@octocodeai/freellama-mcp-server 0.9.0 → 0.12.0

package/docs/BENCHMARK.md

@@ -0,0 +1,381 @@
+# Benchmark
+
+One run, measured on 2026-11-14 on an Apple M2 Pro (32 GB): FreeLlama against Ollama alone,
+llama.cpp, llama-swap and LocalAI, all loading the same hash-verified model file. Every one of the
+360 measured answers was correct and byte-identical for that single extraction fixture.
+This is separate from [result-check accuracy and local tool-calling evidence](RELIABILITY.md).
+The current explicit [managed llama.cpp text path](LLAMACPP.md) is also separate from these
+Ollama-backed FreeLlama measurements.
+
+![Ollama, FreeLlama and the most popular local runners on six measured KPIs](assets/benchmark-overview.png)
+
+Then the [features](#features): what FreeLlama's management changes, each measured against the
+same work without it.
+
+## What the measured scenarios show
+
+FreeLlama manages model selection, admission and receipts around external engines. The
+FreeLlama rows below used Ollama. In this workload, its layer added a few milliseconds and
+used about 30 MB of server RSS. These results describe the recorded configurations:
+
+| Scenario | Setup measured here | Observations in this run |
+|---|---|---|
+| An agent (Claude Code, Cursor, scripts) that offloads bounded work to local models on a machine that also runs editors, browsers and builds | **FreeLlama on Ollama** | Same speed as Ollama (274 vs 272 ms median, first streamed token 95 vs 93 ms); its own work is 3.7 ms per task and its server uses 26–31 MB. Its built-in workflow prompts passed 45/46 completed graded cases versus 25/48 for one-line prompts (planned-denominator limits below), in less time and fewer tokens. An interactive task behind 8 queued ones answered in 1.40 s instead of 3.52 s; repeated code searches and sweeps reuse saved work (55.9 s → 0.13 s, 49.8 s → 18 ms); long tasks return a job id in 29 ms. It refuses a load that would not fit and names what holds the memory (live: needed 7.2 GB, 4.2 GB free, refused), selects eligible installed models with runtime admission, and returns a receipt with per-phase timing. 9.7 ms CPU per task. MCP server, CLI and an Ollama/OpenAI/Anthropic gateway. |
+| One application, one model you chose, little else running | **Ollama alone** | Fastest here by 2 ms (272 ms), one-line install, the largest model library. You size models and watch memory yourself. |
+| Many concurrent requests to one model | **llama.cpp `llama-server`** | Slower per request here (336 ms) but 291 tasks/min at 4 clients against 225 for Ollama, which serves one request at a time unless `OLLAMA_NUM_PARALLEL` is set (not measured). Model loads at process start. |
+| Several models swapped on demand behind one OpenAI endpoint, on any engine | **llama-swap** | 332 ms, the best 4-client throughput (305 tasks/min), 1.58 s to load on first request. Limits are pools you declare, not measured memory. |
+| One server for many modalities and backends (images, speech, video) | **LocalAI** | About 3× slower on this task (802 ms; 541 ms to first token) and 3.39 s cold, in exchange for breadth. |
+| A desktop app for people | LM Studio, Jan | Not measured here. |
+
+FreeLlama and Ollama are not either-or: FreeLlama runs on the Ollama you have, and Ollama's own
+clients keep working through it.
+
+## KPIs
+
+Six setups, the same request, 60 measured requests each over two run orders (forward and
+reversed), after 2 warm-ups per pass.
+
+| KPI | What it measures | How | Ollama alone | FreeLlama | FreeLlama streamed | llama.cpp | LocalAI | llama-swap |
+|---|---|---|---:|---:|---:|---:|---:|---:|
+| Correct answers | The task's JSON, field by field | Exact match against the known answer | 60/60 | 60/60 | 60/60 | 60/60 | 60/60 | 60/60 |
+| Median task time | What one request costs | Client wall clock, warm model, pooled over both orders | **272 ms** | **274 ms** | **273 ms** | 336 ms | 802 ms | 332 ms |
+| p95 task time | The slow tail | Same samples | 290 ms | 303 ms | 299 ms | 375 ms | 911 ms | 372 ms |
+| First token | When text starts to appear | Client clock to the first non-empty content | 93 ms | — | 95 ms | 128 ms | 541 ms | 130 ms |
+| Throughput, 1 client | Tasks finished per minute in sequence | Verified tasks ÷ measured window, mean of both orders | 209 | 206 | 217 | 178 | 76 | 178 |
+| Throughput, 4 clients | Tasks per minute under contention | 4 concurrent clients, 30 requests (forward pass) | 225 | 221 | 222 | 291 | 145 | 305 |
+| Cold request | First answer when the model is not loaded | 3 samples after an unload (`keep_alive: 0`) or a process restart | 1.13 s | 1.21 s | 1.17 s | n/c¹ | 3.39 s | 1.58 s |
+| Decode speed | Engine token rate | Ollama's own timing, median of each order, averaged | 124 tok/s | 125 tok/s | 124 tok/s | — | — | — |
+| Layer overhead | Time added on top of the engine | Client time minus Ollama engine time, 40 requests | — | 3.7 ms (p90 5.6) | — | — | — | — |
+| Layer memory | RAM the management process itself uses | Peak RSS of the FreeLlama server process | — | 26–31 MB | 26–31 MB | — | — | — |
+| Swap growth | Whether the run pushed the host into swap | Swap used at block start versus peak | 0 | 0 | 0 | 0 | 0 | 0 |
+| Layer CPU | CPU the management process uses per task | `ps` CPU time of FreeLlama's server over 100 warm tasks | — | 9.7 ms | — | — | — | — |
+| GPU placement | Whether the model ran fully on the GPU | Ollama `/api/ps` after each answer, from the receipt | — | 59/60 | 60/60 | — | — | — |
+
+¹ llama.cpp loads the model before it reports ready, so its first request after a restart is warm
+(0.33 s); its load time is part of process start, not the request.
+
+Bold marks the three setups within 2 ms of each other. Gaps under about 50 ms changed order
+between earlier runs on a busier host; read them as ties.
+
+## Results
+
+[Full report](assets/benchmark-results-2026-11-14.html): pooled and per-order tables,
+host samples. Between the two run orders LocalAI's median moved by 100 ms (851 → 751 ms) and
+llama.cpp's by 34 ms (311 → 345 ms); Ollama, FreeLlama and FreeLlama streamed stayed within
+268–275 ms in both.
+
+**Concurrent clients** (forward pass, 30 requests per level):
+
+| System | 2 clients: p50, tasks/min | 4 clients: p50, tasks/min |
+|---|---:|---:|
+| Ollama alone | 527 ms, 226 | 1,058 ms, 225 |
+| FreeLlama | 534 ms, 221 | 1,068 ms, 221 |
+| FreeLlama streamed | 535 ms, 220 | 1,074 ms, 222 |
+| llama.cpp | 561 ms, 210 | 777 ms, 291 |
+| llama-swap | 564 ms, 213 | 787 ms, 305 |
+| LocalAI | 913 ms, 128 | 1,621 ms, 145 |
+
+Ollama serves one request at a time by default, so its time per request doubles with two clients;
+FreeLlama keeps its throughput within 2%. llama.cpp's parallel slots overlap requests.
+
+### FreeLlama's overhead, profiled
+
+Each receipt carries `timing`: milliseconds per phase inside `serve`. On 40 warm requests the old
+build spent its time on three things every request paid for: reading Ollama's `/api/tags` for a
+pinned model (21 ms median; Ollama reads every manifest from disk), waiting for a fresh host
+sample every 2 s (55 ms at p90), and writing learned feedback to SQLite (5 ms median). The new
+build caches each endpoint's tag list for 10 s, samples the host and probes Ollama's settings in
+the background, and writes feedback after the response.
+
+| Client time minus Ollama engine time | Old build | New build |
+|---|---:|---:|
+| Median | 22.9 ms | 3.7 ms |
+| p90 | 60.3 ms | 5.6 ms |
+| Max | 81.5 ms | 29.6 ms |
+
+Both builds measured back to back on the same host at the same load (load average about 7).
+
+### Memory gate, live
+
+With the host under pressure (load average 16, swapping, 5.9 GB free of 32 GB), a FreeLlama
+`summarize-text` task was held for its 20 s wait limit and then refused: "needs 7.2 GB incl. 3.6 GB reserve, 4.2 GB
+free; largest memory users: Google Chrome Helper ×54 4.6 GB, Code Helper ×3 2.6 GB, node ×13
+2.4 GB…". Nothing was loaded and nothing swapped. The suite has no matching row for the other
+systems yet (`--hog-gb` exists but was not run).
+
+## Features
+
+![FreeLlama's management features, measured](assets/benchmark-features.png)
+
+Each feature measured against the same work without it, through the MCP server an agent uses,
+with a fresh `serve` on the build above (`scripts/bench/features.mjs`). Same Mac, 2026-11-14. Raw
+numbers: [`benchmark-features-raw-2026-11-14.json`](assets/benchmark-features-raw-2026-11-14.json).
+
+| Feature | Measured against | Without it | With FreeLlama | How |
+|---|---|---:|---:|---|
+| Priority under load | One interactive task sent after 8 background tasks | 3.52 s (Ollama alone, arrival order) | **1.40 s** | Median of 7 rounds; `priority: "interactive"` vs `"background"`. The whole batch of 9 took 3.62 s and 3.67 s (+1%) |
+| Built-in workflow prompts | The one-line prompt an agent would write | 52% of graded cases | **98%** | 24 cases × 2 samples on `bonsai-2:9b`; [per workflow](#built-in-prompts-measured) |
+| Code search index | The first search of `packages/mcp/src` by meaning | 55.9 s | **0.13 s** (repeat) | Job `find-code` with `nearest`; the repeat embeds no unchanged chunk and returns the same top file |
+| Sweep with a baseline | A first sweep: 83 chunks judged one by one by `strands-decider:2b` | 49.8 s | **18 ms** | `sweep {baseline}` reused all 83 unchanged chunks |
+| Long task as a job | Waiting on one call | — | **29 ms** to a job id | `run_task {defer:true}` for `summarize-log` over `CHANGELOG.md`; done in 19.2 s while the caller polled `task_jobs` |
+| Memory gate | Loading a model that does not fit | — | **Refused** | 7.2 GB needed with 4.2 GB free; nothing was loaded ([details](#memory-gate-live)) |
+| GPU placement | — | — | **119/120** verified | Ollama `/api/ps` after each answer showed the whole model in VRAM |
+| Layer cost | — | — | **9.7 ms CPU, 26–31 MB** per task and server | `ps` CPU time over 100 tasks; peak server RSS |
+
+The first index and sweep runs include embedding or judging every chunk; their repeats read saved
+results, so the size of the gain depends on how much changed.
+
+Not a FreeLlama-only gain, so not in the table: a second question about the same 3,144-token
+document took 5.1 s instead of 11.9 s (2,120 of 3,151 prompt tokens from cache). That is Ollama's
+prompt cache; FreeLlama keeps the context size and settings stable between calls so the cache
+stays usable, and Ollama alone gives the same when a client does the same. Ollama reuses it across
+turns only for attention models such as `bonsai-2:9b`, not hybrid ones such as `lfm3`.
+
+### Built-in prompts, measured
+
+FreeLlama's workflows send model-specific prompts evaluated on the dated cases below. To measure what that adds, the
+workflow evaluation (`scripts/local-test/workflows.mjs`) ran the 9 text workflows' graded cases on
+`bonsai-2:9b`, 2 samples each, once with the shipped prompts and once with the one-line prompt an
+agent would write itself ([`baseline-prompts.mjs`](../scripts/local-test/baseline-prompts.mjs),
+naming the answer format where the workflow needs one). Everything else (model, options, grading)
+was the same. On the 24 cases graded both ways:
+
+| | Shipped workflow prompts | One-line prompts |
+|---|---:|---:|
+| Pass among completed graded samples | **45/46 (98%)** | 25/48 (52%) |
+| Pass / nominal 48 planned trials | 45/48 (94%) | 25/48 (52%) |
+| Median time per task | **2.54 s** | 3.29 s |
+| Median output tokens | **46.5** | 72.5 |
+
+Per workflow (pass rate, median time per task):
+
+| Workflow | Built-in prompt | One-line prompt |
+|---|---:|---:|
+| `summarize-text` | 7/7, 1.1 s | 6/8, 1.5 s |
+| `summarize-log` | 9/9, 4.0 s | 2/10, 6.6 s |
+| `summarize-diff` | 4/4, 3.1 s | 4/4, 5.2 s |
+| `extract-json` | 9/10, 2.1 s | 9/10, 1.7 s |
+| `explain-code` | 2/2, 6.9 s | 0/2, 27.7 s |
+| `review-code` | 6/6, 1.6 s | 2/6, 13.2 s |
+| `fix-bug` | 4/4, 3.0 s | 0/4, 2.7 s |
+| `change-code` | 2/2, 2.7 s | 1/2, 2.8 s |
+| `write-tests` | 2/2, 13.6 s | 1/2, 28.6 s |
+
+`extract-json` is a tie: a schema-shaped task the one-line prompt already states.
+
+The one-line prompts failed by padding past length limits (a log summary of 170 words for a 90-word
+bound), inventing causes ("database" in a log that never mentions one), missing the failed step,
+and leaving out the edit block. The retained aggregates contain 46 shipped completions versus 48
+baseline completions over a nominal 48 planned trials per arm. The 98% figure is 45/46 completed
+samples; counting the two absent shipped completions as not passed gives 45/48 (93.75%). The
+individual traces for those absent samples are not in the checked-in aggregates. Earlier notes
+report six memory-held attempts, but the retained data does not reconstruct those attempts or
+explain their relationship to these two missing completions. Treat this as
+incomplete completion evidence, not a production reliability percentage. The newer
+[tool-calling smoke suite](RELIABILITY.md#public-tool-calling-smoke-suite) tests a different scope.
+
+## Scorecard
+
+FreeLlama rated on measured numbers only: each target is parity with Ollama alone (which
+FreeLlama runs on) or a fixed bar, and each value comes from this page.
+
+| KPI | Target | Measured | How | Result |
+|---|---|---|---|---|
+| Quality: benchmark task | Same as Ollama (100%) | 60/60 correct, byte-identical to Ollama | [KPIs](#kpis) | Met |
+| Quality: built-in prompts | Beat one-line prompts | 45/46 completed vs 25/48; nominal planned 45/48 vs 25/48 | [Denominators and limits](#built-in-prompts-measured) | Observed gain; incomplete shipped trials |
+| Time: median task | Within 5% of Ollama | 274 vs 272 ms (+0.7%) | [KPIs](#kpis) | Met |
+| Time: own overhead | p95 ≤ 15 ms | p95 9.8 ms, median 3.7 ms | [Overhead](#freellamas-overhead-profiled) | Met |
+| Time: first token | Within 10% of Ollama | 95 vs 93 ms (+2%) | [KPIs](#kpis) | Met |
+| Load: 4 clients | ≥ 95% of Ollama's throughput | 221 vs 225 tasks/min (98%) | [Results](#results) | Met |
+| Switching: cold request | Within 5% of Ollama | 1.21 vs 1.13 s (+7%) | [KPIs](#kpis) | **Missed** |
+| CPU | < 2% of one core while serving | 9.7 ms CPU per task, 1.6% of a core | below | Met |
+| Memory: the layer | ≤ 50 MB | 26–31 MB server RSS | [KPIs](#kpis) | Met |
+| Memory: under pressure | Refuse rather than swap | Refused 7.2 GB with 4.2 GB free; swap +0 in every block | [Memory gate](#memory-gate-live) | Met |
+| GPU placement | Every receipt verified | 119/120 verified fully on the GPU | below | **Missed** |
+
+The two misses: a cold request costs about 80 ms more through FreeLlama (admission, a host sample
+and the residency watch around the load), and one receipt could not verify placement because the
+model was no longer resident when FreeLlama looked after the answer.
+
+**CPU** was measured as the `ps` CPU time (user + system) of FreeLlama's server process over 100
+warm managed tasks, background host sampling included: 9.7 ms per task, against 10.8 ms for
+Ollama's own server process over the same tasks, with tasks 0.6 s apart. **GPU placement** is the
+receipt's `execution.observation`: Ollama's `/api/ps` after each answer showed the model's whole
+size in VRAM (`processor: gpu`) for 119 of 120 FreeLlama tasks.
+
+## Competitor landscape
+
+Checked on 2026-11-13 using GitHub repository metadata (displayed stars and last push). Stars help
+choose which tools to look at. They do not measure quality or speed.
+
+**Runtimes and servers.** These load and run the model themselves.
+
+| Project | Stars | Status here |
+|---|---:|---|
+| [Ollama](https://github.com/ollama/ollama) | 186.9k | Benchmarked (baseline; FreeLlama's upstream) |
+| [llama.cpp](https://github.com/ggml-org/llama.cpp) | 134.9k | Benchmarked (`llama-server`) |
+| [vLLM](https://github.com/vllm-project/vllm) | 96.9k | Not run: no Apple Metal backend |
+| [GPT4All](https://github.com/nomic-ai/gpt4all) | 77.3k | Not run: desktop app, last push 2025-05 |
+| [LocalAI](https://github.com/mudler/LocalAI) | 50.4k | Benchmarked |
+| [exo](https://github.com/exo-explore/exo) | 48.4k | Not run: built for clusters of several machines |
+| [textgen (oobabooga)](https://github.com/oobabooga/textgen) | 47.7k | Not run: desktop app |
+| [Jan](https://github.com/janhq/jan) | 45.7k | Not run: desktop app |
+| [SGLang](https://github.com/sgl-project/sglang) | 38.8k | Not run: GPU-server engine |
+| [llamafile](https://github.com/mozilla-ai/llamafile) | 26.6k | Not run: same engine family as llama.cpp |
+| [oMLX](https://github.com/jundot/omlx) | 23.6k | Not run: needs a separate MLX artifact and a third-party tap |
+
+**Lifecycle, routing and agent layers.** These are FreeLlama's closest peers. They sit in front
+of an engine.
+
+| Project | Stars | Relation to FreeLlama | Status here |
+|---|---:|---|---|
+| [LiteLLM](https://github.com/BerriAI/litellm) | 63.2k | Gateway for many providers: budgets, retries | Not run: a different layer |
+| [llama-swap](https://github.com/mostlygeek/llama-swap) | 6.2k | Starts engines on demand and swaps models | Benchmarked |
+| [GPUStack](https://github.com/gpustack/gpustack) | 6.0k | Schedules models across GPU clusters | Not run: needs several machines |
+
+Smaller runtimes, not run: [KoboldCpp](https://github.com/LostRuins/koboldcpp) (12.3k) and
+[mlx-lm](https://github.com/ml-explore/mlx-lm) (7.7k) share an engine family or artifact format with
+those above; [Lemonade](https://github.com/lemonade-sdk/lemonade) (6.2k),
+[LM Studio's CLI](https://github.com/lmstudio-ai/lms) (5.4k), [RamaLama](https://github.com/containers/ramalama)
+(3.1k) and [Docker Model Runner](https://github.com/docker/model-runner) (0.7k) were not installed.
+Ollama-specific layers under 1k stars ([ollama-mcp-bridge](https://github.com/patruff/ollama-mcp-bridge),
+[Olla](https://github.com/thushan/olla), [ollama-mcp](https://github.com/rawveg/ollama-mcp),
+[herd](https://github.com/tom-swift-tech/herd), [ollama-queue-proxy](https://github.com/TadMSTR/ollama-queue-proxy))
+cover one piece each: MCP pass-through, load balancing, GPU routing or a priority queue.
+
+Tools at other layers (Open WebUI 158.6k, Fabric 44.6k, llm 12.8k, aichat 10.6k) are clients that
+sit above FreeLlama, not peers. [TensorZero](https://github.com/tensorzero/tensorzero) is now archived,
+and [RouteLLM](https://github.com/lm-sys/RouteLLM) was last pushed in 2024-08.
+
+**How systems were chosen.** A system was benchmarked when it serves an HTTP API on this Mac and
+can load the same GGUF bytes. llama-swap was added as the closest lifecycle manager. oMLX was left
+out by the operator: it would need a third-party Homebrew tap plus a 1.71 GB MLX copy of the model
+(`mlx-community/LFM3-VL-2B-OptiQ-4bit`), and that copy would be a separate cohort.
+
+### Features beyond speed
+
+Checked on 2026-11-14 against each project's README, docs and release notes, not by running
+them. "Partial" means the feature exists with a limit named in the cell.
+
+| System | Memory check before a load | Suggests a model that fits | MCP server for agents | OpenAI / Anthropic / Ollama API | Streaming | Prometheus metrics |
+|---|---|---|---|---|---|---|
+| FreeLlama 0.12.0 | Yes: refuses a cold load that would not fit live free memory and names what holds it | Yes: machine-local calibrated routing, then installed catalog fallback and runtime admission | Yes (13 tools), plus a CLI | Gateway: all three | Yes: managed tasks stream with their receipt (`stream: true`); gateway workflows stream through Ollama | Yes, plus a receipt per task |
+| Ollama 0.40.16 | Partial: splits layers to fit VRAM; refuses only when total RAM is short | Partial: featured models, not fitted to the machine | No (`ollama launch` configures agents) | Yes / partial / native | Yes | No |
+| llama.cpp 1.0.0 | Partial: `--fit` on by default; the router unloads to make room, then loads anyway | No | No (MCP client only) | Yes / yes / no | Yes | Yes (opt-in) |
+| LM Studio 0.4.28 | Yes: resource guardrails, `lms load --estimate-only` | Partial: estimates | No (MCP client) | Yes / yes / no | Yes | No |
+| LocalAI 4.12.1 | Yes: `--vram-budget` cap and context auto-fit | Yes: hardware-based recommendations | Unclear (client documented) | Yes / yes / yes | Yes | Yes |
+| oMLX 0.8.0rc1 | Yes: process memory cap and eviction | Yes: first-run fit helper | No (MCP client) | Yes / yes / no | Yes | No |
+| llama-swap v272 | Partial: evicts by declared pools, not measured memory | No | No | Yes / yes / no (passthrough) | Yes | Yes |
+| vLLM 0.32.0 | Not found in docs | Partial: Metal recipes by Mac memory | No | Yes / yes / no | Yes | Yes |
+| Jan 0.8.8 | Partial: fit check; auto-fit off by default since 0.8.1 | Partial | No (MCP client) | Chat completions / unknown / no | Unknown | Unknown |
+| GPT4All 3.10.0 | No (static RAM labels) | Partial: static labels | No | Partial / no / no | Unknown | No |
+
+FreeLlama inherits Ollama's engine, formats and modalities (vision, embeddings, speech) and adds
+the gate, the pick and the receipts. Memory-aware loading is no longer rare: LocalAI, oMLX, LM
+Studio and llama.cpp's router each have their own. FreeLlama's distinct parts are a gate driven
+by live host signals, [machine-local calibrated picks with catalog fallback](MODEL_SELECTION.md#turn-measurements-into-routing-evidence),
+job prompts, and an MCP server for bounded agent work. These current selection rules were not
+measured by the pinned single-model comparison above.
+
+The dated feature snapshot above lists llama.cpp 1.0.0 (2026-11-13) and Ollama 0.40.16; the measured
+run used llama.cpp build 11960 and Ollama 0.40.15. Do not transfer the results to different versions.
+
+## Setup
+
+| Item | Value |
+|---|---|
+| Host | Apple M2 Pro, 12 CPUs, 32 GiB unified memory; other applications open; 1-minute load average 9.8 at the start, 5.3 at the end |
+| Model | `lfm3-vl:2b-q4_K_M` (LFM3 1.7B, Q4_K_M), one GGUF file, SHA-256 `56dac2e192ff0d9ab8507f4f1aeb4aed0463449bbbefaac79497f041c95f21eb`, checked with `shasum -a 256` |
+| Task | Extract `id`, `team` and `priority` from a support ticket as JSON; temperature 0, seed 42, 64 output tokens, 4,096 context, thinking off |
+| Tokens per task | 75 prompt + 22 output (Ollama, FreeLlama, LocalAI); 74 + 24 (llama.cpp, llama-swap): each server applies the model's chat template its own way |
+| Ollama | 0.40.15 (`/api/chat`) |
+| FreeLlama | 0.12.0 plus the unreleased changes in the [changelog](../CHANGELOG.md): managed tasks (`/_freellama/v1/tasks`), streamed with `stream: true` |
+| llama.cpp | `llama-server` 0.8.0, build 11960 (`8642541b2`), `-ngl 99 --jinja --reasoning-budget 0` |
+| llama-swap | v272 (`802f1c4`) over the same `llama-server` |
+| LocalAI | 4.12.1, `llama-cpp-development` backend, `gpu_layers: 99`, reasoning off |
+| Procedure | One system at a time; 2 warm-ups, 30 measured requests per pass; forward and reversed order; then 2- and 4-client levels and 3 cold requests (forward only). After each Ollama-backed block the harness unloads the model it loaded |
+
+llama.cpp 1.0.0 and Ollama 0.40.16 are outside this measured version cohort.
+
+## Limits
+
+- One machine, one small model, one short task. Larger models, long prompts and long outputs can
+  order the systems differently.
+- Ollama ran with its defaults (`OLLAMA_NUM_PARALLEL` unset); with parallel slots its concurrent
+  throughput would change. Not measured.
+- Whole-host CPU includes other applications, so it is recorded in the report, not compared.
+- Process RSS undercounts GPU buffers on unified memory; only FreeLlama's own server RSS is quoted.
+- The memory-pressure comparison (`--hog-gb`) and the agent-token meter are built but not run.
+- Prompt quality was measured on one model (`bonsai-2:9b`), 2 samples per case, with deterministic
+  graders; other models gain differently. The one-line baseline is one reasonable agent prompt, not
+  the best an agent could write.
+- The feature runs are one sample of each scenario on a busy workstation (load average 10–17),
+  except priority (7 rounds). An earlier priority run gave the same interactive gain (3.80 s →
+  1.57 s) but a 17% slower batch; the batch cost varies with load.
+- CPU and the overhead before/after were measured outside the benchmark run, on the same host, with
+  the same request, minutes apart.
+- Targets in the scorecard were set by the FreeLlama team: parity with Ollama, or the RFC's bars.
+
+## Reproduce
+
+1. Install the engines: `bash scripts/bench/setup-macos.sh --install` (llama.cpp, LocalAI, oMLX).
+   Get llama-swap from its [GitHub release](https://github.com/mostlygeek/llama-swap/releases) and
+   check the archive's SHA-256. No model is downloaded automatically.
+2. Set up LocalAI from
+   [`localai-model.example.yaml`](../scripts/bench/localai-model.example.yaml). If the stable
+   `llama-cpp` backend rejects your model's architecture (4.12.1 rejected `lfm3`), use
+   `llama-cpp-development`. Keep
+   `reasoning.disable`, or LocalAI spends the 64-token cap on reasoning and returns an empty answer.
+   Set up llama-swap from [`llama-swap.example.yaml`](../scripts/bench/llama-swap.example.yaml).
+3. Add `start` commands, or start each service yourself, then run preflight, a forward pass and a
+   reversed pass:
+
+```sh
+node scripts/bench/preflight.mjs --config CONFIG.json --hash-model --out NEW_PREFLIGHT_DIR
+node scripts/bench/run.mjs --config CONFIG.json --samples 30 --out NEW_FORWARD_DIR
+node scripts/bench/run.mjs --config CONFIG.reversed.json --samples 30 --out NEW_REVERSED_DIR
+node scripts/bench/report.mjs --report NEW_FORWARD_DIR/summary.json --report NEW_REVERSED_DIR/summary.json --out NEW_REPORT.html
+```
+
+With two or more `--report` files, the page shows a pooled table (p50/p95 from each run's raw
+`trials.jsonl`) and one table per run order. Preflight reports vendors with `start` as
+`owned_start` and checks only their start command.
+
+After each Ollama-backed block, the runner unloads the benchmark model it warmed (`keep_alive: 0`
+at `ollama_base_url`, confirmed by `/api/ps`); `--keep-loaded` skips this. Optional coverage flags,
+described in the [bench README](../scripts/bench/README.md#run-options):
+
+- `--max-load-per-cpu X [--strict-load]`: record load and top CPU processes at each block start;
+  warn above X, or skip the block as `not_run`.
+- `--concurrency 2,4,8`: C1, N concurrent clients per level on the warm model.
+- `--cold [--cold-samples N]`: C3, unload or restart, then time the first request.
+- `--hog-gb N [--force-hog]` with `--cold`: C2, hold N GiB during each block. Refused above
+  estimated available minus 2 GiB unless forced.
+
+Keep the background workload fixed and note it. Before claiming a difference, compare it with the
+drift between passes. The runner stops every later vendor if an inference outcome is uncertain or
+if it cannot stop a server it started.
+
+The [features](#features) run against a fresh `serve` with its own home and the installed models
+(`--chat-model`, `--decide-model`); it pulls nothing:
+
+```sh
+node scripts/bench/features.mjs --binary target/release/freellama --out FEATURES.json
+```
+
+Local evidence (Git-ignored): `.octocode/bench/live-2026-11-14-v2/` holds the configs, both
+passes (`run-forward/`, `run-reversed/`), raw `trials.jsonl`, cold and concurrency samples, server
+logs, host samples and `background-load.txt`. The charts are drawn from it by
+`node scripts/bench/readme-charts.mjs --run DIR --run DIR --facts docs/assets/benchmark-facts-2026-11-14.json --out docs/assets`.
+
+## Verification
+
+Historical verification recorded on 2026-11-14 after the benchmark build (before the later
+calibration, engine and refactor changes): `yarn test:js` **1,280 pass** (1 skipped), `yarn
+test:bench` **112 pass**, MCP integration against live Ollama **160 pass**, and the Rust suite
+all unit tests pass (427) with **3 integration-contract failures** (293 pass), all in tests the
+earlier staged branch work did not update (digest-verified
+placement and no inference retry after a 500); none touches the changes measured here. `cargo
+clippy -D warnings` is clean. The production gate (`yarn verify:production`) was not run for that snapshot. These counts and
+failures are preserved as historical context; they are not the current checkout's verification
+status. Use [Testing](TESTING.md) for the current gate commands.

package/docs/BUILD_AND_NOTES.md

@@ -1,6 +1,8 @@
 # Build from source and operating notes
 
-Moved from the root README.
+Published npm installs need Node.js 22+ and the matching prebuilt platform package; they do not
+compile Rust. Use the [installation guide](../README.md) for that path. The commands below are for
+developing this checkout or building your own executable and addon.
 
 ## Good to know
 
@@ -13,19 +15,26 @@ Moved from the root README.
 - Without CA certificates, as in slim containers, FreeLlama still serves plain-HTTP Ollama. Install
   `ca-certificates` for HTTPS upstreams.
 - The service binds to loopback. For remote access, see [Production](PRODUCTION.md).
+- Default loopback ports: 11434 Ollama, 11435 `freellama serve` (`--listen`), 11436 the CPU-only
+  Ollama that `serve --spawn-cpu-backend` starts (`--cpu-backend-port`), 11437 `freellama gateway`
+  (`--port`).
 - Port 11435 is also Ollaya's default ([Ollaya API](https://github.com/ollaya-dev/ollaya/blob/main/docs/api.md)).
   To run both, use `serve --listen` on another port and set the MCP server's
   `FREELLAMA_SERVE_ENDPOINT` to match.
 
 ## Build from source
 
-Requires Rust 1.89+ (`rust-toolchain.toml` pins the compiler) and Node.js 22.12+ with Corepack.
+Requires Node.js 22.12+ with Corepack and the Rust compiler selected by `rust-toolchain.toml`
+(currently 1.99.0). `Cargo.toml` declares a 1.89 minimum; that floor is separate from the pinned
+development compiler and needs its own check before claiming minimum-version compatibility.
 
 ```bash
 corepack enable && yarn install && yarn build
 ./target/release/freellama serve
 node packages/mcp/dist/index.js        # MCP server from this checkout
+node packages/cli/bin/freellama.js run --help   # the agent commands (run, chat, gateway…)
 ```
 
 npm releases can lag this checkout. [Testing](TESTING.md) describes the checks; the release gate
-is `yarn verify:production`.
+is `yarn verify:production`. It includes live-system tiers and all-platform release artifacts;
+building successfully on the current host does not satisfy that complete gate.

package/docs/CLI.md

@@ -5,6 +5,9 @@ work. It exposes diagnostics, model routing, managed execution, resource control
 and Ollama-compatible passthrough. Use the [README CLI section](https://github.com/bgauryy/FreeLlama/blob/main/README.md#from-the-terminal) for first
 use and this reference to look up commands and controls.
 
+To use local models from a shell the way `ollama run` works, with job picks, workflows and files,
+use the [agent commands](AGENT_CLI.md) (`freellama run`, `chat`, `gateway`…). [Every API](API.md) puts all the surfaces on one page.
+
 ## Choose a command
 
 ```mermaid
@@ -73,6 +76,12 @@ model; see [client keys and other engines](PRODUCTION.md#client-keys). The runti
 `[preload]` table loads models at startup and predicts the next one; see
 [Preload](SCOPES_AND_WARMING.md#preload).
 
+An upstream with an explicit `[upstream.llamacpp]` configuration also supports bounded managed
+text tasks through an exact `--model NAME/MODEL` assignment. It requires fixed context/output
+limits and a conservative memory reservation; physical placement and benchmark confidence stay
+unknown. It does not participate in automatic job selection or Ollama model lifecycle controls.
+See [managed llama.cpp](LLAMACPP.md) for setup and supported fields.
+
 `serve` exposes both managed routes under `/_freellama/v1/*` and byte-preserving Ollama routes under
 `/api/*` and `/v1/*`. Use `proxy` when you need only the latter. See
 [Architecture](ARCHITECTURE.md) for the request flow.
@@ -93,10 +102,12 @@ Or let `serve` start and own the CPU process (stopped when `serve` exits; a port
 is refused):
 
 ```bash
-npx @octocodeai/freellama serve --spawn-cpu-backend --cpu-model nomic-embed-text:latest \
-  [--cpu-backend-port 11436] [--ollama-binary /path/to/ollama]
+npx @octocodeai/freellama serve --spawn-cpu-backend \
+  --cpu-model nomic-embed-text:latest --cpu-backend-port 11436
 ```
 
+Use `--ollama-binary /absolute/path/to/ollama` if the intended executable is not `ollama` on `PATH`.
+
 FreeLlama rejects a CPU upstream with no assignments, `--cpu-model` without `--cpu-upstream` or
 `--spawn-cpu-backend`, both of those together, and a CPU upstream that is the same Ollama endpoint as
 `--upstream`, including loopback aliases on one port.
@@ -121,7 +132,7 @@ runtime-file value it defaults to 2 units per `OLLAMA_NUM_PARALLEL` slot of
 the Ollama server (read from its process environment where visible; Ollama's own default is 1, so 2).
 `freellama config` shows the effective value and where it came from.
 `--cpu-max-concurrent-tasks` controls the independent CPU pool and defaults to 1. Default base costs
-are embedding 1, chat 2, and vision 4. Embeddings multiply their base by `ceil(input_items/4)`;
+are embedding and decision 1, chat 2, and vision 4. Embeddings multiply their base by `ceil(input_items/4)`;
 the acquired charge is capped to the selected backend's pool. Configure overrides in `[task_costs]`;
 see [monitoring controls](MONITORING.md#adaptive-concurrency). A saturated GPU pool does
 not consume CPU permits. `--max-queue-wait-seconds` defaults to 120; when no permit becomes
@@ -129,6 +140,8 @@ available, FreeLlama refuses the task with HTTP 503 (`admission_timeout`) instea
 A task can ask for a shorter wait with `max_wait_seconds` (`maxWaitSeconds` in MCP); it can never
 exceed the configured one. Omit it to use the server's waiting budget.
 That deadline covers weighted admission, host-resource waiting, and model-transition locking together.
+`FREELLAMA_MAX_REQUEST_BYTES` also caps each managed Ollama response at 64 MiB by default, including
+embedding, decision, error and NDJSON wire bytes; an oversized response fails with nonretryable 502.
 `--max-queued-tasks` (default 16) and `--cpu-max-queued-tasks` (default 8) also bound how many
 parsed requests can be retained per backend; a full queue receives 429 (`admission_queue_full`)
 immediately, and cancelled clients release their waiter. Every capacity refusal carries a
@@ -160,15 +173,15 @@ unknown measurements, and reserved memory. These flags do not change Ollama or O
 When effective available RAM does not cover already-reserved bytes, the task's estimated bytes and the
 reserve, FreeLlama first unloads idle unpinned runners, then holds the new load with
 `held_insufficient_capacity` until capacity returns, refusing with 503 at its deadline. This is the
-default on every platform: Linux and Windows available RAM already counts reclaimable cache (and is
-the hard cap in a container), and on macOS effective available RAM is free memory plus a bounded
-reclaimable estimate. A load already resident is never held by it. `--low-memory advisory` (or
+default on every platform: Linux uses kernel `MemAvailable` capped by a smaller cgroup limit,
+Windows uses CIM `FreePhysicalMemory`, and macOS adds a bounded reclaimable estimate to free RAM.
+A load already resident is never held by this capacity check. `--low-memory advisory` (or
 `FREELLAMA_LOW_MEMORY=advisory`) instead dispatches such a load: the receipt reports
 `status: "admitted_low_memory_estimate"` and `assessment.memory_advisory: "below_reserve"` with
 `required_with_reserve_bytes` and `effective_available_bytes`, and Ollama and the OS decide whether
 it fits.
 OS memory pressure holds new loads where the OS reports it (macOS, and Linux when the kernel
-exposes `/proc/pressure/memory`; Windows has no host collector yet), and holds already-resident
+exposes `/proc/pressure/memory`; Windows does not expose this signal through its collector), and holds already-resident
 runners too only at the critical level: a warning is what a large load itself raises. Active
 swapping and low RAM hold new loads only: an already-resident runner adds no memory and stays
 admitted, for raw passthrough requests too when the request's `model` is in Ollama's `/api/ps`. The CPU-load hold (one-minute load at
@@ -181,15 +194,17 @@ request on a runner already resident on the GPU proceeds. The receipt's `resourc
 lists the holds that applied to the request.
 Both `serve` and `proxy` accept `--resource-telemetry-policy`:
 
-- `require-memory` (default on Linux and macOS): wait or refuse local inference when available RAM is
+- `require-memory` (default on Linux, macOS and Windows): wait or refuse local inference when available RAM is
   unknown, including raw requests with zero forecast bytes. Missing thermal data alone does not block
   this policy.
 - `require-all`: require all signals applicable to the collector. The built-in Linux collector reads
   OS memory pressure from `/proc/pressure/memory` when the kernel provides it, but has no thermal
   signal and marks it as unsupported; custom collectors must provide it. Unsupported signals are not
   evidence of healthy hardware.
-- `best-effort`: explicitly permit unknown telemetry unless an observed pressure hold is active. This
-  is the default on Windows, which has no built-in memory collector.
+- `best-effort`: explicitly permit unknown telemetry unless an observed pressure hold is active.
+  Unsupported platforms without a collector use this default. Windows instead requires its bounded
+  PowerShell/CIM memory observation; its CPU-load, OS-pressure, thermal, swap and GPU signals remain
+  unknown, so `require-all` holds work there. See [Windows memory telemetry](MONITORING.md#read-the-live-status).
 
 Remote upstreams bypass local host policy. Raw metadata and empty unload requests remain available
 during a hold. Route receipts separate `queue_readiness` from `resource_readiness`, and include
@@ -291,6 +306,10 @@ Useful task options include:
 - `--stop-on-repetition` to stream the task and stop once the output repeats itself, keeping all
   generated text.
 - `--defer` to return a job ID immediately.
+- repeatable `--fallback-model TAG` for an ordered fallback chain (up to 4): each model runs only
+  after the one before it was refused before anything executed (no room by the wait deadline, input
+  over its context window, not installed or ineligible, circuit open), never after a model started.
+  The receipt's `fallback` lists every attempt ([fallback chains](FEATURES.md#fallback-chains)).
 
 Inspect deferred work with `jobs`, retrieve a result with `jobs --id <uuid>`, and cancel one task
 with `jobs --id <uuid> --cancel`. Remove one task with `jobs --id <uuid> --remove`: active work is
@@ -300,10 +319,13 @@ through MCP `task_jobs`.
 Cancellation and removal wait for local execution permits to be released; they do not unload a shared model
 or confirm that the physical runner has stopped.
 
-Jobs stay in server memory: at most 64 receipts, 16 MiB per input (64 MiB across queued and running
+By default jobs stay in server memory: at most 64 receipts, 16 MiB per input (64 MiB across queued and running
 jobs), and 2 MiB per retained result or error. Finished receipts expire after 10 minutes and can be evicted earlier to make room for new
 work. A restart discards all jobs. Admission queue limits still apply, and deferred work can fail
 after acceptance. See [monitoring](MONITORING.md) for states and waiting reasons.
+Opt-in durable receipts retain terminal results and idempotency records in the file-backed store;
+they never replay prompts after restart. See [durable deferred jobs](DEFERRED_JOBS.md) before
+enabling `FREELLAMA_DURABLE_JOBS` or using an idempotency key.
 
 Create a session with `session`, then pass its identifier to related `route` or `task` calls. A
 route preview honors an existing affinity but does not create or change one. Only a successfully
@@ -345,6 +367,15 @@ npx @octocodeai/freellama bench-all --output benchmark-report.json
 npx @octocodeai/freellama serve
 ```
 
+Without `--output`, `bench-all` writes `benchmark-report.json` in the FreeLlama home so the next
+`serve` start discovers it automatically. Use repeated `--include EXACT_TAG` flags to bound the run.
+Reports qualify only their measured host/hardware, endpoint and artifact digests; old or foreign
+reports do not qualify routing. Schema-2 reports record a hashed machine identity and observed
+hardware; missing identity or remote endpoints cannot establish machine-local throughput. A primary
+endpoint report does not qualify a separately assigned CPU runner. Throughput screens do not
+establish job accuracy; that requires task-quality evidence and verification. See
+[machine calibration](MODEL_SELECTION.md#turn-measurements-into-routing-evidence).
+
 `serve` resolves each file in this order: the explicit flag, its environment variable, the working
 directory, then the FreeLlama home (`~/.octocode/freellama`, or `--home` / `FREELLAMA_HOME`). The home also holds
 `runtime.toml`, read when neither `--runtime-config` nor `FREELLAMA_RUNTIME_CONFIG` is set.
@@ -362,6 +393,10 @@ schema_version = 1
 [policies.code_repair]
 qualified_models = ["qwen3.8:27b-mlx"]   # best first; quality-qualified only
 
+# Preserve the digest recorded by the evaluation, never substitute today's tag digest.
+[policies.code_repair.model_digests]
+"qwen3.8:27b-mlx" = "sha256:REPLACE_WITH_EVALUATED_DIGEST"
+
 # Optional request defaults for one exact tag on this task: think and num_predict (>= 1).
 [policies.code_repair.model_options."qwen3.8:27b-mlx"]
 think = false
@@ -374,11 +409,12 @@ num_predict = 512
 [platform.example.toml](https://github.com/bgauryy/FreeLlama/blob/main/platform.example.toml).
 
 The MCP server autostarts `serve` with its own environment and an arbitrary working directory, so
-put a machine-wide policy in the config directory or set the variables in the MCP client's `env`.
+put a machine-wide policy in the FreeLlama home or set the variables in the MCP client's `env`.
 `serve` prints each file it picked up.
 
 `policy from-eval` reads `pass_at_1`, not the throughput produced by `bench-all`. It refuses
-expired aggregates and fewer than three trials unless `--allow-smoke` is explicit, and it skips
+expired aggregates and fewer than three trials unless `--allow-smoke` is explicit. That flag
+produces smoke-only output for inspection, which `serve` refuses as a routing policy. It skips
 models that are not installed on `--ollama-endpoint` (default `FREELLAMA_OLLAMA_ENDPOINT`, then
 `http://127.0.0.1:11434`). This prevents speed data from being mislabeled as quality evidence.
 It accepts the task kinds `route` accepts, with `-` or `_` (`code-repair` or `code_repair`), writes
@@ -386,7 +422,16 @@ to stdout unless `--out` is given, and reports a refusal as one `error: ...` lin
 
 ## Compare the CLI and MCP surfaces
 
-MCP-only operations are `delegate_research` and the online `models { view: "library" }` view.
-CLI-only operations include `serve`, `proxy`, `bench-all`, `policy from-eval`, and `auth-token`.
-
-For the MCP tool contracts, read the [MCP package reference](https://github.com/bgauryy/FreeLlama/blob/main/packages/mcp/README.md).
+| Need | Operator commands | MCP | Agent commands (`run`, `chat`, `gateway`, `tool`) |
+|---|---|---|---|
+| Run the service, the proxy, benchmarks, policies, tokens | `serve`, `proxy`, `bench-all`, `policy from-eval`, `auth-token` | — | — |
+| One task with a job's pick, prompts and files | `task` (by task kind, no jobs) | `run_task {job\|workflow, files}` | `freellama run -j\|-w … -f …` |
+| Batches, embeddings with `nearest`, decisions | `task` (one item) | `run_task {tasks}`, `nearest`, `questions` | `freellama run -j find-code`, `--ask`; batches through `freellama tool run_task` |
+| Model library search and pulls | — | `models {view:"library"}`, `ollama_pull` | `freellama tool models\|ollama_pull` (approval still applies) |
+| Research, sweeps, speech, saved responses | — | `delegate_research`, `sweep`, `audio`, `read_response` | `freellama tool NAME JSON` |
+| Ollama, OpenAI and Anthropic clients | `serve` passthrough (real tags) | — | `freellama gateway` (virtual models and real tags) |
+
+`freellama tool` reaches all 13 MCP tools with the same schemas, because both surfaces register from
+one list. For the MCP tool contracts, read the
+[MCP package reference](https://github.com/bgauryy/FreeLlama/blob/main/packages/mcp/README.md)
+and [Every API](API.md#mcp-tools).