUntrackedDiffs. The latter resolve the repository
once for a batch instead of per file, taking a panel
~340ms on an 80-file working tree.
- small-model: structured output across four wire forma
and abort signal, and an onOverflow policy so an oversized prompt fails
loudly instead of being silently clipped. A provider
remembered so the prompt-side fallback goes first next time.
- models.dev metadata: surface structured_output as tri
false blocks a model, a missing field does not, because the catalog omits it
for roughly half of all models.
Desktop and tablet only: VS Code serves Git through its
these routes, and the mobile shell does not consume the surface registry.
Docs: packages/docs walkthrough page in English and all eight locales.
Skills listing ignored the active project when the OpenCode client
directory was unset, so project `.agents/skills` files were created but
never shown. Prefer the active project path (matching Commands/Agents)
and soft-fall back to it on skill API routes when directory is omitted.
Co-authored-by: Serhii Dziupin <makeittech@users.noreply.github.com>
Require an API key or {env:VAR} on client and server, add edit/prefill for
existing custom providers, save auth before config, and surface incomplete
auth plus disconnect after partial failures. Add VS Code parity tests and
drop the unused allProvidersConnected locale key.
Co-authored-by: Serhii Dziupin <makeittech@users.noreply.github.com>
Custom provider upsert uses PUT /api/provider, which was skipped by the
selective express.json allowlist and always saw an empty body.
Co-authored-by: Serhii Dziupin <makeittech@users.noreply.github.com>
Allow Settings → Providers to define custom providers (id, name, base URL,
API key, models, headers) without code changes. Persist config via OpenCode
layers, store keys through auth.set, and keep web/VS Code parity.
Co-authored-by: Serhii Dziupin <makeittech@users.noreply.github.com>
* perf: optimize session loading and startup
* fix(chat): stabilize history prepend virtualization
* perf: unblock first session open from startup network contention
Opening the first session after app start waited seconds for its message
fetch. Three independent contributors, each measured via CDP network
capture and Chromium net-log against the packaged desktop app:
- The active-session watchdog fired an uncapped per-directory status poll
and child-session discovery burst at startup, and other subsystems
(git checks, global session pages, command/skill discovery) fanned out
alongside it, saturating the browser's ~6 HTTP/1.1 sockets per origin.
Add a shared background-network gate (concurrency 3) and route the
watchdog, poll-shaped git reads (also priority: low), global session
pages, command/skill loads, and the background update check through it.
- The packaged renderer is cross-origin to the loopback backend, so every
API call needs a CORS preflight; a few slow OpenCode-proxied requests
held the whole pool while preflights and interactive traffic queued
behind them. Lift Chromium's per-host connection cap for loopback via
ignore-connections-limit in the Electron shell.
- OpenCode initializes each directory lazily on its first request, so the
first click paid that cost interactively. Warm the last-used directory
and the three most recently opened projects right after OpenCode
readiness, sequentially and best-effort, overlapping UI startup.
Validation: new background-network tests, lifecycle warmup test, focused
store/sync tests, UI type-check and lint, dead-code report, node --check
plus electron type-check/lint, and CDP first-open measurements on the
packaged app (message fetch socket queue 5.4s -> 0.03s).
* fix(ui): keep interactive git reads out of background queue
---------
Co-authored-by: Bohdan Triapitsyn <artmore@protonmail.com>
* feat: add Windows ARM64 support with x64-baseline CLI workaround
Windows ARM64 native opencode.exe fails with a Bun FFI/TinyCC dlopen
error (anomalyco/opencode#19130). As a temporary workaround:
- Bundle x64-baseline OpenCode CLI on ARM64 instead of native ARM64
(prepare-opencode-cli.mjs, env-runtime.js)
- Disable OpenCode self-upgrade on ARM64 in server, VS Code, and UI
(upgrade-capability.js, opencode-upgrade-runtime.ts, useUIStore.ts,
OpenCodeCliSettings.tsx, search.ts, SettingsView.tsx, platform.ts)
- Add ARM64 Windows cross-compile builds to release and smoke workflows
(release.yml, release-desktop-smoke.yml)
- Refactor Windows latest.yml to use combine-electron-manifests pattern
matching macOS, since two arches now produce per-arch manifests
The CI ARM64 build itself is permanent; only the x64-baseline CLI
bundling and upgrade disablement are temporary and should be reverted
when the upstream issue is resolved.
* fix(desktop): select Windows updater by architecture
---------
Co-authored-by: Bohdan Triapitsyn <artmore@protonmail.com>
Sérgio Pedro