Skip to content

CHANGELOG.md

Latest commit

 

History

History
846 lines (704 loc) · 61.3 KB

CHANGELOG.md

File metadata and controls

846 lines (704 loc) · 61.3 KB
Raw

Changelog

Past-tense record of shipped changes.

Unreleased

v2.7.0 - 2026-08-02

Coordinated LocalX release.

v2.6.0 - 2026-07-27

Coordinated LocalX release.

  • One-line install for the whole stack. localbox is now installed alongside the rest of LocalX by a single command, with each archive checked against its published SHA-256 before it is unpacked:

    curl -fsSL https://raw.githubusercontent.com/C0deGeek-dev/LocalPilot/main/install/install.sh | sh

    The train cuts every tool to one version and they are only tested together, so they are installed together; localpilot update --all re-runs it. Nothing in this repository changed to support it — the release archives and manifest.json already published were the whole contract.

v2.5.0 - 2026-07-27

Coordinated LocalX release.

  • Releases now ship prebuilt binaries. A tagged release previously carried release notes and nothing else, so the only way to get this tool was to build it from source with a Rust toolchain. Each release now attaches an archive per platform — Linux x86-64 (glibc and a static musl build that runs anywhere), Linux arm64, macOS Apple Silicon, and Windows x86-64 — with a SHA-256 beside each archive and a manifest.json indexing the release.

    Publishing happens once, only when every platform built. A partial release is worse than a failed one: a download cannot tell the difference. The checksums prove an archive was not corrupted in transit; they do not prove who produced it, which needs signing.

v2.4.0 - 2026-07-26

Coordinated LocalX release.

  • Classic draft-model speculative decoding is a first-class catalog surface. A model entry can name a DraftModule (a drafter GGUF in its repo); localbox launch <model> --draft resolves it like the vision projector — shown in dry-run as present or will-download, fetched on demand, a failed download stops the launch — and the server runs it as draft-simple speculation. A drafter cannot combine with an MTP spec-type (one speculation engine per launch; refused up front). A drafter the server rejects for speculation (e.g. a draft/target vocab-type mismatch) does not fail the server — it runs without speculation — so the launcher detects that from the server log and warns that the drafter is loaded but idle. Both Bonsai entries wire their DSpark Q4_1 drafters; as of the current Prism release the server rejects those for exactly that mismatch, which the catalog descriptions now state.

  • Added Bonsai 27B 1-bit (bonsai27b, GGUF Q1_0, ~3.8 GB) to the shipped catalog. Q1_0 is upstream llama.cpp, so it runs in plain native mode on the pinned mainline build — no Prism fork required — making it the working fallback when the Prism engine is unavailable; vision via its own mmproj.

  • Shipped configuration no longer goes stale on existing installs. The two shipped layers (defaults.json, llm-models.example.json) now refresh to match the installed binary — previously they were seeded once and never updated, so new release pins and engine settings silently never reached an existing ~/.local-llm (user overrides belong in settings.json, which always wins). Your editable llm-models.json is still never overwritten: localbox update now reports shipped models your catalog predates, and a new localbox update --merge-models adds exactly those missing entries — additive only, existing entries and aliases untouched, --check previews.

  • The Prism engine now installs on Linux (x64/arm64): CUDA archives matched to the driver's CUDA major (newest toolkit build within it), Vulkan on an AMD GPU, the plain CPU archive otherwise; rocm and KleidiAI archives are deliberately not selected. Windows/macOS selections are unchanged. The installer now also warns, on every platform, when the only available Prism CUDA build's major does not match the driver's — that pairing can emit garbage output, which the launch smoke test catches before an agent sees it. The .build-stamp variant is derived from the actually selected asset (cuda-12.4 / cuda-12.8 / vulkan / cpu / metal) instead of a per-OS hardcode, and the shipped pin table covers the Linux archives.

  • Engine pins now age loudly and advance deliberately. localbox update --check reports per pinned mode whether the pinned release tag is behind the latest upstream release (informational — nothing auto-installs), and a new localbox update --mode <m> --refresh-pins advances one mode's pin to the latest release: it installs this host's assets, verifies each download against the GitHub release's published sha256 digest (a mismatch refuses to install or record), and records the new tag plus asset hashes in settings.json, which wins layer precedence and survives upgrades. The PrismML engine pin moved to the current fork release, whose Windows binary and macOS Metal archive hashes are re-pinned in defaults.json.

  • Fixed Prism responses being misreported by LocalPilot as truncated streams. The vendored shared proxy now synthesizes a missing Anthropic content_block_start and flushes held-back visible text before content_block_stop, while genuine EOFs still remain truncations.

  • Added the Ternary Bonsai 27B runtime path end-to-end. The catalog now marks tbonsai27b as requiring the new prism engine, so CLI, guided launches, saved defaults, and AutoBest cannot route its low-bit weights through an incompatible llama.cpp build. The pinned PrismML release installs the Windows x64 CUDA 12.4 binary plus runtime DLLs, or the standard Apple Silicon Metal archive; KleidiAI and DSpark are intentionally out of scope.

  • Configured vision projectors now download on demand when --vision is used. Dry-run shows whether the mmproj is present or will download, and a failed projector download stops the launch instead of silently claiming vision or degrading to text-only.

  • Fixed runaway GPU memory on localbox serve (and guided "for other tools" launches): those paths omitted --parallel, and llama-server's own default is now multi-slot auto (-np -1 → 4 slots), which allocates the full configured context per slot — roughly 4× the KV-cache memory the launch was sized for, OOMing models that fit fine under launch. Every launch path (CLI launch/serve, the guided launcher, and the native retry) now applies the same single-session defaults --parallel 1 --cache-reuse 256 through one shared finalizer, restoring v1.x behaviour. The guided launcher also picks up the settings.json launch tunables (LlamaCppAgentParallel, LlamaCppAgentCacheReuse, LlamaCppNCpuMoe, LlamaCppMlock, LlamaCppNoMmap) it previously ignored. To deliberately use llama-server's auto slot count, set LlamaCppAgentParallel to -1 (a non-positive value is not emitted, leaving the server default in charge).

v2.3.0 - 2026-07-07

Coordinated LocalX release.

  • The shared crate tier advanced: the no-think proxy's gateway key check is now constant-time (SHA-256 digest comparison instead of a short-circuiting string equality), and its docs state the LAN posture honestly — the key gates access over plain HTTP, and /health deliberately stays open.
  • Fixed the keyed LAN launch end-to-end: the reply-path smoke test now authenticates with the gateway key, and the Claude/LocalPilot agent environment carries that key instead of the loopback placeholder — a --lan --password launch previously aborted at the smoke stage against a fresh key-enforcing proxy (and the agent would have been refused anyway).
  • Proxy reuse is now posture-aware: an existing proxy is reused only when its listen host and key match the plan, read from the live process itself. A keyless proxy can no longer be silently reused behind a "key required" banner, and an earlier open-LAN gateway no longer keeps serving after a loopback-only launch — mismatches restart the proxy with the wanted posture. Existing proxies started before this change restart once on the next launch.
  • --lan --keep-thinking is refused up front: --keep-thinking routes the agent straight at the loopback-only server, so the LAN gateway it announced never actually started.
  • A failed reply check now stops the server (and any proxy) that launch had just started instead of stranding them; a reused proxy from an earlier launch keeps serving.
  • The guided launcher resolves VRAM through the same shared ladder as the CLI path: the VRAMGB setting is honored and a host with no detectable GPU falls back to the shared default instead of 0 GB (which painted every quant over-budget).
  • The native retry after a failed fork reply check re-derives its launch request with AutoBest off — as its own docs already claimed — instead of carrying fork-tuned params/quant/context into the native build (fork-only KV-cache types made that retry fail outright).

v2.2.0 - 2026-07-06

Coordinated LocalX release.

  • A LocalPilot launch no longer wipes hand-added .localpilot.toml sections (e.g. [mcp.servers.*] for Playwright/context7). Previously every launch overwrote the whole file with just the provider/permissions blocks LocalBox generates; it now merges those into the existing file, preserving every other section untouched. A pre-existing file that fails to parse as TOML now fails the launch loudly instead of being silently discarded.
  • The guided launcher's "Save as my default" now persists the Images (vision) toggle in DefaultLaunch. A saved vision-on recipe replays for that same model after restarting LocalBox, while other models still start text-only unless vision is explicitly enabled.
  • Docs now describe the weight-vs-VRAM fit estimate accurately: it is an advisory picker colour (green/yellow/red), not a guard that blocks a launch, and KV-cache pressure is not computed. Corrected the README and architecture notes that overstated it.
  • The LAN-gateway refusal message now names the real CLI flags (--password, --lan, --allow-public-no-auth) instead of retired PowerShell-module flags and a LOCAL_LLM_SERVE_PASS env var that was never read.
  • Clarified that GGUF weight downloads are integrity-checked at the reply-path smoke test, not by a per-file checksum (only the llama.cpp binary zips are sha256-pinned).
  • Restoring the shell env after a launch now returns a variable that was set to an empty string before launch verbatim, instead of removing it.
  • Raised the default LocalModelMaxOutputTokens from 4096 to 16384. Agentic coding replies routinely hit the old cap and stopped mid-word with no error (a legitimate finish_reason: length that nothing in the pipeline surfaced) — the no-think proxy's <think>-tag holdback was not at fault. The cap only bounds decode time for replies that actually need it; it does not reserve extra VRAM/KV-cache. docs/troubleshooting.md now documents the symptom and how to raise it further in settings.json.
  • Fixed the actual cause of replies vanishing silently mid-sentence: the no-think proxy's <think>-tag stripper discarded everything after a <think> tag that never got a matching </think> before the turn ended — and did so incrementally on every chunk, not just once at the end — with no warning and no effect on the stream's terminal marker, so the client saw a clean, seemingly-complete reply that was actually missing real content. Any local model that ever emits a stray/unclosed <think>-like substring mid-turn triggered this, independent of platform or model. Unterminated <think> spans are now flushed as visible text instead of discarded. (Shared localx-llama fix; re-pinned.)
  • Fixed a second, more common cause of the same symptom: an unterminated <think> span was only recovered once the stream reached its own natural end, so a false-positive <think>-like substring in ordinary (non-reasoning) text — confirmed via llama-server's own log to leave generation completely healthy, with no further reasoning activity — could freeze the visible reply for the rest of the turn if the caller gave up on the stream first. The no-think proxy now bails out of a <think> span that runs past 32KB without closing, flushing it as visible text and resuming normal streaming instead of waiting indefinitely. (Shared localx-llama fix; re-pinned.)

v2.1.5 - 2026-07-04

Coordinated LocalX release.

v2.1.4 - 2026-07-04

Coordinated LocalX release.

v2.1.3 - 2026-07-03

Coordinated LocalX release.

v2.1.2 - 2026-07-03

Coordinated LocalX release.

v2.1.1 - 2026-07-03

Coordinated LocalX release.

  • Streamed local-model replies are no longer truncated by up to ~7 characters. The no-think proxy held back the last few bytes of the visible text for split-<think>-tag safety and only released them after the stream's end marker ([DONE] / message_stop) — where the agent had already stopped reading — so every reply lost its final characters (e.g. "…What can I do for you?" arrived as "…What can I do f"). The proxy now flushes the held-back tail in-band, just before the end marker. (Shared localx-llama proxy fix; re-pinned.)

v2.1.0 - 2026-07-03

Coordinated LocalX release.

  • Agent launches now apply the permission/bypass gate again. On first launch of an agent that supports it (Claude, Codex, LocalPilot) LocalBox asks once whether to let the model act without per-action approval — defaulting off, saving your answer, and never enabling a bypass in a non-interactive session.
  • New --auto-best flag for serve/launch: applies the saved localbench profile (best-<model>.json) — tuned quant/context/mode plus KV-cache and n-cpu-moe — on a headless launch, the way the guided launcher already does.
  • The documented settings.json tunables now actually take effect: LlamaCppMlock, LlamaCppNoMmap, LlamaCppAgentParallel, LlamaCppAgentCacheReuse, LlamaCppNCpuMoe, LocalModelMaxOutputTokens, and the health/smoke timeouts were previously read from nothing.
  • An AMD GPU with no NVIDIA driver now installs and runs the Vulkan build instead of silently falling back to CPU.
  • A turboquant/mtpturbo build that fails its reply check now falls back to native llama.cpp instead of hard-stopping.
  • localbox stop spares the embedding server, so a chat-model stop no longer takes down the embeddings endpoint LocalMind uses.
  • A model server that exits during startup (e.g. OOM) now fails in seconds with its log path, instead of waiting out the full readiness timeout.
  • A configured VRAMGB now overrides the auto-detected value.
  • Removed phantom settings and the stale MCP page; corrected example model keys, the SECURITY version table, and a retired command name in a proxy error. An existing saved setting or a LOCAL_LLM_* env override is honored, and the chosen posture is printed before hand-off.

v2.0.2 - 2026-07-02

Coordinated LocalX release.

  • Guided launcher rebuilt around one live band. 2.0.1 drew a new inline viewport per menu and printed summaries between them, so screens stacked as leftovers. Every state now renders as one composed frame in a single terminal band: a hardware banner (GPU name + memory via nvidia-smi, then AMD tools; plain CPU-only fallback), the plan summary as a standing panel beside the menu, and content-sized boxes. Notices insert above the band into scrollback; launching hands the screen back cleanly and narrates its stages (server start + log path, model load, ready, reply check, agent start). Ctrl+C leaves the launcher from any menu.
  • Model rows are informative: key · name · size · context-as-words · images, size from the catalog or the downloaded file, with the fit traffic light coloring only the size segment.
  • Customize polish: the cursor survives toggles, "Save these as my default" returns to the launch menu, and the saved recipe's Strict toggle actually replays (the resolver never read it — fixed, per-model like quant/context).
  • Auto-tune is a full tuning menu: optimize-for, workload, engine, quality, memory, trial budget, runs per trial, cached-vs-fresh measurements, and save-vs-preview — every setting opens an option sub-menu with plain descriptions, technical names, the default and the current selection marked. An ℹ row explains every setting; the tune passes the current plan's quant/context so "Launch now" replays the winner, and it actually runs localbench findbest.

v2.0.1 - 2026-07-02

Coordinated LocalX release.

  • 1.x leftovers are now detected and named. A 1.x install.ps1 deployment survives the 2.0.0 upgrade badly: a symlink-mode install leaves a dangling ~/.local-llm/LocalLLMProfile.ps1 link (a red "Could not find file" error at every PowerShell start), a copy-mode install keeps running the stale 1.x launcher silently — and the cleanup logic lived in the deleted install.ps1, so nothing repaired either. localbox status and the guided launcher now detect the deployed 1.x launcher file, the lib/ module tree, and a PowerShell profile that still dot-sources LocalLLMProfile.ps1, and print one bounded notice with the remedy (localbox::migrate, unit-tested).
  • "Upgrading from 1.x" documented. docs/install.md gains the manual migration steps (remove the profile hook, delete the deployed 1.x files, install the binary) and the full 1.x → 2.x command mapping; the README links it and no longer cites the retired info -Commands.

v2.0.0 - 2026-07-02

Coordinated LocalX release.

  • Retired the PowerShell launcher, the Python no-think proxy, and the .NET TUI; LocalBox is now a single native binary. The local-llm/lib profile module, localbox-proxy, tui/, the Pester suite, and install.ps1 are gone — every behaviour they carried ships in the Rust binary, each hard-won invariant pinned by a golden test before its logic was ported. The no-think proxy runs in-process (the binary hosts it via localbox nothink-proxy), the guided launcher is localbox itself, and the first run seeds ~/.local-llm (defaults plus an editable catalog, never overwriting). Install is cargo install --path crates/localbox --locked or a release binary — LocalBox no longer needs PowerShell, .NET, or Python at runtime on any platform. CI keeps the tri-platform Rust gate and the cross-repo launcher-contract check.

  • Catalog inspection and cleanup in the binary. localbox info lists the configured models by tier (or one in detail; any of a model's names resolves), localbox purge stops servers and deletes downloaded model files strictly under the GGUF root, and localbox log tails the most recent server log.

  • Driver policy + resilience (localbox-tui::driver). Scrollback safety by construction: the terminal options pin an inline fixed-height viewport (no alternate screen, no whole-screen clear — content above the live band stays in native scrollback). Failures render as one bounded plain-language warning line (first line only, no backtrace vocabulary); a non-TTY stdout or an explicit plain request degrades to numbered-line menus with zero escape sequences; and UTF-8 console output is ensured best-effort on Windows so box-drawing renders without profile edits.

  • The Customize path (localbox-tui::customize). Progressive disclosure over the resolved plan: every row shows its current value and choosing one merges the delta back into the overrides. Auto-tune owns Engine + KV — with it on, those rows lock with a plain explanation (never a silent block) and any manual KV override is dropped so it cannot strand against the tuner's choice; turning it off restores manual control. Save is target-gated: serve cannot be saved as the default launch, refused with the friendly name in the reason.

  • Guided-launcher widgets (localbox-tui::ui). The model picker (key · DisplayName, dim [strict], [Show all tiers]/[Cancel] footers), the recommended-plan summary panel, and the exact five-item confirm menu (Launch now / Customize / Auto-tune / What do these mean? / Back to models) — backend-agnostic ratatui widgets whose rendered screens are pinned by TestBackend snapshot tests, including the no-jargon contract on the summary. Quant rows get the fit-aware traffic light (green fits / yellow tight / red over; unknown stays uncolored) from the shared fit classifier.

  • Guided-launcher core in Rust (localbox-tui: vocabulary + plan). The plain-language vocabulary (Run with / Quality / Memory / Speed / Images friendly names, memory-as-words at tokens×0.75, quality-as-hint+GB, the glossary) and the recommended-plan summary with the no-jargon contract as a test (quant/AutoBest/turboquant must not appear; the friendly labels must). The plan resolver pins the precedence: explicit overrides > the saved DefaultLaunch's cross-model preferences (target/engine/auto-tune) > the per-model definition > hard defaults, with quant/context applying from the recipe only for the model it was saved for; a returning user's plan IS their last-good plan, previews are non-committing by construction, and the .llm-default workspace override resolves by walk-up with nearer markers shadowing outer ones.

  • Security posture + serve guard in Rust (localbox-launcher::posture). The one-screen posture (permission/bypass gate status without prompting, the loopback-only agent proxy, serve-gateway exposure and auth, and the download pin posture incl. the trust-on-first-use spelling), the public-HTTP classifier that spares loopback/RFC-1918/link-local/localhost while flagging public addresses and DNS names on plain HTTP, and the serve guard: open no-auth HTTP on a public-looking address is refused with the remedy unless explicitly opted into; password-protected public HTTP stays allowed and visible. Discovery defaults derive from the home directory — never a hardcoded machine path.

  • Launch planning in Rust (localbox-launcher::orchestrate). One read-only resolution — model → context → expected GGUF path (no download) → vision projector (opt-in AND actually resolved) → free server port → full llama-server argv → proxy config → provider TOML → agent env plan — produces the inspectable plan that serves both DryRun (print it; commits no session state, creates no files) and the live launch. Vision is honest end to end: --mmproj enters the argv and supports_vision is declared only when a projector resolved; a present projector stays unused unless requested. The smoke-fallback rule is pinned: a non-native fork failing the smoke retries on native llama.cpp; native failing is a hard stop.

  • Launch smoke test in Rust (localbox-launcher::smoke). The degenerate-output detectors with thinking stripped FIRST (a reasoning model may legitimately fill its budget with <think> before a short answer): the [no output] marker, 8+ runs of flood punctuation (the GPU/driver-mismatch ////// signature), and 10+ consecutive identical tokens; verdicts name their failure mode (no response / think-only / degenerate), the Anthropic reply body parses fail-closed, and failure rendering is collapsed and bounded for the operator.

  • .localpilot.toml provider wiring in Rust (localbox-launcher::localpilot_config). The generated provider block pins the coupling rules by test: provider kind and endpoint move together (anthropic behind the no-think proxy, openai-compatible direct; base_url always ends in /v1), api_key_env follows the kind, the model's context is declared as providers.local.context_window — never a [harness] context_token_limit key — supports_vision is auto-declared only when the launch actually loaded a projector, max_tokens/context_window are omitted at zero, and bypass is a [permissions] profile = "bypass" config block, never argv.

  • Permission/bypass gating in Rust (localbox-launcher::permissions). The three agent gates (Claude skip-permissions, LocalPilot bypass, Codex bypass) resolve fail-closed: env override → persisted per-machine setting → a one-time first-run prompt that defaults OFF. A non-interactive session never silently enables anything and never persists; a preview/DryRun resolve is read-only; an explicit "no" persists a literal false (decided-no, not never-asked); Codex bypass is never defaulted; LocalPilot bypass is written to .localpilot.toml as [permissions] profile = "bypass", never argv. The tests build from a REAL empty settings file and assert the observable gate decision, and the settings store refuses catalog-only keys.

  • Agent env envelope in Rust (localbox-launcher::env). One plan function computes the launch's environment variables — it is both the DryRun snapshot and the live setter's source, so the two can never drift (the PowerShell version hand-mirrored the list in two places). The envelope saves every canonical variable before mutating and restores after: values present before come back, values absent before are removed. The canonical set carries its reasons in place (the 1800000ms SDK timeout for slow local prefill, the auto-memory and experimental-betas/ToolSearch disables, the no-think trio skipped under keep-thinking, the paired context/auto-compact caps, the conditional per-model image ceiling, and the LOCALBOX_* telemetry exports), and a test pins that every setter variable is inside the envelope. All over a store seam so tests never race the real process environment.

  • No-think proxy lifecycle in Rust (localbox-launcher::proxy). The launch-time proxy orchestration over the shared tri-state machine, with every effect behind a mockable seam: reap-before-probe (a dead-upstream orphan still answers /health and must never read as a live match), reuse-on-match with the gateway-logs rule (an owned proxy restarts so logs are captured; a foreign matching one is refused with the llm-stop remedy), repoint-on-mismatch (tear down and restart at the wanted target, never fail into a silent direct route), kill-stale-listener-first before binding (Windows SO_REUSEADDR), bounded readiness polling that tears down a never-ready start, owned-vs-any teardown (llm-stop reaps every owner on the port), and the netstat/lsof socket→PID parsers. The shared health state now distinguishes proxy-down (upstream up, proxy missing) from fully-down.

  • Native Rust workspace seeded (crates/localbox-launcher). The launcher contract now has its Rust implementation alongside the shipping PowerShell module (which stays authoritative until each piece is retired): the three-layer config load (defaults.json < llm-models.json < per-machine settings.json) over the shared precedence engine with catalog-only keys locked, and LlamaLauncher implementing the shared Launcher trait — catalog-backed model/quant/context resolution delegating to the shared domain crate, on-disk GGUF and mmproj resolution (configured module wins, deterministic auto-detect otherwise; a missing GGUF fails with the install remedy), per-mode install roots and llama-server binary resolution, KV-type capability per fork, free-port/wait/stop lifecycle with a recorded backend session, and ~/%VAR% path expansion. The version envelope passes the shared compatibility gate (api 1 / export 1, LocalBox/llamacpp declared).

  • The model catalog is now per-user, not shipped. llm-models.json is your own list (gitignored); the repo ships llm-models.example.json as the template. install.ps1 seeds ~/.local-llm/llm-models.json from it on first run and never overwrites an existing catalog, so your models stay yours and aren't published for everyone. Existing installs are unaffected; a fresh checkout falls back to the template. Add models with addllm or by editing your local catalog.

v1.2.1 - 2026-07-01

Coordinated LocalX release.

  • New default llm experience: a guided, plain-language launcher. When Spectre is available, llm now walks you through launching in plain words — pick a model, then Launch now (recommended) or Customize (Run with / Quality / Memory / Speed / Images), with a What do these mean? help screen — and returns to the model list when the agent exits. Recommended settings are pre-filled to fit your machine, so the common path is model → Launch now. The developer-oriented inline board is opt-in via LOCALBOX_BOARD=1; LOCALBOX_LEGACY_WIZARD=1 restores the old step wizard; llmc/llms remain the explicit classic/Spectre entrypoints.
  • Registered ornith35hapex in the model catalog — a decensored (Heretic) + APEX GGUF of Ornith-1.0-35B (4 tiers: Quality/Compact/I-Quality/I-Compact + vision mmproj, Parser qwen36, contexts capped for a 24 GB GPU, experimental, not the default). It reuses the existing serving glue (proxy 11435→8080, --mmproj). On wikitext-2 perplexity it is on par with the q3635ba3bapex baseline (6.96 vs 6.91) at comparable throughput and VRAM, and adds vision; trunk-only (no MTP draft head). Repo points at the public HuggingFace release C0deGeek/Ornith-1.0-35B-heretic-APEX-GGUF. Serve with llmdefaultserve ornith35hapex.

v1.2.0 - 2026-06-30

Coordinated LocalX release.

  • Start-LocalPilot -UseVision now actually loads the vision projector, and auto-declares it to LocalPilot. The LocalPilot agent launch previously hardcoded an empty projector path, so -UseVision set the AutoBest profile but never passed --mmproj to llama-server — an image-capable model ran text-only. It now resolves the model's mmproj.gguf (mirroring the Claude Code launch), guarded by an availability check so a missing projector gives a clear message and a text-only launch rather than a broken --mmproj. A real launch downloads the projector on demand; a -DryRun preview resolves the expected path without downloading. On that vision launch LocalBox also writes supports_vision = true into the [providers.local] block of the generated .localpilot.toml, so LocalPilot accepts image input zero-config. The default (no -UseVision) path loads no projector and declares nothing — unchanged. See docs/harness-mode.md → "LocalPilot harness".

v1.1.0 - 2026-06-29

Coordinated LocalX release.

  • A CPU-only embedding server (llmembedserve). A small, self-contained sibling of llmdefaultserve that serves a GGUF embedding model through llama-server's OpenAI-compatible POST /v1/embeddings on a dedicated loopback port (8090 by default), forced onto the CPU (-ngl 0) so it costs zero GPU VRAM. That CPU rule is load-bearing for fair benchmarking: a GPU-resident embedding model would steal VRAM from a chat model running alongside it, so a benchmark pairing the two would see a degraded chat model on the embeddings side only — keeping embeddings on the CPU leaves the chat model byte-identical. The server has its own port, process, and lifecycle (llmembedstop), independent of the chat server. Default model: Qwen3-Embedding-0.6B (GGUF Q8_0, Apache-2.0, 1024-dim, --pooling last), acquired on first run into the models dir (never committed) and overridable via EmbedModelRepo/EmbedModelFile/ EmbedModelRoot/EmbedPort/EmbedPooling in settings.json. -WhatIf renders the exact served command without acquiring a model or launching anything; Test-LocalLLMEmbedEndpoint probes a running server and returns the vector dimension. See docs/harness-mode.md → "CPU embedding server".

  • no-think proxy v0.4.0 — normalizes system messages so strict (qwen-family) chat templates accept Anthropic agentic clients. llama.cpp's qwen3 template hard-rejects any system message that is not the first message (raise_exception('System message must be at the beginning')), surfacing as 400 Unable to generate parser for this template. Anthropic clients such as Claude Code put the base prompt in the top-level system field and can inject a second role: system message inside messages (e.g. a SessionStart hook); llama.cpp renders both, so the in-array one lands second and the template raises. The proxy now folds any in-array system message into the top-level system field (preserving existing content blocks and their cache_control) and removes it from messages; for OpenAI-form requests (no top-level system) it collapses misplaced/duplicate system messages into a single leading system message. Default-on; opt out with NO_THINK_PROXY_MERGE_SYSTEM=0. defaults.json NoThinkProxyRequiredVersion bumped to 0.4.0. This unblocks driving the default local model from Claude Code (and any Anthropic-native harness) through llmdefaultserve. (The no-think proxy remains deprecated and kept one release for non-LocalPilot clients.)

v1.0.0 - 2026-06-24

Coordinated LocalX 1.0 release. First stable launcher surface.

  • Dead-code cleanup: removed 7 unused PowerShell helpers. No call sites remained for Convert-ToPosixPath, Get-HuggingFaceModelFiles, Get-LlamaCppTemplatesDir, Set-LocalBoxTuiSetting, Invoke-LocalBoxTuiLaunch, Get-LocalBenchLauncherBestConfig, or Get-LocalBenchLauncherBestConfigCandidates. Also dropped the stale local-llm/bench-history.jsonl .gitignore entry (nothing writes it). No behaviour change. (The Ensure-LlamaBenchExe / Ensure-LlamaPerplexityExe finders/installers were left in place: they are part of the LocalBench launcher contract surface, not dead.)

  • llmdefaultserve -DryRun now previews the recipe that actually launches. When the DefaultLaunch recipe selects a non-default quant, the dry run used to show the model's default quant (e.g. APEX-Balanced.gguf with q8 KV-cache args) while the live launch ran the selected one (e.g. APEX-I-Quality.gguf with turbo3 args) — the quant was applied only on the live path. The selected quant is now resolved for both paths, so the preview renders the same GGUF + AutoBest/KV-cache recipe the live launch consumes; a dry run reverts the change afterwards, so previewing still commits no session state.

  • Stale no-think proxy is now diagnosable instead of a bare 502. When the no-think proxy is up but the upstream model server is down, a request returned only 502 Bad Gateway. A bounded, non-blocking health probe now distinguishes that stale state from a fully-down or healthy stack and recommends the fix (llmstop; llmdefaultserve); it is surfaced when a headless serve's smoke test fails, and never blocks the launch.

  • llmdefaultserve — headless model serve for CLI / agent / CI. Brings up the DefaultLaunch model as a background llama-server + no-think proxy (loopback) with a visible-response smoke test and does not attach an interactive agent or tear the endpoint down on exit — unlike llmdefault, whose agent-attach (Start-LocalPilot) stops the server and proxy when the attached process exits. For driving the model from a separate localpilot/claude process. Distinct from the LAN serve gateway llmserve (binds 0.0.0.0 with auth); this one is loopback-only. Stop with llmstop.

  • Fixed the -DryRun launch-plan "Agent argv" error. An empty extra-args list arrives as a nested empty array element (helpers return ,$extras); the real launch splats it away, but the preview's Format-LocalLLMArgvLine failed to bind it as a string. The formatter now flattens nested elements, so the preview renders the agent command line cleanly.

  • Removed the legacy bench history viewer. The obench command and its 70-bench.ps1 module are gone — LocalBench owns benchmarking. The old ~/.local-llm/bench-history.jsonl (written by the retired ospeed helper) is no longer read; use LocalBench for benchmarking and tuning.

v0.3.0-beta.3 - 2026-06-18

2026-06-19 - Launch safety

  • Bypass is no longer a default. Launching LocalPilot through LocalBox no longer passes --bypass unconditionally, and Codex no longer defaults --dangerously-bypass-approvals-and-sandbox to on. Each is now a first-run, persisted decision (LocalPilotBypass / CodexBypassApprovalsAndSandbox) that defaults off and fails closed in non-interactive sessions, mirroring the existing LocalModelSkipPermissions prompt. The active posture is shown in Show-LocalBoxSecuritySummary and every -DryRun launch plan, and an env override (LOCAL_LLM_LOCALPILOT_BYPASS / LOCAL_LLM_CODEX_BYPASS) applies per launch. See docs/settings.md. This restores parity with LocalPilot's "bypass is never the default" posture for users who enter through LocalBox.

  • Codex bypass default fully enforced at config load. Import-LocalLLMConfig no longer injects CodexBypassApprovalsAndSandbox = $true for an unset key, which had quietly re-enabled the default-on posture for a fresh machine even though the launch path itself never prompted. The key now stays absent when unset — exactly like LocalPilotBypass — so the security summary reads "undecided" and the resolver reaches its first-run prompt (interactive) or fails closed (non-interactive). A regression test loads the real merged config (not a stubbed one) to keep this from regressing again.

  • Fixed the LocalPilot install hint thrown by Start-LocalPilot: the crate is cargo install localpilot (not localpilot-cli), matching the troubleshooting doc.

  • Repinned the turboquant llama.cpp build to the forked tqp-v0.2.0 (b9901, lazy-grammar), fixing the q3635ba3bapex 400 empty grammar stack after <think> failure on constrained-decoding requests. Dropped the stale TheTom turboquant refs.

2026-06-18 - Versioning policy

  • Documented the SemVer policy in docs/versioning.md: break → major, add → minor, fix → patch, plus the pre-1.0 rules (breaking changes bump the minor 0.MINOR; features and fixes bump the patch) and the deprecation-vs-removal distinction. Linked from the docs index.

2026-06-18 - Deprecation

  • Marked localbox-proxy/no-think-proxy.py deprecated (scheduled for removal). LocalPilot strips <think> blocks natively on both provider paths and suppresses the thinking request shape itself, so the out-of-band proxy is no longer needed for the LocalBox → LocalPilot path.

2026-06-17 - Documentation restructure

  • Split the long top-level README.md into a lean overview plus owned docs/ pages (install, harness mode, llama.cpp modes, usage, model management, settings, MCP, auto-tuner, wizard/TUI, architecture, troubleshooting), indexed by docs/README.md.
  • Fixed the LocalPilot install hint in troubleshooting: the crate is cargo install localpilot (not localpilot-cli).
  • Added an in-repo wiki source (docs/wiki/) that is one-way CI-synced to the GitHub Wiki, plus an offline link check over the docs.

v0.3.0-beta.2 - 2026-06-15

Coordinated LocalX beta release.

  • Serve gateway now discloses its safe operating posture (LAN/VPN-only, HTTPS in front) in the client instructions and the security banner; the guard still refuses open public HTTP.

  • Conformed the bundled tuner best-config fixture to the current tuner-best-config schema (added the required profile field) so producer and consumer agree on the versioned contract.

  • llm-update now refreshes installed artifacts after source updates. LocalPilot is reinstalled from its updated checkout when it fast-forwards, so the localpilot CLI on PATH no longer lags behind the repo. Added -RefreshInstalled to force redeploy/rebuild from already-current checkouts; Update-LocalPilot -RefreshInstalled does the same for the standalone helper.

2026-06-12 - v0.3.0-beta.1 LocalX release train

  • Aligned LocalBox with the coordinated LocalX v0.3.0-beta.1 beta release.
  • Kept the LocalBench launcher contract binding and updated the no-think proxy required version marker for this release train.

2026-06-03 - Serve gateway rename

  • Remote launch target renamed to Serve. The TUI action contract now uses serve, the command is llmserve, the password environment variable is LOCAL_LLM_SERVE_PASS, and the docs describe the gateway as serving any Anthropic-compatible agentic client, with LocalPilot as one example.

2026-05-24 — Terminal.Gui TUI and catalog polish

  • llmtui documented as the explicit Terminal.Gui preview path. README now covers publish/install, profile resolution, core navigation keys, and the LocalBench handoff key.
  • Spectre model catalog layout fixed. The dashboard now uses explicit column widths and a compact quant list so the context column stays readable instead of wrapping to one or two characters.
  • MTP catalog contexts expanded. q3535ba3bmtp and genesisv2 now expose an explicit 256k context key while keeping default context at 128k.
  • Quant naming clarified. README notes that labels like mtp, mtp-apex, and mtp-q8kp are model-local quant keys that map to concrete GGUF filenames.
  • TUI packaging added. install.ps1 -InstallTui publishes LocalBox.Tui and LocalBench.Tui when available; llm-update -InstallTui refreshes installed TUI binaries.
  • lbtui wrapper added. LocalBox now exposes a LocalBench.Tui entrypoint alongside llmtui, and inline help documents both.

2026-05-24 — Dropped Ollama backend (breaking)

Breaking changes

  • Ollama support has been removed. LocalBox now targets llama.cpp's llama-server exclusively. The -Backend ollama parameter, the Ollama process control (Start-OllamaApp, Wait-Ollama, Stop-OllamaModels, …), Modelfile-based alias creation, Ollama strict siblings, the Ollama remote gateway path, and every shortcut that pulled or rebuilt Ollama aliases (init, initmodel, ostop, qkill, ops, cleanorphans, listorphans, ospeed) are gone. If you still need that path, check out the ollama-classic git tag.
  • -Chat and -Q8 flags removed from per-model shortcuts. llama-server has no chat REPL, and Q8 KV was the Ollama env var (OLLAMA_KV_CACHE_TYPE=q8_0). Use -KvK q8_0 -KvV q8_0 (or the wizard's KV-cache step) for the equivalent llama.cpp setting.
  • Catalog scalars removed. MinOllamaVersion, OllamaAppPath, OllamaCommunityRoot, KeepAlive, RequireAdvertisedTools, LlamaCppCoexistOllama are stripped from defaults.json and from any merged settings at load time. SourceType: remote model entries are no longer supported (all in-catalog entries are gguf).
  • ollama-proxy/ folder renamed to localbox-proxy/. The deployed location is now ~/.localbox-proxy/. The Python no-think proxy itself is unchanged in behavior; it now targets 127.0.0.1:8080 by default instead of 127.0.0.1:11434. The enforcer-claude.ps1 Ollama wrapper has been deleted.
  • Save-LLMDefaultLaunch schema simplified. No more Backend, UseQ8. Existing DefaultLaunch entries with those fields will ignore the obsolete keys; re-save via the wizard if you want the schema clean.

Other

  • Invoke-Backend no longer takes -Backend; it always dispatches to llama.cpp. launch-chat action removed.
  • llm-status is now just Invoke-LlamaCppStatus. Show-OllamaStatus is replaced by Show-LocalBackendStatus + Show-ConfiguredGgufQuants in the dashboard.
  • The wizard's backend step is now a mode picker (native / turboquant / mtpturbo). Default is LlamaCppDefaultMode (still native).
  • Codex launches always use the OpenAI-compatible llama-server provider; the --oss --local-provider ollama path is gone.

2026-05-08 - Codex target and default launch recipes

  • ostop now leaves Ollama stopped. It no longer restarts the Ollama app after teardown.
  • llm-stop. Added a hyphenated all-backend stop command alongside llmstop / unloadall.
  • Codex launch target. Model shortcuts and the wizard now support -Codex / Codex as a peer to Claude Code and LocalPilot. Ollama launches use Codex's local Ollama provider; llama.cpp launches pass a custom OpenAI-compatible provider pointed at the selected llama-server /v1 endpoint.
  • Default launch recipes. The wizard can save the selected model, target, backend, context, quant, strict/Q8 flags, llama.cpp mode, KV cache, and AutoBest profile into DefaultLaunch so llmdefault can replay the full recipe.
  • Wizard default is native selectable. llm/llmmenu now use an in-repo arrow-key picker by default. llms opens the Spectre wizard explicitly, and $env:LOCAL_LLM_USE_SPECTRE=1 opts llm back into Spectre.
  • Context menu noise reduced. Removed visible fast, deep, and bare 128 context aliases from the catalog and new-model defaults. Use 32k, 64k, 128k, and 256k; legacy aliases still resolve for old commands and saved AutoBest profiles.

2026-05-07 - Suite updater, LocalPilot cleanup

Added

  • llm-update / Update-LocalLLMSuite. Checks LocalBox, LocalPilot, and LocalBench when they are installed as git checkouts, fetches upstream state, and fast-forwards only when an update is available. Missing companions are skipped, current checkouts are reported as current, and diverged/no-upstream checkouts are left untouched with a reason.
  • LocalBoxRoot setting. install.ps1 now records the source checkout used for installation so deployed copy-mode profiles can find the LocalBox repo for future self-updates.
  • LocalPilot-only command surface. Removed the old shorthand and pre-rename aliases. Model launches now use the explicit -LocalPilot switch, and the default shortcut is llmdefaultlocalpilot.

2026-05-03 — Wizard back-step nav, full-quant backfill, install fix

Added

  • updatellm <key> / Update-LocalLLMModelQuants. Backfills missing quants on an existing GGUF entry by re-fetching its HF repo and merging any quant codes not already present. Existing Quants, QuantSizesGB, and QuantNotes entries are preserved verbatim — only new keys are added (auto-generated note via New-LocalLLMQuantNoteText). -DryRun previews the additions without writing. Applied to the catalog: qcoder30 3 → 23, qcodernext 3 → 23, q27heretic 4 → 6, q27hauhau 2 → 10. q36plus (HF gated, returns 401 on the API) and q36heretic (subdirectory-organized repo layout) were left untouched.

Fixed

  • LM wizard couldn't go back one step. Both Start-LLMWizardClassic and Start-LLMWizardSpectre were a flat while ($true) / continue loop where every [[Back]] returned to the top (re-pick model). The Spectre quant menu had no Back at all (only [[Keep current: …]]). Both wizards are now step-state machines ('model' → 'quant' → 'context' → 'action' → 'q8' → 'launch'); each step's $null return walks back exactly one step. Quant menu in Spectre now has both [[Keep current: <q>]] and [[Back]]; the classic quant menu uses a letter shortcut (k = keep current, 0 = back) via a new -LetterChoices hashtable on Read-LLMChoiceIndex. The q8 prompt also accepts b (classic) / [[Back]] (Spectre) to walk back to action selection.
  • addllm could pick up imatrix calibration files. Top-level *.imatrix.gguf (mradermacher's calibration data, not a quantized model) used to pass the file filter and only got dropped because Get-HuggingFaceQuantCode returned $null for it. Now excluded explicitly in both Add-LocalLLMModel and Update-LocalLLMModelQuants so a future quant-code regex change can't accidentally include them.
  • install.ps1 failed when -Profile was passed. The [switch]$Profile parameter shadowed PowerShell's $PROFILE automatic variable inside the script, so $PROFILE.CurrentUserAllHosts resolved to [switch].CurrentUserAllHosts (no such property → $null), then Test-Path $null threw Value cannot be null. Renamed the parameter to [switch]$SetupProfile with [Alias("Profile")] so existing -Profile invocations still bind. Also collapsed the convoluted installFiles flag computation into one line, and fixed a cosmetic message bug where \$PROFILE was meant as a literal but PowerShell escapes with backtick (`), not backslash — the message accidentally interpolated the auto-variable.

2026-05-03 — Spectre wizard

Added

  • Spectre-rendered llm wizard. Start-LLMWizard now dispatches to Start-LLMWizardSpectre when PwshSpectreConsole is available, falling back to Start-LLMWizardClassic otherwise. The Spectre flow uses Read-SpectreSelection for model / quant / context / action picks and Read-SpectreConfirm for the -Q8 toggle, and renders the full Show-ModelCatalogSpectre table above the model picker so quant fit / size / built status stay visible while choosing. Same env switch as the dashboard: $env:LOCAL_LLM_NO_SPECTRE=1 forces the classic wizard.
  • llmc escape hatch. New llmc function calls Start-LLMWizardClassic directly, bypassing Spectre regardless of availability — useful when a Spectre render bug makes the rich wizard unusable.
  • Wizard error trap. Each Spectre prompt is wrapped in Invoke-LLMWizardStep; on exception, Save-LocalLLMWizardError records the full trace (timestamp, context tag, exception type/message, InvocationInfo.PositionMessage, ScriptStackTrace, inner exception) to ~/.local-llm/wizard-errors.log and pauses with Press Enter to continue so a Spectre live-display refresh can't scroll the trace off screen. Inspect with llmlogerr [-Lines 80]; reset with llmlogerrclear.

Fixed

  • Spectre markup parse errors crashed wizard prompts. Choice labels like [Back], [Cancel], [Show all tiers], [Keep current: …], and the fit tags [fits]/[tight]/[over]/[?] were passed straight to SelectionPrompt, which interprets […] as Spectre markup — every render frame threw Encountered malformed markup tag … and the live-display refresh hid the trace. Sentinels are escaped as [[…]]; user-supplied text (display names, quant/context keys, notes, $ModelKey titles) is routed through ConvertTo-LocalLLMSpectreSafe; fit tags became proper colored markup ([green]fits[/], [yellow]tight[/], [red]over[/], [grey50]?[/]).
  • Renderables leaked into captured pipeline output. Format-SpectrePanel / Format-SpectreTable return a Spectre.Console.Renderable; PwshSpectreConsole's format.ps1xml only renders that to ANSI at Out-Default. Inside Show-ModelCatalogSpectre (called from Select-LLMModelKeySpectre, captured by $modelKey = …), the Panel and Table objects bubbled up into $modelKey, producing arrays like [Panel, Table, "qcoder30"] and tripping Cannot convert value to type System.String on the next Get-ModelDef -Key $modelKey. All six Format-Spectre* call sites now pipe to | Out-Host so the renderable renders eagerly and emits nothing to the caller's pipeline. Side effect: the catalog table that was silently swallowed during the wizard now appears as intended.

2026-05-02 — 256k Qwen3-Coder profile, VRAM-aware tradeoffs, per-quant/context notes

Added

  • qcoder30 256k context. Added "256": 262144 to the Qwen3-Coder-30B-A3B Heretic model and a new iq4xs quant (Qwen3-Coder-30B-A3B-Instruct-Heretic.i1-IQ4_XS.gguf, ~16.5 GB). The 256k profile only fits a 4090 with IQ4_XS weights + q4_0 KV cache (~6 GB at 256k); use qcoder -Ctx 256 -Quant iq4xs.
  • qcodernext (experimental). New entry pointing at mradermacher/Huihui-Qwen3-Coder-Next-abliterated-i1-GGUF — the 80B/3B-active hybrid DeltaNet+Attention coder. Quants: iq1m, iq2s, iq3xxs. Only iq1m (~18.1 GB) fits a single 4090 with any KV headroom; flagged as experimental and "tight on 4090" in the display name.
  • Per-model Description, QuantNotes, ContextNotes catalog fields. Free-form strings keyed by quant/context name. Backfilled across the existing catalog so users can see file sizes, KV pressure, and "when to pick this" guidance without leaving the launcher.
  • info / llmdocs / llm wizard surfaces. Show-ModelCatalog, Show-LLMDynamicModelSummary, Select-LLMModelKey, Select-LLMQuantKey, and Format-LLMContextLabel all render the new notes inline. The current default quant is marked with * in the per-quant list.
  • addllm -Description, -QuantNotes, -ContextNotes. Optional params on Add-LocalLLMModel / addllm that round-trip into the catalog entry. Notes are hashtables (@{key='note'}) keyed by the same quant/context shortname.
  • -Q8 + long-context guard. Invoke-ModelShortcut refuses -UseQ8 whenever the resolved num_ctx exceeds the Q8KvMaxContext ceiling. The error message tells the user to drop -Q8, lower -Ctx, or raise the threshold.
  • VRAM-aware recommendations. New top-level VRAMGB setting plus Get-LocalLLMVRAMInfo helper. Auto-detects via nvidia-smi --query-gpu=memory.total (largest GPU on a multi-card box). Override via Set-LocalLLMSetting VRAMGB 32. The dashboard surfaces the resolved value + source (configured / auto / fallback).
  • QuantSizesGB per-quant numeric field. Drives a [fits] / [tight] / [over] badge next to each quant in info and the wizard, computed against the host's VRAM (weight-budget heuristic: [fits] when the model leaves >=7 GB headroom for KV, [tight] when only ~2 GB headroom, [over] otherwise). Backfilled across qcoder30, qcodernext, q36plus, q36heretic, q27heretic, q27hauhau.
  • Q8KvMaxContext now scales with VRAM by default. Removed the explicit 131072 literal from the catalog. The guard derives (VRAMGB - 16) * 16384 (floored at 64k) when not pinned, so a 5090 (32 GB) gets ~256k while a 4090 (24 GB) gets ~128k. Override still works via Set-LocalLLMSetting Q8KvMaxContext.
  • Quant notes rewritten to be VRAM-agnostic where possible. The hand-written notes describe quality/use-case (no longer "partial offload on a 4090"); the per-quant [fits]/[tight]/[over] badge is the live verdict for the host's actual VRAM.

Why

Picking a quant and context blindly was costing real time — Q4_K_M is fine at 64k but cannot fit 256k KV; Q6_K is too heavy at any long context; -Q8 looks free until it OOMs at 128k+. The notes encode the tradeoff directly next to the selector, and the guard prevents the worst foot-gun (-Q8 -Ctx 256) from ever launching.

VRAM auto-detection was the next cliff: every recommendation in the catalog implicitly assumed a 24 GB 4090. A 5090 user (32 GB) should see Q5_K_M as [fits], not "partial offload"; a 4080 user (16 GB) should see most 35B variants flagged [over] and not waste time downloading them. The fit badge gives a per-host verdict without the user having to do KV-cache arithmetic.

The catalog gained one realistic 256k coder option (qcoder30 -Ctx 256 -Quant iq4xs) and one aspirational one (qcodernext) so the "uncensored 256k on a 4090" question has a documented answer instead of trial-and-error.

2026-04-30 — Per-machine settings + auto-install LocalPilot

Added

  • ~/.local-llm/settings.json — per-machine overlay for the catalog. Top-level scalars (LocalPilotRoot, OllamaAppPath, Default, KeepAlive, RequireAdvertisedTools, NoThinkProxyPort, LocalModelTools, LocalPilotRepoUrl, etc.) load from llm-models.json first, then any matching keys in settings.json override. Models and CommandAliases are catalog-only and protected from override.
  • Set-LocalLLMSetting <Key> <Value> — writes to settings.json and reloads. Pass $null/"" to remove a key. Refuses Models/CommandAliases.
  • LocalPilotRepoUrl config field, defaulting to https://github.com/C0deGeek-dev/LocalPilot.
  • Ensure-LocalPilotInstalled — called by Invoke-LocalPilotCli before doing anything. If the configured LocalPilotRoot doesn't contain src/entrypoints/cli.tsx, it prompts Clone <url>? [y/N] and runs git clone on confirmation. Aborts with a clear instruction otherwise.
  • settings.json added to .gitignore so per-machine config never lands in the repo.
  • install.ps1 prints a tip pointing at Set-LocalLLMSetting for fresh-machine setup.

Why

Cloning the public repo onto a different machine should not require editing llm-models.json to fix LocalPilotRoot (and risking merge conflicts with future pulls). LocalPilot launches should do the obvious thing on a fresh machine instead of failing because no checkout is around.

2026-04-30 — LocalPilot rename

The external harness fork was renamed to LocalPilot. Propagated through this project:

  • JSON config field renamed to LocalPilotRoot.
  • Internal CLI wrapper renamed to Invoke-LocalPilotCli.
  • Switch parameter renamed to -LocalPilot on Start-ClaudeWithOllamaModel, Invoke-ModelShortcut, and the per-model shortcut functions.
  • User-visible labels updated: launcher banner, wizard action label, install diagnostics, README, quick reference (llmdocs).
  • Existing local folder paths were not renamed and still work as the configured LocalPilotRoot.

2026-04-29 — second-pass refactor

Reviewed the project, then ran a single-day refactor pass guided by an explicit plan (plan.md, retired into this changelog).

Bugs fixed

  • Persona pollution. LocalLLMProfile.ps1 had a hardcoded "You are Qwen, created by Alibaba Cloud" prepended to every model launch — wrong for Devstral and even somewhat wrong for the Qwen variants whose GGUF templates already self-identify. Removed the persona layer entirely; the system prompt now contains only universal tool-use rules, plus an opt-in deferred-tool-schema block (gated on LimitTools).
  • enforcer-claude.ps1 rewritten. The wrapper used to hardcode qcoder30 and bypass the no-think proxy by pointing at localhost:11434. Now it reads Default from llm-models.json (or $env:ENFORCER_MODEL), routes through the proxy on 11435, self-starts the proxy if needed, and sets the same thinking/caching/attribution env stack as the main launcher.
  • Legacy harness stub deleted. It was a one-liner that called the old harness wrapper with no args and ignored everything. The flag-based LocalPilot launch path covers it now.
  • Tool-support detection rewritten. Test-OllamaModelSupportsTools used to grep ollama show text for the literal word "tools" — which could match unrelated lines. Now POSTs to /api/show and checks the structured capabilities array. Falls back to the regex if the API is unreachable.
  • Devstral parser confirmed correct. Parser: "none" was the right call (its GGUF self-templates with persona, [SYSTEM_PROMPT]/[TOOL_CALLS] tags, and capabilities=[completion,vision,tools]). Documented inline via a ParserNote field.
  • init -Stale parameter shadow bug. Initialize-LocalLLM declared [switch]$Stale; the body did $stale = @(Get-StaleModelAliases). PowerShell variables are case-insensitive, so the assignment tried to coerce an array into a SwitchParameter and failed silently, leaving $stale as the boolean $true. Renamed the local to $staleEntries.

Added capabilities

  • Per-model Tools allowlist. Start-ClaudeWithOllamaModel now takes -Tools; Invoke-ModelShortcut reads the optional Tools field from the model def, falling back to the global LocalModelTools. No models populated yet — capability only.
  • Auto-generated alias prefixes. Added ShortName field per model. Register-ModelShortcuts walked ShortName × Contexts × actions and registered PowerShell aliases. Pruned the 30 hand-maintained CommandAliases entries to {}.
  • Parser-version stamping. New-OllamaModelFromSource now writes a sha256-hash sidecar at <profile-root>\parser-versions\<aliasname>.txt. Test-ModelAliasFresh, Get-StaleModelAliases, init -Stale, and the info dashboard surface stale aliases (parser config drifted since build).
  • Default model. Added "Default" field at the top of llm-models.json. Get-DefaultModelKey reads it (with a recommended-tier fallback). New shortcuts: llmdefault, an LocalPilot default shortcut, and llmdefaultchat. Used by the enforcer.
  • ThinkingPolicy per model. Either strip (default) or keep. keep mode bypasses the no-think proxy, points ANTHROPIC_BASE_URL at Ollama directly, and skips the thinking-disable env vars. Set on q36opus47abl. Launcher banner shows the active mode.
  • Configurable OLLAMA_KEEP_ALIVE. Top-level KeepAlive field; Set-OllamaRuntimeEnv reads it (defaults to "-1").
  • Wait-Ollama resilience. Deadline bumped 20s → 60s. After 5s of waiting, prints Waiting for Ollama and adds a . every 2s.
  • Bench history persistence. Test-OllamaSpeed now appends to <profile-root>\bench-history.jsonl per run. Show-LLMBenchHistory [-Model] [-Last N] and the short obench alias display recent runs.
  • Header truth. File header lost the "+ LM Studio" advertisement (LM Studio support was never implemented). Now states "Windows / PowerShell only — does not work in WSL/bash."

Architectural changes

  • Flag-based shortcut scheme (Option C). Replaced ~135 multi-suffix functions with 9 model functions: dev, qcoder, q36, q36hau, q36p, q36h, q27, q27hau, qop. Each takes -Ctx, -LocalPilot, -Chat, -Q8, and (where applicable) -Quant. Introduced Get-ModelShortcutName and Unregister-AllModelShortcuts; Register-ModelShortcuts is now idempotent.

Deferred

  • Diagnostic logging on tool-call failure. Naive stderr-tee breaks Claude Code's interactive terminal; needs better design (probably a debug-mode flag rather than a wrapper).

Pre-2026-04-29

Project predates this changelog. The state at the start of this round:

  • Single 2,506-line LocalLLMProfile.ps1 engine with JSON catalog (llm-models.json).
  • Per-(model, context) Ollama aliases.
  • Hand-maintained CommandAliases map.
  • HTTP proxy on 11435 stripping Anthropic thinking/reasoning fields.
  • Hardcoded "You are Qwen" persona prepended to every launch.
  • enforcer-claude.ps1 hardcoded to qcoder30 and pointing at the wrong port.