feat(opencode): local LAN provider discovery + auto-discover models - #27554
Open
androidand wants to merge 350 commits into
Open
androidand wants to merge 350 commits into
androidand wants to merge 350 commits into
Conversation
github-actions
Bot
added
the
needs:compliance
Contributor
|
The following comment was made by an LLM, it may be inaccurate: Based on the search results, here are the potentially related PRs: Most Related:
Related by Pattern:
Note: PR #27554 (the current PR) appears as the top result in all searches, which is expected. The most directly related duplicate candidate is #26756, which already implements discovery from |
11 tasks
androidand
force-pushed
the
dev
branch
2 times, most recently
from
May 14, 2026 14:53
c368353 to
a56a8fe
Compare
github-actions
Bot
removed
the
needs:compliance
Contributor
|
Thanks for updating your PR! It now meets our contributing guidelines. 👍 |
androidand
changed the title
github-actions
Bot
added
needs:compliance
Contributor
|
Thanks for updating your PR! It now meets our contributing guidelines. 👍 |
Author
|
Please have a look, I think people will appreciate this, especially in combination with the updated llama-swap which adds feature parity (and more) with ollama. |
|
It would be very much appreciated. Thanks! |
|
This PR contains too much irrelevant data. You're adding local skills and theme content, along with debugging data from your local system. |
This was referenced Jun 12, 2026
6 tasks
llama-skein streams model-load "loading theme" flavor as reasoning_content SSE deltas tagged with a top-level `skein_loading: true`. opencode persisted them as reasoning, ballooning the session DB to ~16 GB and filling the disk (SqlError → TUI crash). The Vercel ai-sdk discards unknown top-level fields, so the marker is only visible on the RAW chunk. stripSkeinLoading() is a line-buffered SSE transform on the llama-skein local provider's stream that drops skein_loading events before the ai-sdk — they never enter the reasoning/message/persistence path, so they can never be stored, by construction. Their text is forwarded to the SkeinLoading transient channel for live display (a no-op until a display subscriber is wired — that's the remaining half of the task). Unit-tested incl. an event split across stream chunks. Fixes the persistence half of skein-duey (P1); live themed-display is the follow-up. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The turn-level loop-detection streak in prompt.ts resets on every run, so a session that breaks out of a stuck loop, auto-recovers via compaction, and breaks again can cycle forever (build -> compact -> build -> ...). Track consecutive breaks per session across runs in SessionCompaction; after 3 consecutive breaks the session is marked loop-dead and auto-compaction refuses to run, surfacing the terminal error instead of another silent recovery cycle. Manual /compact is unaffected, and any genuine turn clears the dead state. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…peated text The loop-detection block in runLoop only ever compared turns when neither had a tool call, so any turn with a tool call unconditionally reset the streak — a model stuck retrying the exact same tool call with the exact same arguments (e.g. re-sending an identical peer message after a failure) was invisible to it, and this is the more expensive failure mode since each repeat fires a real side effect. Extract the comparison into session/loop-detect.ts (pure, tested): a turn with tool calls is compared by exact call signature (tool + sorted-key arguments) against the previous turn's; a turn with no tool calls is compared by bigram text similarity as before. The two are never cross-compared, so switching between tool calls and plain text resets the streak rather than counting as a repeat. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…de Code SessionStatus and Permission (the only two signals statusFrom reads besides loop state) are both per-process in-memory state. A session another opencode process is actively driving was therefore indistinguishable from one this process knows to be genuinely idle: both are simply absent from `statuses`. resolvePeers (the `peers` tool's awareness roster) then dropped every real opencode sibling, leaving only the Claude Code peer — which has an actual cross-process registry (peer/claude/registry.ts) — visible, even though send_peer_message could already reach those siblings directly by session id. Fall back to treating a session with unresolved idle status as busy when it was updated within the last 45s (well above realistic turn/tool-call cadence). Scoped to resolvePeers only: resolveMessageTargets already treats idle as its intended target, and send_peer_message reads `status` back out to refuse delivery into a session it believes is genuinely mid-turn, so guessing "busy" there would make delivery to a real cross-process peer unreliable instead of fixing anything. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
task_id is meant to be an opaque id this same tool previously returned,
but a model — especially a weaker one — will sometimes invent its own
human-readable label instead (e.g. "review-changes-1"), believing it is
naming the task rather than resuming one. SessionID.make() throws a raw
schema validation error ("Expected a string starting with 'ses', got
...") for anything not shaped like a real session id, and that throw
happened while constructing sessions.get()'s argument — before the
existing catchCause around sessions.get could apply — so it escaped as a
confusing tool error instead of falling through to "create a fresh task"
like an unresolvable-but-well-formed task_id already does.
Guard with Schema.is(SessionID) first, and reword the parameter
description to make clear it must be reused verbatim from a prior
result, not invented.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…es clickable permission.mode (the auto-approve toggle, currently only reachable via <leader>p or the command palette) was missing the slashName that /auto was explicitly freed up for when the queue/backlog command renamed away from that name — "Auto mode" is upstream's name for permission auto-approve, not the queue driver, but nobody ever finished wiring it. Add slashName: "auto" so it behaves like every other slash-visible command. Also make the footer's agent-name label and "● auto" badge clickable (mouse-up, matching the existing loop/backlog badge's ClickText idiom) instead of plain, non-interactive text: clicking the agent name opens the same picker /agents does, and clicking the auto badge dispatches the same permission.mode command /auto and <leader>p already run. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
With only a couple of built-in agents, agent_cycle_reverse (shift+tab) was barely distinguishable from agent_cycle (tab) — cycling backward or forward through 2-3 items lands you in nearly the same place. Free shift+tab from agent cycling (still reachable via /agents or the command palette) and give it to permission_mode (auto-approve toggle) instead, alongside the existing <leader>p. Matches the muscle memory of similar tools' mode-cycling without overloading Tab's meaning with state-dependent behavior. Also add a rotating tip for /auto (permission.mode's new slashName), matching the existing tip for agent.cycle — the toggle was otherwise undocumented anywhere a new user would look. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
fork:verify flagged four files as unregistered divergence after today's commits: session/loop-detect.ts and its test are new fork-only files (owned); tips-view.tsx and test/tool/task.test.ts are the first fork edits to previously-untouched upstream files (patched, with a marker unique to each edit). Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
model-gallery-ui task 6.1: adds the section shell design.md decision 7
describes without changing existing behavior. SettingsModelsV2 becomes a
container with three sub-tabs:
- Installed: the pre-existing provider-grouped model list, moved verbatim
into models-installed.tsx (SettingsModelsInstalledV2) — zero behavior
change, same search/filter/visibility-toggle logic.
- Discover: real but deliberately narrow (models-discover.tsx) — lists
llama-skein hosts via the already-shipped GET /gallery/hosts endpoint
(model-gallery-ui task 5.7), online/offline status, installed-model count.
Not the full catalog search/filter/candidate-card experience (task 6.2) —
proof the gallery backend is reachable from this panel, nothing more.
- Operations: an honest empty state (models-operations.tsx). Real
operation-list data needs the app-side operations client task 7.2 builds;
no fabricated content in the meantime.
Added settings.models.{installed,discover,operations} tab labels and the
Discover panel's strings to en.ts, plus a new plural key
(settings.models.discover.hostInstalledCount) registered in
context/language.tsx's PluralKey union.
tsgo --noEmit -p packages/app: 0 errors (unchanged from baseline, this
package already typechecked clean). oxlint: 1 pre-existing-pattern warning
(narrowing assertion on the tab-value onChange, same shape as elsewhere in
this file tree), 0 errors.
Not yet verified in a running browser — no visual/E2E check performed this
pass; say so explicitly rather than claiming a look that didn't happen.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…er the gallery API
… owner's sidecar; owners publish busy/idle to the registry; sidecars run without Claude Code
…, confirmed install, /downloads shows llama-skein operations; SDK regenerated
… every hypothetical-fit call 404'd live
…irmed install) and Operations (progress, cancel, refresh on success)
…ed-store detection, load/unload, hide vs delete, copy/move between hosts; TUI /manage
…ment; register app gallery files
…lete with shared-store warning, copy/move between hosts
…decar registry — Session.list is project-scoped
…; regenerate llama-skein client
…rce_peer; provenance re-download is the fallback
- graphify-out/ (analysis-tool cache, ~32MB) was never gitignored and leaked into git; removed and ignored. - Three stray debug .log files (openapi-ts-error-*.log, debug-storybook.log) leaked local absolute paths; removed, *.log ignored. - artifacts/glm52-rise-video/ (personal AI-video experiment output, unrelated to opencode-skein) and HANDOFF*.md (personal home-lab GPU/VRAM tuning notes) removed — no product value to a public reader, out of scope for this repo. - .opencode/opencode.jsonc: skein MCP command now resolves via PATH instead of a hardcoded personal home directory. - family.ts: provenance comment now says ~/dev/skein instead of the full personal path. Follow-up not done here: several archived openspec/changes/** planning docs still reference real home-lab hostnames (rocky/z4/m3/m5) tied to real incident narratives. Left alone — needs careful per-doc editing to genericize without garbling the technical content, not a blind find-replace.
…scan scanLlamaSwap() opened a Bonjour browser, waited a fixed 1s window, tore it down, and returned whatever arrived — every /local request paid that latency, and any host that answered a beat late was silently missing from that response with no way to get it back short of another request. Adds startPeerRegistry()/getPeerRegistrySnapshot()/stopPeerRegistry() to local/mdns.ts: one Bonjour browser opened at server startup and kept running for the process lifetime, updated by "up"/"down" events as they happen, with LAN/localhost probing and per-peer model refresh moved to background intervals instead of per-request work. Reading a snapshot is synchronous and never waits on network I/O; a peer that misses one refresh cycle is marked `stale`, not dropped. Wired in via a new peer-registry-node.ts (mirrors LocalProviderSync's node shape) added to the app-runtime layer graph. The /local HTTP handler now reads the snapshot instead of calling scanLlamaSwap. scanLlamaSwap itself is untouched and still available for one-shot CLI use (sync.ts's local-provider-sync job). Verified against the real network (localhost + LAN probes found 5 real peers within one refresh cycle); tsgo --noEmit clean.
Replace real hostnames (rocky, z4, m3, m5, proxmox) with generic role-based labels (gpu-host-1..5) across ~14 archived planning docs, preserving the technical narrative (VRAM sizing, GPU model, incident details) — only the hostname changes.
The README was unmodified upstream content: install instructions pointed at opencode.ai/anomalyco packages, and nothing explained the fork's purpose or its pairing with llama-skein. Add a "Why this fork exists" / "Getting started" section up front with the fork's actual positioning and a working curl install command, and flag that the untouched sections below (brew/npm/nix/etc.) install upstream opencode, not this fork.
A peer review found the earlier cleanup was incomplete: /Users/andreas still appeared in 7 files, m5-llamaswap in 1. Fixes: - ECOSYSTEM.md: ~/dev/... instead of the full personal path. - Register packages/opencode/src/local/peer-registry-node.ts in fork/manifest.json (bun run script/fork-verify.ts was failing on it as an unregistered divergence). - 4 archived openspec docs: /Users/andreas/ -> ~/. - Test fixtures (presence-claude, peers, sync): personal paths and a real hostname (m5-llamaswap) swapped for generic equivalents. Not done here, flagged separately: several 192.168.1.x IPs remain in tracked test fixtures, at least one of which (192.168.1.126) matches a real, currently-live host on the owner's network, not a coincidental example address. Left alone pending a decision on scope, same as the larger history-rewrite question this repo now also has (see report).
…nsistent The earlier scrub renamed one m5 occurrence to gpu-host-1 but not the rest of this file's rocky/m3/m5 slugs, and canonicalName strips the -llamaswap suffix, so the input/assertion pair no longer matched (caught by a peer review before this shipped further). Reverted the premature partial rename here and kept the file using the real hostnames consistently — the upcoming history rewrite's replace-text pass handles renaming all occurrences everywhere in one atomic step, which is the right place to do it, not piecemeal per-file edits. Anonymized this file's IPs to the 192.0.2.0/24 documentation range (RFC 5737) while here.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Issue for this PR
Closes #6231
Closes #27553
Type of change
What does this PR do?
Adds
Local (LAN)discovery in/connectfor local OpenAI-compatible servers.Auto-discovers models for OpenAI-compatible providers from
/models, including limits like context window and max output tokens when available.Compatible with and preserves manually configured/edited providers and models and respects
discoverModels: true|false.Adds context
current / context_lengthwhen available (see llama-swap PR below) to Context in sidebar and footer.Fixes unrelated bugs introduced by rebase, --agent parameter would crash opencode.
Related
Server-side companion: mostlygeek/llama-swap#755 adds
context_lengthandmax_output_tokensto the/v1/modelsresponse. When those fields are present, this PR reads them to configure each model's context window and output token limits automatically — no manual config needed.How did you verify your code works?
bun run typecheckbun test test/provider/provider.test.ts --timeout 30000bun turbo typecheck/connect -> Local (LAN)against multiple local OpenAI-compatible servers on localhost and LAN.Screenshots / recordings
Provider config example
Checklist