CLI: Rename unsloth connect to unsloth start#6613
Conversation
|
Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits. |
for more information, see https://pre-commit.ci
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request renames the CLI command unsloth connect to unsloth start across the entire codebase, updating CLI entry points, helper scripts, tests, and frontend components. It also introduces a new UI section in the frontend settings to display and copy the coding agent command. Feedback on the changes suggests fixing a comma splice in the newly added English translation string to improve grammatical correctness and readability.
Important
The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.
|
@codex review |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: dacc19461e
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| const KEY_PLACEHOLDER = "sk-unsloth-YOUR_KEY"; | ||
| const MODEL_FALLBACK = "unsloth/gemma-4-E4B-it-GGUF:UD-Q5_K_XL"; | ||
|
|
||
| const AGENT_COMMAND = "unsloth start claude"; |
There was a problem hiding this comment.
When this panel is showing a non-default local port or a tunneled/remote base URL, the copied agent command still contains neither UNSLOTH_STUDIO_URL nor --api-key. find_studio_server() defaults to http://127.0.0.1:8888 unless UNSLOTH_STUDIO_URL is set, and _agent_api_key() refuses automatic minting for non-loopback servers without an explicit key, so users following this snippet can connect to the wrong local server or fail instead of launching against the server shown in the UI. Build the snippet from the selected base/available key or show the required env vars for these cases.
Useful? React with 👍 / 👎.
The API keys panel showed a hardcoded `unsloth start claude`. `unsloth start` defaults to 127.0.0.1:8888 and only mints a key for a loopback server, so a non-default port or a tunnel/remote base would target the wrong server or fail to mint. Build the command from the panel base/key (and emit a key for non-loopback), matching the other snippets in the panel.
Avoids breaking existing scripts and docs that still call `unsloth connect`.
The test builds _SameTaskStreamingResponse via __new__, so set the attribute that __call__ now reads.
|
Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits. |
|
Pushed three commits to wrap up the feedback here. 1. The coding-agent command now targets the server the panel is showing. 2. 3. Fixed an unrelated pre-existing failure in Repo tests (CPU). The grammar note on Verified locally: |
|
Pushed three commits to wrap up the review feedback. 1. Frontend loopback check now matches the CLI's 2. 3. Made the touched frontend files biome-clean by wrapping the long Verified locally: One item left before merge: this branch is about 20 commits behind |
# Conflicts: # tests/studio/test_stream_cancel_registration_timing.py
for more information, see https://pre-commit.ci
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 87a9f64bb9
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
|
|
||
| // The bare `unsloth start` only auto-discovers 127.0.0.1:8888 (localhost/::1 alias it). | ||
| function isDefaultLocalHost(host: string): boolean { | ||
| return host === "127.0.0.1" || host === "localhost" || host === "::1"; |
There was a problem hiding this comment.
URL.hostname is bracketed for IPv6 literals (for http://[::1]:9000 it is [::1]), so this check never treats Studio's IPv6 loopback URL as loopback. In that local IPv6/non-default-port case the command builder falls through and appends --api-key sk-unsloth-YOUR_KEY once the reveal card is dismissed, even though the CLI can auto-mint for ::1; copying the shown command then fails instead of connecting. Normalize [::1] to ::1 before calling these predicates.
Useful? React with 👍 / 👎.
|
Decided to drop Removed in 6e543d7:
After this, |
|
Reverting my last call: keeping In 89b2cdc: re-added the hidden Verified: |
What
|
| Command | Agent | Notes |
|---|---|---|
unsloth start claude |
Claude Code | sets ANTHROPIC_BASE_URL / ANTHROPIC_AUTH_TOKEN / ANTHROPIC_MODEL |
unsloth start codex |
OpenAI Codex | writes a ~/.codex profile; needs a GGUF model on llama-server |
unsloth start openclaw |
OpenClaw | writes ~/.openclaw/openclaw.json |
unsloth start opencode |
OpenCode | writes the opencode config |
unsloth start hermes |
Hermes (Nous Research) | writes ~/.hermes/config.yaml |
What it does under the hood
- Discovers the Studio server from
UNSLOTH_STUDIO_URL(defaulthttp://127.0.0.1:8888). If none is running, it tells you to start one. - Gets an API key: for a local loopback server it mints one automatically and remembers it per server; for a remote server you pass
--api-key(orUNSLOTH_API_KEY). - Resolves the model:
--modelloads that model on the server, otherwise it uses whatever Studio currently has loaded. - Writes that agent's config/env and launches the agent's binary, forwarding any extra args straight through.
Shared flags
--model,-m: model to use; defaults to the one loaded in Studio.--api-key: Studio API key (auto-minted locally, required for remote). Also readsUNSLOTH_API_KEY.--launch/--no-launch:--no-launchprints the env and command instead of running it (handy for remote shells or WSL).- Trailing args pass straight to the agent, e.g.
unsloth start claude -p "...".
Workflow and examples
# 1) Start a Studio server serving a model (one terminal)
unsloth run -hf unsloth/gemma-3-270m-it-GGUF
# or open the Studio app and load a model from the dropdown
# 2) Launch a coding agent against it (another terminal)
unsloth start claude # Claude Code on the loaded model
unsloth start codex # OpenAI Codex (GGUF model required)
unsloth start openclaw
unsloth start opencode
unsloth start hermes
# Use a specific model (loads it on the server if needed)
unsloth start claude --model unsloth/Qwen3-4B-GGUF
# Print env + command instead of launching (remote shells / WSL)
unsloth start claude --no-launch
# Point at a remote or tunneled Studio
UNSLOTH_STUDIO_URL=https://abc.trycloudflare.com unsloth start claude --api-key sk-unsloth-...
# Pass extra flags straight to the agent
unsloth start claude -p "summarize this repo"
# Backwards compatible: the old name still routes to the same thing
unsloth connect claudeThe Studio API-keys panel (Settings -> API) also shows a ready-to-copy unsloth start command targeting whatever server it is displaying.
…context-length, load-in-4bit, tensor-parallel) (unslothai#6613)
|
Made
They are forwarded to unsloth start claude --model unsloth/Qwen3-4B-GGUF --gguf-variant UD-Q4_K_XL --context-length 8192Intentionally not included: the server-startup flags ( All five agents and the |
for more information, see https://pre-commit.ci
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: cded5f9aed
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
|
|
||
| // Non-default server: set the URL; non-loopback also needs an explicit key. | ||
| let cmd = bare; | ||
| if (!loopback && key) cmd += ` --api-key ${key}`; |
There was a problem hiding this comment.
When the panel is using a non-loopback base (Cloudflare tunnel or LAN URL) after the reveal card is dismissed, UsageExamples passes sk-unsloth-YOUR_KEY as key, and this line turns that placeholder into an explicit --api-key. Running the copied command then saves and sends the placeholder, so unsloth start fails against /v1/models instead of either auto-minting locally or prompting for a real key; only append --api-key when a real revealed key is available, or show a template that clearly must be edited.
Useful? React with 👍 / 👎.
| ) -> dict: | ||
| models = _loaded_models(base, key) | ||
| match = next((m for m in models if m["id"] == requested), None) | ||
| if requested and match is None: |
There was a problem hiding this comment.
When the requested model id is already present in /v1/models, this condition skips /api/inference/load entirely, so the newly added load knobs are silently ignored. For example, if unsloth/foo-GGUF is loaded with one GGUF variant or context length, unsloth start claude --model unsloth/foo-GGUF --gguf-variant Q8_0 --context-length 8192 still connects to the old loaded instance; the backend's load path already compares these settings and would reload when they differ, but it is never reached here.
Useful? React with 👍 / 👎.
…daction Three review findings: 1. CI could leak a live key. agent-guides-drive.sh printed the raw 'unsloth start --no-launch' transcript (which carries export UNSLOTH_API_KEY / ANTHROPIC_AUTH_TOKEN lines) to the Actions log on both the failure path and the success path before redact() ran. Add cat_redacted() and use it for those two prints, so the key is scrubbed on the way to the log while the on-disk file stays intact for the env parsing that follows. 2. Outages masqueraded as bad keys. _key_accepted caught a broad Exception and returned False, so a 5xx or timeout while checking a cached key looked like a rejection: it discarded a good key and minted extra ones (local) or reported 'no saved key' (remote). Only treat HTTP 401/403 as a rejection; let other errors propagate so a real outage surfaces. 3. Codex preflight could leave the auto-started server up. _require_gguf_for_codex runs after _connect may have auto-started Studio but before _run installs its teardown finally, so a preflight rejection (e.g. a transformers-backend model) left the server holding the port/GPU until the atexit backstop. Tear it down explicitly at the point of failure. Tests: a 5xx on a saved key surfaces without minting; a non-GGUF codex preflight tears down the auto-served server.
|
Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits. |
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: fc2734e371
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| const host = normalizeHost(url.hostname); | ||
| const loopback = isLoopbackHost(host); | ||
| // Default local server (127.0.0.1/localhost:8888): bare command auto-discovers it. | ||
| if (isDefaultLocalHost(host) && url.port === DEFAULT_STUDIO_PORT) return bare; |
There was a problem hiding this comment.
When the API panel's base is http://[::1]:8888 (for example, Studio bound to IPv6 loopback), this branch returns the bare unsloth start claude command even though the CLI default probes http://127.0.0.1:8888. Users copying the command from that local IPv6 server will fail to find Studio, or could attach to a different server on IPv4; keep emitting UNSLOTH_STUDIO_URL for ::1 instead of treating it as the bare default.
Useful? React with 👍 / 👎.
| # home-scope the session to leave the user's real ~/.pi untouched. The key | ||
| # rides in the config rather than the env. | ||
| write_pi_config(base, key, entry, home / ".pi" / "agent" / "models.json") | ||
| env = {"HOME": str(home)} |
There was a problem hiding this comment.
When the user's shell already has PI_CODING_AGENT_DIR set, Pi ignores the HOME relocation here: upstream Pi's getAgentDir() reads process.env[ENV_AGENT_DIR] (for Pi, PI_CODING_AGENT_DIR) before falling back to homedir(). Because _launch preserves that inherited variable, the child can skip the Unsloth-written models.json and start with the user's config instead, so the unsloth provider/key may be missing; set or clear PI_CODING_AGENT_DIR for this session as well.
Useful? React with 👍 / 👎.
| # Auto-start a local server only for an interactive launch with a model to serve, and | ||
| # only for a loopback target (never stand in for an explicit remote UNSLOTH_STUDIO_URL). | ||
| if serve and launch and model and is_loopback_url(expected): | ||
| return expected, _start_studio_server(expected, model, load or LoadOptions()) |
There was a problem hiding this comment.
If UNSLOTH_STUDIO_URL is a loopback URL with no explicit port, e.g. http://127.0.0.1, this passes that unnormalized base into _start_studio_server; the child is launched on parsed.port or 8888, but readiness still polls the original port-80 URL via _studio_healthy(base), so unsloth start ... --model ... waits until the 15-minute timeout even though the server came up on 8888. Normalize expected to the actual host:port before starting/polling.
Useful? React with 👍 / 👎.
| *_yolo_command_flags("pi", yolo), | ||
| *ctx.args, | ||
| ] | ||
| install_hint = "npm install -g @earendil-works/pi-coding-agent" |
There was a problem hiding this comment.
When Pi is missing and the user accepts the auto-install prompt, _install_agent executes this hint as-is. Pi's current install docs and README document the npm path with --ignore-scripts as part of their supply-chain hardening, but this command omits it, so accepting the prompt runs package lifecycle scripts during a global install instead of following the upstream safe install recipe.
Useful? React with 👍 / 👎.
… install recipe Four review findings: 1. Pi ignored the session config when PI_CODING_AGENT_DIR was already set. Pi's getAgentDir() reads process.env.PI_CODING_AGENT_DIR before falling back to $HOME/.pi/agent, so a value inherited from the user's shell sent Pi to their real config and skipped our provider/key (the HOME relocation alone was not enough). Pin PI_CODING_AGENT_DIR at the session's .pi/agent dir; it is path-valued so the WSL bridge translates it automatically. 2. Pi install hint dropped Pi's documented --ignore-scripts. Pi's README installs with 'npm install -g --ignore-scripts @earendil-works/pi-coding-agent' and notes it needs no install scripts, so accepting the prompt now follows that safe recipe. 3. Auto-start ignored a portless UNSLOTH_STUDIO_URL. unsloth run binds to 'parsed.port or 8888', so http://127.0.0.1 launched the child on 8888 but the health poll (and the returned base) still used port 80, stalling until the startup timeout. Normalize the base to host:8888 (IPv6-safe) before starting and polling. 4. API-panel command mistook IPv6 loopback for the bare default. The bare 'unsloth start' only probes 127.0.0.1:8888 on the IPv4 stack, so http://[::1]:8888 must carry an explicit UNSLOTH_STUDIO_URL. Drop ::1 from the bare-default host set while keeping it a loopback host (URL emitted, no key needed). Tests: PI_CODING_AGENT_DIR is set to the session dir; _effective_base normalizes portless/IPv6 bases; a portless UNSLOTH_STUDIO_URL auto-serves on :8888.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: d165401d92
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| const host = normalizeHost(url.hostname); | ||
| const loopback = isLoopbackHost(host); | ||
| // Default local server (127.0.0.1/localhost:8888): bare command auto-discovers it. | ||
| if (isDefaultLocalHost(host) && url.port === DEFAULT_STUDIO_PORT) return bare; |
There was a problem hiding this comment.
When Studio reports http://localhost:8888 (for example from --host localhost, especially on systems where localhost resolves to IPv6), this branch returns the bare unsloth start claude. The CLI's bare default only probes http://127.0.0.1:8888 unless UNSLOTH_STUDIO_URL is set, so copying the shown command can miss the server displayed in the UI or attach to a different IPv4 process. Only elide the URL for the literal 127.0.0.1:8888 default, or keep UNSLOTH_STUDIO_URL for localhost too.
Useful? React with 👍 / 👎.
| expected = _effective_base(expected) | ||
| return expected, _start_studio_server(expected, model, load or LoadOptions()) |
There was a problem hiding this comment.
When UNSLOTH_STUDIO_URL is an HTTPS loopback URL and no Studio is running, this branch still auto-starts unsloth run without any TLS mode but keeps expected as https://.... The readiness loop then probes the plain HTTP child over HTTPS, so unsloth start ... --model ... waits the full startup timeout before tearing it down. Restrict auto-serve to HTTP loopback URLs or normalize the returned base to the actual HTTP endpoint.
Useful? React with 👍 / 👎.
…command From a fresh multi-reviewer pass over the merged head plus the latest Codex bot review: 1. Load knobs now always consult the server. _resolve_model matched on model id alone, so --gguf-variant / --context-length / --no-load-in-4bit / --tensor-parallel were silently ignored whenever the id was already loaded (asking for UD-Q4_K_XL kept a Q8_0 serving). With any explicit knob the CLI defers to /api/inference/load, whose already-loaded dedup answers without reloading when variant and settings match, so a second session running the same command still attaches without evicting the first. 2. OpenCode --yolo and the session model pin now ride in OPENCODE_CONFIG_CONTENT. A project's own opencode.json outranks OPENCODE_CONFIG, so a repo config could silently override the session model and the --yolo permission block; OPENCODE_CONFIG_CONTENT outranks project config. The API key stays in the private file, never in printed env. 3. The --no-launch recipe's last line is a self-contained one-liner (inline VAR=value assignments before the command, conflicting vars blanked). People copy just the last line, and a bare codex/claude there ran against the user's real ~/.codex or Anthropic credentials with zero isolation, e.g. inheriting a pre-existing damaged ~/.codex state DB and blaming the recipe. The CI drive script scrubs the key from the one 'invoking:' echo this adds. 4. The auto-serve log is 0600 and the parent handle is closed. It sat world-readable in the shared tempdir under a predictable name while carrying the minted sk-unsloth- key from the unsloth run banner. 5. _key_accepted fails with a clean message on outages. Non-auth errors (5xx, network, timeout) surfaced as a raw traceback; 401/403 still mean a rejected key. 6. _effective_base strips URL paths, and https loopback targets never auto-serve. http://127.0.0.1:8888/studio polled /studio/api/health (404) and https://127.0.0.1 polled the wrong scheme, both spinning until the 15-minute startup timeout. 7. API-panel command: only literal 127.0.0.1:8888 earns the bare command. localhost can resolve to ::1, which the bare CLI never probes, so it keeps UNSLOTH_STUDIO_URL. 8. CI artifact sweep covers redacted-configs/ and agent-workdir/, not just logs/. Tests: 125 CLI tests pass (new coverage for each fix), 156 backend tests pass, ruff clean. Adds an unsloth connect alias regression test.
for more information, see https://pre-commit.ci
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 8629227120
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| # pin (and --yolo permissions) would silently lose to a repo config. Carry the | ||
| # settings that must win in OPENCODE_CONFIG_CONTENT, which outranks project | ||
| # config; the API key stays in the private file, never in the printed env. | ||
| inline_config: dict = {"model": f"unsloth/{entry['id']}"} |
There was a problem hiding this comment.
Fresh evidence: this fix moves only model (and yolo permissions) into OPENCODE_CONFIG_CONTENT; the provider.unsloth endpoint/key still lives only in the lower-precedence OPENCODE_CONFIG file written above. OpenCode's config docs say OPENCODE_CONFIG is loaded before project opencode.json, while OPENCODE_CONFIG_CONTENT is the later runtime override, so in any repo that already defines provider.unsloth the project provider can replace this Studio base/key and model: unsloth/... will still target the project's provider instead of the current Studio session. Use a collision-proof provider id or put the provider override in the same higher-precedence layer.
Useful? React with 👍 / 👎.
| typer.echo(" ".join(_powershell_quote(arg) for arg in command)) | ||
| return |
There was a problem hiding this comment.
On native Windows --no-launch, this branch prints the $env: setup lines and then echoes only the bare agent command before returning, so copying just the last line skips the relocated config/key env (CODEX_HOME, OPENCODE_CONFIG, ANTHROPIC_AUTH_TOKEN, etc.). With the new session-scoped config dirs, that partial copy runs against the user's real agent config or credentials instead of the Studio session; emit a PowerShell-scoped one-liner like the POSIX path or otherwise make the final command include the env.
Useful? React with 👍 / 👎.
Pi paints inline from wherever the cursor sits: its first render assumes a clean screen instead of clearing or entering the alternate screen itself (current Pi never emits a clear at startup). Launched under unsloth start, that left the session starting mid-scroll beneath the connection output. Clear the screen (click.clear, cross-platform, no-op without a TTY) right before the Studio banner so Pi opens exactly one line down on a clean viewport. Launch path only: --no-launch recipes and piped output are never wiped, and alternate-screen agents are left alone.
for more information, see https://pre-commit.ci
Pi now opens on a clean screen (260da17)Feedback from testing: Root cause: current Pi never clears at startup. Its first TUI paint explicitly assumes a clean screen (no clear sequence, no alternate screen; verified against Pi 0.79.10 and 0.80.3 by capturing raw pty output and Pi's own redraw debug log, which shows only "first render" with no clear). Bare Fix: the launcher now hands Pi the clean screen its first render assumes. Right before the Verified end to end in tmux against a live Studio: scrollback markers gone from the viewport, banner on line 1, Pi's UI directly below, and a real tool-call turn completed in the session. 128 CLI tests pass including three new ones (Pi launch clears exactly once before exec, --no-launch never clears, claude launch never clears). |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 987cf40922
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| # No start.py recipe; the agent's documented package name. The CLI moved | ||
| # from the now-deprecated @mariozechner scope to @earendil-works (the old | ||
| # scope is frozen, so installing it would test a stale Pi against the API). | ||
| npm_retry "@earendil-works/pi-coding-agent" \ |
There was a problem hiding this comment.
Now that unsloth start pi has a real install hint (npm install -g --ignore-scripts @earendil-works/pi-coding-agent in start.py), this branch is no longer a hand-written exception. The CI installer still calls npm_retry without that flag; I checked npm install --help, which lists [--ignore-scripts] for install, so this cell both runs package lifecycle scripts and stops exercising the exact recipe users are prompted to run. Pass the flag through for Pi so guide drift is caught.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Fixed in f7f4cff: npm_retry passes flags through and the pi cell installs with --ignore-scripts, exactly matching the start.py hint; the stale no-recipe comment is refreshed too. Validated on a staging clone where the connection (pi) cell passed with the flagged install.
Hermes refuses to initialize when the served model's context window is under 64,000 tokens, and a second copy of the same check rejects the compression model mid-session. write_hermes_config previously pinned the real window, so any small local model (e.g. 40,960) failed at startup with manual config.yaml instructions. For windows below the floor the recipe now claims 65,536 in model.context_length, scales compression.threshold so compaction still fires at 90% of the real window, and sets auxiliary.compression.context_length to cover the mid-session check. Windows at or above the floor keep the exact previous behavior.
|
Pushed 5f158fc: Hermes refuses to initialize when the served model reports a context window under 64,000 tokens, and the same check runs a second time against the compression model mid-session. The recipe previously pinned the real window, so any smaller local model (for example Qwen3-1.7B at 40,960) failed at startup and pointed the user at manual config.yaml edits. For windows below the floor,
Windows at or above the floor keep the exact previous behavior, and the config stays idempotent. Verified end to end against a llama-server serving Qwen3-1.7B at 40,960 context: a fresh session now initializes with no floor error and completes interactive file tool turns; before the change the same setup reproduced the hard init refusal. CLI test suite passes (129), and the Local Agent Guides connection matrix stays green, so its model-choice comment no longer needs to cite the Hermes floor. |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 5f158fccf5
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| loaded = _http_json( | ||
| "POST", | ||
| f"{base}/api/inference/load", | ||
| key, | ||
| payload, |
There was a problem hiding this comment.
Fresh evidence: this POST is still not enough for an already-active transformers/non-GGUF model, because _load_model_impl returns already_loaded for that path based only on backend.active_model_name (studio/backend/routes/inference.py:3051-3054) before comparing max_seq_length or load_in_4bit. In that context, unsloth start ... --model same --context-length 8192 can print that it is ensuring the requested settings but still attach the agent to the old context/quantization; force a reload/unload for non-GGUF load overrides or make the backend guard compare these forwarded options.
Useful? React with 👍 / 👎.
The pi cell predates the pi recipe in start.py and still installed the package with lifecycle scripts enabled, so CI stopped exercising the exact command users are prompted to run. npm_retry now passes extra flags through, the pi branch mirrors the install hint verbatim, and the stale no-recipe comment is refreshed.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: f7f4cff30a
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| const host = normalizeHost(url.hostname); | ||
| const loopback = isLoopbackHost(host); | ||
| // Default local server (127.0.0.1/localhost:8888): bare command auto-discovers it. | ||
| if (isDefaultLocalHost(host) && url.port === DEFAULT_STUDIO_PORT) return bare; |
There was a problem hiding this comment.
When the panel base is https://127.0.0.1:8888, this condition still returns the bare unsloth start ... command because it checks only host and port. The CLI's bare default probes http://127.0.0.1:8888, so copying the recipe from a local HTTPS endpoint can either miss the displayed server or attach to a different plain-HTTP process; only elide UNSLOTH_STUDIO_URL for the literal http://127.0.0.1:8888 default.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Fixed in 6e653f8: the bare-command elision now also requires the http: scheme, so an HTTPS loopback on the default port keeps its explicit UNSLOTH_STUDIO_URL.
| shutil.rmtree(path, ignore_errors = True) | ||
| else: | ||
| path = _agents_config_root() / agent | ||
| shutil.rmtree(path, ignore_errors = True) |
There was a problem hiding this comment.
When --no-launch is used, the printed recipe points the agent at this stable per-agent directory (CODEX_HOME, HERMES_HOME, OPENCLAW_STATE_DIR, etc.). If the user runs unsloth start <same-agent> --no-launch again while a previously printed command is still running, this unconditional removal deletes that live session's config/state home, which can make agents that keep sessions or SQLite/state under that directory fail or lose state mid-run; use a per-recipe directory or avoid deleting a home that may be active.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Fixed in 6e653f8: the stable no-launch home is no longer wiped on re-run; it is created if missing and reused, since every config writer already merges idempotently into an existing home and a previously printed recipe may still be running an agent whose sessions or sqlite state live there. Test pins that a marker file survives a second run.
| load = load._replace(gguf_variant = variant) | ||
| base, server = _require_studio(model, load, serve = serve, launch = launch) | ||
| try: | ||
| key = _agent_api_key(base, api_key) |
There was a problem hiding this comment.
When no Studio is running and the user has UNSLOTH_API_KEY exported for some other server, _require_studio can auto-start a fresh local server for --model, but this line still forces that stale key into the first /v1/models request. The new local server can mint/use its own local key, yet the agent launch fails with auth errors solely because an unrelated key was present in the environment; for server is not None, validate the explicit key before using it or fall back to the local mint/cache path.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Fixed in 6e653f8: for a server this run auto-started, an explicit or exported key is validated first and, if the fresh server rejects it, the launch falls back to the local mint/cache path instead of failing (and the foreign key is never remembered for that base). A key the fresh server accepts, e.g. one persisted in the Studio home's auth db, is honored exactly as before. Both cases are covered by new tests.
The empty-string guards ran after appending /config.toml or /config.yaml, so they could never fire: crosscheck_contract silently skipped its contract checks and patch_hermes_tools died on the root path with a bare traceback. Check the raw variable first and guide_fail with the real cause.
|
Pushed bc7498b: hardened the guide-drive script's relocation-var handling. The empty-string guards ran after appending /config.toml or /config.yaml, so a missing CODEX_HOME or HERMES_HOME in the connect output could never trip them; crosscheck_contract silently skipped its contract checks and patch_hermes_tools died on the root path with a bare traceback. The raw variable is now checked first and fails with the real cause. Validated on a staging clone: all six connection cells green, Windows and macOS parity green. |
|
You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard. |
for more information, see https://pre-commit.ci
|
Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits. |
|
Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits. |
5 participants
The unsloth connect command has been renamed to unsloth start and expanded to launch multiple coding agents (Claude Code, OpenAI Codex, OpenClaw, OpenCode, Hermes) against a running Unsloth Studio server.