opencode-llmstack 0.9.0__tar.gz → 0.9.2__tar.gz

opencode_llmstack-0.9.2/CHANGELOG.md

@@ -0,0 +1,117 @@
+# Changelog
+
+All notable changes to `opencode-llmstack` are documented here.
+
+---
+
+## [0.9.2] — 2026-05-11
+
+### Fixed
+- `classify()` now counts only **user-role messages** when evaluating the
+  multi-turn floor (`n_turns`). Previously `len(messages)` counted system
+  prompts, assistant turns, and tool-result messages, causing the floor to
+  fire after just a few real exchanges and permanently blocking `code-fast`
+  routing for the rest of any session.
+- Multi-turn floor threshold raised from **6 → 10** user turns. `code-fast`
+  is now a hosted Bedrock model (Haiku 4.5) that tool-calls reliably, so
+  the old 3B-model rationale no longer applies. Sessions with fewer than 10
+  user turns will now correctly step down to `code-fast` past 32k tokens.
+- Log label corrected: `(tools floor)` → `(user-turns=N>=10 floor)`.
+- `__version__` corrected from `"0.1.0"` to `"0.9.2"`.
+- CI release pipeline now runs lint (`ruff`) + `pytest` across Python
+  3.11/3.12/3.13 before building the wheel. Previously the `test` job was
+  documented in comments but never implemented.
+- Added `LICENSE` (MIT) file to the repository root.
+- README routing table updated: high-fidelity ceiling corrected (8k → 12k),
+  tools-floor condition updated to reflect user-turn counting, `ROUTER_MULTI_TURN`
+  default corrected (6 → 10).
+- UPGRADING.md corrected: `llmstack install` does **not** regenerate
+  `llama-swap.yaml` — that is `llmstack restart`'s job. Three places in the
+  doc had this wrong.
+- README layout tree: repo root label corrected (`opencode/` → `llmstack/`),
+  `models.ini` moved to its correct location inside the package, `shell.py`
+  (deleted) removed, `reload.py` and `LICENSE` added.
+- `iter_downloads` reference in UPGRADING.md corrected to `iter_download_targets`.
+- Bundled `llmstack/models.ini` header comment paths updated from legacy
+  locations to current `.llmstack/` state-dir layout.
+- `assert` statements in production code (`app.py`, `bedrock.py`) replaced
+  with explicit `RuntimeError` / `TypeError` raises so `-O` optimisation
+  does not silently swallow them.
+- `UPGRADING.md` and `LICENSE` added to the sdist via `pyproject.toml`
+  `package-data` / `tool.setuptools` config.
+- `[tool.pytest.ini_options]` added to `pyproject.toml` with `testpaths`
+  and `addopts`.
+- Python 3.14 classifier added to `pyproject.toml`.
+
+### Added
+- `classify()` end-to-end test coverage: step-down ladder (short/mid/long
+  context), multi-turn floor, plan-signal routing, ultra-trigger routing,
+  uncensored-trigger routing, plan ctx-size overflow fall-through.
+- Generator tests: `build_config()` coverage for gguf tiers, bedrock tiers,
+  `use_next`, `small_model` wiring, agent wiring, `auto` ctx derivation.
+- `X-LLMStack-Tokens` response header on every `/v1/chat/completions` and
+  `/v1/completions` response so opencode (and curl) can see the estimated
+  token count the router used to make its routing decision.
+
+---
+
+## [0.9.1] — 2026-05-11
+
+### Fixed
+- `classify()` multi-turn floor: count only `role == "user"` messages
+  (not all messages). This was the primary fix preventing `code-fast`
+  from ever being reached in long sessions.
+- Multi-turn threshold raised 6 → 10 (see 0.9.2 for full rationale).
+- Log label `(tools floor)` corrected to `(user-turns=N>=10 floor)`.
+
+---
+
+## [0.9.0] — 2026-05-08
+
+### Changed
+- Plan tiers now strip `tools` from the request body before dispatch.
+  Previously a plan-routed request carrying a `tools` array would fail
+  on Bedrock (Converse rejects tool configs on non-agent models).
+- Long-context fall-through to `code-fast` is now allowed even when
+  `tools[]` is present in the request body. The tools-presence check
+  was removed from the floor condition; only turn count matters now.
+- `plan` tier ctx-size overflow: when estimated tokens exceed the
+  planner's `ctx_size`, the request falls through to the coding ladder
+  instead of being sent to a planner whose window can't hold it.
+- `HIGH_FIDELITY_CEILING` raised to 12 000 (was 8 000).
+
+---
+
+## [0.8.0] — 2026-05-07
+
+### Changed
+- Fidelity-ceiling overhaul: each ceiling is now exactly half of the
+  corresponding tier's `ctx_size` (the "comfortable headroom" invariant).
+- `code-ultra.ctx_size` set to 24 000 (2× high ceiling of 12 000).
+- `code-smart.ctx_size` set to 64 000 (2× mid ceiling of 32 000).
+- `code-fast.ctx_size` set to 128 000 (YaRN ×4 from native 32k).
+- `HIGH_FIDELITY_CEILING` env var added; overrides the 12 000 default.
+- `MID_FIDELITY_CEILING` env var added; overrides the 32 000 default.
+
+---
+
+## [0.7.3] — 2026-05-06
+
+### Added
+- Per-tier Bedrock alternatives in `models.ini`: every tier now ships a
+  commented-out Bedrock block directly beneath its GGUF block.
+- All Bedrock tiers anchored to `eu-west-3`; `plan-uncensored` pinned to
+  `us-west-2` (Llama 405B has no EU deployment).
+- `aws_model_id_next` / `aws_region_next` support for Bedrock upgrade
+  pre-staging (mirrors gguf `hf_file_next`).
+
+### Fixed
+- `models.ini` comment cleanup: removed stale references to old model names.
+
+---
+
+## [0.7.2] — earlier
+
+### Fixed
+- Soft-fail when `llama-server` binary is missing at startup.
+- PowerShell activation hook: fixed `Invoke-Expression` quoting.

opencode_llmstack-0.9.2/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2024 llmstack contributors
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

{opencode_llmstack-0.9.0 → opencode_llmstack-0.9.2}/PKG-INFO

@@ -1,9 +1,30 @@
 Metadata-Version: 2.4
 Name: opencode-llmstack
-Version: 0.9.0
+Version: 0.9.2
 Summary: Multi-tier local LLM stack: llama-swap + FastAPI auto-router + opencode wiring.
 Author: llmstack
-License: MIT
+License: MIT License
+        
+        Copyright (c) 2024 llmstack contributors
+        
+        Permission is hereby granted, free of charge, to any person obtaining a copy
+        of this software and associated documentation files (the "Software"), to deal
+        in the Software without restriction, including without limitation the rights
+        to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+        copies of the Software, and to permit persons to whom the Software is
+        furnished to do so, subject to the following conditions:
+        
+        The above copyright notice and this permission notice shall be included in all
+        copies or substantial portions of the Software.
+        
+        THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+        IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+        FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+        AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+        LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+        OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+        SOFTWARE.
+        
 Project-URL: Homepage, https://github.com/rohitgarg19/llmstack
 Project-URL: Issues, https://github.com/rohitgarg19/llmstack/issues
 Keywords: llm,llama-cpp,llama-swap,opencode,router,local-ai
@@ -17,9 +38,11 @@ Classifier: Programming Language :: Python :: 3
 Classifier: Programming Language :: Python :: 3.11
 Classifier: Programming Language :: Python :: 3.12
 Classifier: Programming Language :: Python :: 3.13
+Classifier: Programming Language :: Python :: 3.14
 Classifier: Topic :: Software Development
 Requires-Python: >=3.11
 Description-Content-Type: text/markdown
+License-File: LICENSE
 Requires-Dist: fastapi<1.0,>=0.110
 Requires-Dist: httpx<1.0,>=0.27
 Requires-Dist: uvicorn[standard]<1.0,>=0.30
@@ -32,6 +55,7 @@ Requires-Dist: pytest>=7; extra == "dev"
 Provides-Extra: bedrock
 Requires-Dist: boto3>=1.35; extra == "bedrock"
 Requires-Dist: botocore>=1.35; extra == "bedrock"
+Dynamic: license-file
 
 # llmstack — multi-tier local LLM stack for Mac M4 Max / 64 GB
 
@@ -118,9 +142,9 @@ First match wins:
 | 1 | last user msg contains `[nofilter]`, `[uncensored]`, `[heretic]`, or starts with `uncensored:` / `nofilter:` | `plan-uncensored` | explicit opt-in |
 | 2 | `[ultra]` / `[opus]` / `ultra:` trigger AND `code-ultra` tier configured | `code-ultra` | explicit top-tier opt-in |
 | 3 | plan verbs (*design, architect, approach, trade-off, should we, explain why, …*) AND no code blocks / agent verbs / tools | `plan` | pure design discussion (orthogonal track) |
-| 4 | estimated input ≤ 8 000 tokens | `code-ultra` *(or `code-smart` if ultra unwired)* | top tier — context still being built, latency/$ are best here |
+| 4 | estimated input ≤ 12 000 tokens | `code-ultra` *(or `code-smart` if ultra unwired)* | top tier — context still being built, latency/$ are best here |
 | 5 | estimated input ≤ 32 000 tokens | `code-smart` | mid-context, local heavy coder is at its sweet spot |
-| 6 | otherwise (long context) AND (`tools[]` OR ≥ 6 turns) | `code-smart` | floor: 3B model tool-calls unreliably |
+| 6 | otherwise (long context) AND ≥ 10 user turns | `code-smart` | floor: deep agentic loop, keep the heavy model |
 | 7 | otherwise (long context) | `code-fast` | 128k YaRN window + always-resident + free |
 
 Token estimates are `chars / 4` over all message text + `prompt`. The
@@ -185,12 +209,12 @@ from any directory you previously ran `install` in.
 ## Layout
 
 ```
-opencode/                       # repo root
+llmstack/                       # repo root
 ├── pyproject.toml              # package metadata + `llmstack` console script
 ├── README.md                   # this file
 ├── UPGRADING.md                # how to swap any tier for a newer/better model
 │                                  + how to upgrade the Python toolchain itself
-├── models.ini                  # SINGLE SOURCE OF TRUTH for tiers + sampler
+├── LICENSE                     # MIT
 └── llmstack/                   # the python package (importable, installable)
     ├── __init__.py
     ├── __main__.py             # `python -m llmstack`
@@ -199,6 +223,7 @@ opencode/                       # repo root
     ├── shell_env.py            # spawn the env-prepared subshell + activate hooks
     ├── app.py                  # FastAPI auto-router (~280 lines)
     ├── tiers.py                # parse models.ini -> Tier dataclasses
+    ├── models.ini              # SINGLE SOURCE OF TRUTH for tiers + sampler (bundled template)
     ├── check_models.py         # snapshot tool (HF metadata + drift check)
     ├── AGENTS.md               # opencode agent template (shipped as package data)
     ├── generators/
@@ -213,9 +238,9 @@ opencode/                       # repo root
         ├── install_llama_swap.py
         ├── download.py
         ├── start.py
-        ├── shell.py
         ├── stop.py
         ├── restart.py
+        ├── reload.py
         ├── status.py
         ├── check.py
         └── activate.py
@@ -544,7 +569,7 @@ All knobs are env vars; defaults are picked up by `llmstack start`.
 | `ROUTER_UNCENSORED_MODEL` | `plan-uncensored` | `[nofilter]` triggers → here |
 | `ROUTER_HIGH_FIDELITY_CEILING` | `12000` | tokens; at or below this, route to top tier (ultra → smart fallback). Paired with `code-ultra.ctx_size = 24000` (2x). |
 | `ROUTER_MID_FIDELITY_CEILING` | `32000` | tokens; at or below this, route to `code-smart`; beyond, step down to `code-fast`. Paired with `code-smart.ctx_size = 64000` (2x). |
-| `ROUTER_MULTI_TURN` | `6` | turn count that floors the long-context rung at `code-smart` |
+| `ROUTER_MULTI_TURN` | `10` | user-turn count that floors the long-context rung at `code-smart` |
 | `ROUTER_HOST` / `ROUTER_PORT` | `127.0.0.1` / `10101` | listen address |
 | `LOG_LEVEL` | `info` | router log level |
 

{opencode_llmstack-0.9.0 → opencode_llmstack-0.9.2}/README.md

@@ -83,9 +83,9 @@ First match wins:
 | 1 | last user msg contains `[nofilter]`, `[uncensored]`, `[heretic]`, or starts with `uncensored:` / `nofilter:` | `plan-uncensored` | explicit opt-in |
 | 2 | `[ultra]` / `[opus]` / `ultra:` trigger AND `code-ultra` tier configured | `code-ultra` | explicit top-tier opt-in |
 | 3 | plan verbs (*design, architect, approach, trade-off, should we, explain why, …*) AND no code blocks / agent verbs / tools | `plan` | pure design discussion (orthogonal track) |
-| 4 | estimated input ≤ 8 000 tokens | `code-ultra` *(or `code-smart` if ultra unwired)* | top tier — context still being built, latency/$ are best here |
+| 4 | estimated input ≤ 12 000 tokens | `code-ultra` *(or `code-smart` if ultra unwired)* | top tier — context still being built, latency/$ are best here |
 | 5 | estimated input ≤ 32 000 tokens | `code-smart` | mid-context, local heavy coder is at its sweet spot |
-| 6 | otherwise (long context) AND (`tools[]` OR ≥ 6 turns) | `code-smart` | floor: 3B model tool-calls unreliably |
+| 6 | otherwise (long context) AND ≥ 10 user turns | `code-smart` | floor: deep agentic loop, keep the heavy model |
 | 7 | otherwise (long context) | `code-fast` | 128k YaRN window + always-resident + free |
 
 Token estimates are `chars / 4` over all message text + `prompt`. The
@@ -150,12 +150,12 @@ from any directory you previously ran `install` in.
 ## Layout
 
 ```
-opencode/                       # repo root
+llmstack/                       # repo root
 ├── pyproject.toml              # package metadata + `llmstack` console script
 ├── README.md                   # this file
 ├── UPGRADING.md                # how to swap any tier for a newer/better model
 │                                  + how to upgrade the Python toolchain itself
-├── models.ini                  # SINGLE SOURCE OF TRUTH for tiers + sampler
+├── LICENSE                     # MIT
 └── llmstack/                   # the python package (importable, installable)
     ├── __init__.py
     ├── __main__.py             # `python -m llmstack`
@@ -164,6 +164,7 @@ opencode/                       # repo root
     ├── shell_env.py            # spawn the env-prepared subshell + activate hooks
     ├── app.py                  # FastAPI auto-router (~280 lines)
     ├── tiers.py                # parse models.ini -> Tier dataclasses
+    ├── models.ini              # SINGLE SOURCE OF TRUTH for tiers + sampler (bundled template)
     ├── check_models.py         # snapshot tool (HF metadata + drift check)
     ├── AGENTS.md               # opencode agent template (shipped as package data)
     ├── generators/
@@ -178,9 +179,9 @@ opencode/                       # repo root
         ├── install_llama_swap.py
         ├── download.py
         ├── start.py
-        ├── shell.py
         ├── stop.py
         ├── restart.py
+        ├── reload.py
         ├── status.py
         ├── check.py
         └── activate.py
@@ -509,7 +510,7 @@ All knobs are env vars; defaults are picked up by `llmstack start`.
 | `ROUTER_UNCENSORED_MODEL` | `plan-uncensored` | `[nofilter]` triggers → here |
 | `ROUTER_HIGH_FIDELITY_CEILING` | `12000` | tokens; at or below this, route to top tier (ultra → smart fallback). Paired with `code-ultra.ctx_size = 24000` (2x). |
 | `ROUTER_MID_FIDELITY_CEILING` | `32000` | tokens; at or below this, route to `code-smart`; beyond, step down to `code-fast`. Paired with `code-smart.ctx_size = 64000` (2x). |
-| `ROUTER_MULTI_TURN` | `6` | turn count that floors the long-context rung at `code-smart` |
+| `ROUTER_MULTI_TURN` | `10` | user-turn count that floors the long-context rung at `code-smart` |
 | `ROUTER_HOST` / `ROUTER_PORT` | `127.0.0.1` / `10101` | listen address |
 | `LOG_LEVEL` | `info` | router log level |