minima-cli 0.5.0__tar.gz → 0.7.0__tar.gz

minima_cli-0.7.0/AGENTS.md

@@ -0,0 +1,58 @@
+# AGENTS.md
+
+Guidance for both human and AI agents working in this repo.
+
+## What this repo is
+
+- **`src/minima/`** — Minima: a recommend-only LLM cost-optimization service (FastAPI,
+  Python 3.11+). It does NOT proxy LLM calls. It tells you which model to run, you run
+  it, judge quality, and feed the outcome back via `POST /v1/feedback` so its Mubit
+  memory sharpens. See `docs/concepts.md` for the algorithm.
+- **`client_sdk/minima_client/`** — bundled sync + async Python client. Ships inside the
+  same wheel.
+- **`packages/tui/`** — the `minima` CLI/TUI (TypeScript/Bun, shipped via Homebrew:
+  `brew tap mubit-ai/minima`). It replaced the old Python harness (`src/minima_harness`,
+  removed in v0.7.0). See `packages/tui/README.md`.
+
+## The Minima loop (always respect this)
+
+```
+recommend  ->  run the model yourself  ->  judge quality  ->  feedback
+```
+
+Minima never runs the model. Any harness owns model dispatch + quality judging. Feedback
+MUST include realized `input_tokens` / `output_tokens` / `actual_cost_usd` / `latency_ms`
+so the cost basis can climb `estimate -> observed -> rescaled` (the biggest accuracy
+lever). Outcome thresholds: `success >= 0.8`, `partial >= 0.4`, else `failure`.
+
+## Conventions
+
+- `from __future__ import annotations` at the top of every module.
+- Pydantic v2 `BaseModel` for serializable schemas; `@dataclass(slots=True)` for internal
+  types; `StrEnum` for enums; `Protocol` for seams (`Memory`, `Reasoner`, `Provider`).
+- Async-first. Bridge sync SDKs (Mubit) off the event loop via `anyio`/threadpools.
+- `structlog` via `get_logger("minima.<sub>")`.
+- ruff `line-length=100`, mypy `py311`. Never break the hot path: bookkeeping failures are
+  logged-and-swallowed (`except Exception: # noqa: BLE001`).
+- NO comments unless asked. Match existing style.
+
+## Commands
+
+```bash
+uv sync --extra dev                       # install (uv is the package manager)
+uv run ruff check . && uv run ruff format .  # lint + format
+uv run mypy src/minima                     # typecheck
+uv run pytest                              # unit + integration (hermetic, offline)
+uv run pytest -m live                      # needs a running Mubit (make run-mubit)
+make run                                   # local Minima on :8080
+```
+
+Offline tests MUST stay hermetic: the `conftest.py` autouse fixture neutralizes `.env`.
+Use `tests/factories.py:FakeMemory` + `create_app(...)` + `TestClient` for in-process
+tests without Mubit.
+
+## Agent harness
+
+The Python harness (`src/minima_harness`) was removed in v0.7.0. Its successor is the
+TypeScript harness + TUI in `packages/tui` (the shipped `minima` CLI) — see
+`packages/tui/README.md` for architecture, commands, and hard rules.

{minima_cli-0.5.0 → minima_cli-0.7.0}/CHANGELOG.md

@@ -4,6 +4,66 @@ All notable changes to Minima are documented here. The format follows
 [Keep a Changelog](https://keepachangelog.com/), and the project aims to follow
 [Semantic Versioning](https://semver.org/).
 
+## [0.7.0] - 2026-07-05
+
+### Added
+- **`GET /v1/capabilities`** — server feature handshake (api version, honored constraint
+  fields, feature flags) so clients can gate enforce-mode features on what the server
+  actually supports instead of guessing. (#54)
+- **Sub-agent tree visualization** — a live `▸ step [status] $cost` panel (`/tree`) fed by
+  tagged child events during multi-agent runs. (#55)
+- **Git worktree isolation for parallel sub-agents** — `isolation: "workdir"` delegations run
+  in a temporary `git worktree` (with a dirty-tree warning and automatic cleanup), so parallel
+  children editing the same files can't clobber each other. (#55)
+- **Delegation ops-rules + BLOCKED convention** — sub-agents get explicit operational rules,
+  boundary refusals surface as `BLOCKED` → `partial` outcome so the parent can tell, and the
+  task tool carries anti-retry guidance (no more re-running an identical failed DAG). (#59)
+
+### Fixed
+- **Every Claude call timed out in the harness** — the Anthropic SDK was handed a timeout in
+  seconds where it expects milliseconds (a 30–60 ms budget), so every Claude stream/complete
+  call died with "Request timed out" and fed quality-0 failures into routing. (#56)
+- **Quality-score fabrication in the feedback loop (M-J2)** — unjudged successes no longer
+  default to quality 0.9 in the server decision log; the harness now sends a `judged` flag and
+  unjudged rows reconcile with `quality = NULL`, so calibration and savings metrics stop
+  trusting values that were never measured. (#54)
+- **Judge hardening** — empty-output guard (scores 0 without an API call), head+tail
+  truncation so tail requirements survive long outputs, `<response>` delimiters + scoped-trust
+  system prompt against response-embedded injection, and a `parseScore` fix for
+  "X out of 10" phrasing. (#57)
+- **Gemini reasoner starvation** — advisory JSON calls (classify/rank) no longer burn their
+  entire token budget on thinking before emitting JSON; RANK_SYSTEM invariants hardened. (#58)
+
+### Notes
+- `minima-cli` on PyPI resumes at 0.7.0: the project version had lagged at 0.5.0, so the
+  0.6.0 release build was silently skipped by `--skip-existing` and 0.6.0 was never published.
+
+## [0.6.0] - 2026-07-02
+
+Agent-core release (milestones M-A…M-I): cost-aware sub-agent orchestration — `task` tool
+with sequential and parallel DAG fan-out under a semaphore, per-child routing, budgets and
+timeouts; DB-backed `BudgetLedger` with graduated enforcement; SQLite persistence spine
+(runs/events/decisions + rehydration); recovery ladder walking server-supplied rungs on
+failure; effort routing Phase A + fleet metrics; feedback-loop poisoning fixes; feature-vector
+classifier with strongest-signal scoring. *(Never published to PyPI — see the 0.7.0 note.)*
+
+## [0.5.2] - 2026-07-02
+
+Fixed TUI viewport-overflow rendering corruption; docs rebrand (Manrope, "minima by Mubit"
+nav logo, new favicon).
+
+## [0.5.1] - 2026-07-01
+
+Mubit memory in the TS harness (recall-before-route + outcome write-back with stable session
+ids); benign routing diagnostics no longer render as errors; live model catalog; docs migrated
+to Vocs with a Homebrew install guide.
+
+## [0.5.0] - 2026-07-01
+
+The TypeScript/Bun TUI becomes the shipped CLI: one-click `minima auth` (browser login,
+per-repo project mapping, workspace provisioning), the Homebrew tap ships the compiled binary
+instead of a Python venv formula, and a comprehensive routing/tool test suite lands.
+
 ## [0.4.10] - 2026-06-26
 
 ### Changed

{minima_cli-0.5.0 → minima_cli-0.7.0}/Makefile

@@ -1,7 +1,7 @@
-.PHONY: install run test lint fmt live eval seed refresh-catalog harness-demo harness-live harness-test harness tui-install tui-test tui-check tui-build tui tui-dev
+.PHONY: install run test lint fmt live eval seed refresh-catalog tui-install tui-test tui-check tui-build tui tui-dev
 
 install:
-	uv sync --extra dev   # dev pulls server + harness + tui for the full test suite
+	uv sync --extra dev   # dev pulls server + reasoner extras for the full test suite
 
 run:
 	uv run --extra server uvicorn minima.main:app --reload --host $${MINIMA_HOST:-0.0.0.0} --port $${MINIMA_PORT:-8080}
@@ -26,20 +26,6 @@ eval:
 seed:
 	uv run minima-seed --limit $${LIMIT:-2000} --lane $${LANE:-minima:default}
 
-# --- minima_harness convenience targets (creds live in gitignored .env.harness) ---
-
-harness-demo:
-	uv run python examples/harness_warmup.py
-
-harness-live:
-	uv run --env-file .env.harness python examples/harness_warmup.py --live --rounds $${ROUNDS:-1}
-
-harness-test:
-	uv run --env-file .env.harness pytest tests/harness -m live -v
-
-harness:
-	uv run --extra harness --extra tui --env-file .env.harness minima-harness $(ARGS)
-
 # --- TS TUI (packages/tui) — run from the repo root so .env.harness auto-loads --------
 
 TUI := packages/tui

{minima_cli-0.5.0 → minima_cli-0.7.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: minima-cli
-Version: 0.5.0
+Version: 0.7.0
 Summary: Minima CLI: cost-aware LLM model routing — recommend cheaper models, backed by Mubit memory.
 Project-URL: Homepage, https://docs.minima.sh
 Project-URL: Documentation, https://docs.minima.sh
@@ -30,7 +30,6 @@ Provides-Extra: dev
 Requires-Dist: anthropic>=0.40; extra == 'dev'
 Requires-Dist: fastapi>=0.115; extra == 'dev'
 Requires-Dist: google-genai>=0.3; extra == 'dev'
-Requires-Dist: keyring>=24; extra == 'dev'
 Requires-Dist: mypy>=1.10; extra == 'dev'
 Requires-Dist: psycopg2-binary>=2.9; extra == 'dev'
 Requires-Dist: pytest-asyncio>=0.23; extra == 'dev'
@@ -38,11 +37,7 @@ Requires-Dist: pytest>=8.2; extra == 'dev'
 Requires-Dist: redis>=5.0; extra == 'dev'
 Requires-Dist: respx>=0.21; extra == 'dev'
 Requires-Dist: ruff>=0.5; extra == 'dev'
-Requires-Dist: textual>=3.0; extra == 'dev'
 Requires-Dist: uvicorn[standard]>=0.30; extra == 'dev'
-Provides-Extra: harness
-Requires-Dist: anthropic>=0.40; extra == 'harness'
-Requires-Dist: google-genai>=0.3; extra == 'harness'
 Provides-Extra: reasoner-anthropic
 Requires-Dist: anthropic>=0.40; extra == 'reasoner-anthropic'
 Provides-Extra: reasoner-gemini
@@ -55,9 +50,6 @@ Requires-Dist: fastapi>=0.115; extra == 'server'
 Requires-Dist: psycopg2-binary>=2.9; extra == 'server'
 Requires-Dist: redis>=5.0; extra == 'server'
 Requires-Dist: uvicorn[standard]>=0.30; extra == 'server'
-Provides-Extra: tui
-Requires-Dist: keyring>=24; extra == 'tui'
-Requires-Dist: textual>=3.0; extra == 'tui'
 Description-Content-Type: text/markdown
 
 # Minima
@@ -176,7 +168,6 @@ Sync + async clients and zero-code `autocapture`: **[Python Client SDK](docs/cli
 | [Multi-Tenancy](docs/multi-tenancy.md) | One deployment, many orgs, per-org Mubit instances. |
 | [Operations](docs/operations.md) | Deployment, health, degradation, monitoring, secrets. |
 | [Examples](docs/examples.md) | Guided tour of the runnable examples. |
-| [Agent Harness](docs/harness.md) | `minima_harness`: a Minima-routing port of PI's agent toolkit. |
 
 ## Examples
 
@@ -191,33 +182,21 @@ Runnable, progressively advanced — in **[`examples/`](examples/)**:
 | 5 | [`05_autocapture.py`](examples/05_autocapture.py) | Zero-code intake via `mubit.learn`. |
 | 6 | [`06_routed_llm_call.py`](examples/06_routed_llm_call.py) | Routing a real Claude call + feedback. |
 | 7 | [`07_multitenant_admin.py`](examples/07_multitenant_admin.py) | Provision an org, call as that tenant. |
-| 8 | [`harness_warmup.py`](examples/harness_warmup.py) | The `minima_harness` agent loop (demo mode needs no keys). |
 
-## Agent harness
+## The `minima` CLI
 
-[`minima_harness/`](src/minima_harness) is a lean Python port of
-[`@earendil-works/pi`](https://github.com/earendil-works/pi)'s agent toolkit, made
-Minima-native: an `Agent` runtime with tool calling **plus** a `MinimaAgent` that routes
-every prompt through Minima and feeds the realized tokens/cost/quality back. It is the
-"run the model yourself" half of the Minima loop, packaged.
-
-```python
-from minima_harness.minima import MinimaAgent, HarnessConfig
-
-agent = MinimaAgent(HarnessConfig.from_env())   # MINIMA_URL, candidates, judge policy
-await agent.prompt("Summarize this incident.", task_type="summarization", slider=3)
-# -> Minima picked the model, the agent ran it, judged quality, fed the outcome back
-```
-
-Try it with no keys via the in-process demo:
+The `minima` CLI/TUI — a coding agent that routes every prompt through Minima and feeds
+realized tokens/cost/quality back — is a TypeScript/Bun binary that lives in
+[`packages/tui/`](packages/tui/) and installs via Homebrew:
 
 ```bash
-uv run python examples/harness_warmup.py          # demo (in-process Minima + fake provider)
-uv run python examples/harness_warmup.py --live   # real Minima + real providers
+brew tap mubit-ai/minima
+brew install minima
 ```
 
-Full architecture, the loop mapping, and extension guide:
-**[Agent Harness](docs/harness.md)**.
+The PyPI package (`pip install minima-cli`) is **not** the CLI: it ships the Python
+client SDK (`minima_client`) and the server + operator tooling (`minima-seed`,
+`minima-calibration-report`). The old Python harness/TUI was removed in v0.7.0.
 
 ## Configuration
 
@@ -254,7 +233,7 @@ src/minima/
   tenancy/          runtime · registry · context · keys · secrets
   seeding/          routerbench · synthetic · run_seed (minima-seed CLI)
   schemas/          common · recommend · workflow · feedback · models_catalog · strategies · admin
-src/minima_harness/   ported pi-ai (ai/) + pi-agent-core (agent/) + Minima integration (minima/) — see docs/harness.md
+packages/tui/         the `minima` CLI/TUI (TypeScript/Bun, ships via Homebrew)
 client_sdk/minima_client/   client (sync+async) · autocapture · errors
 docs/               full documentation       examples/   runnable examples
 tests/              unit · integration (FakeMemory) · live (-m live) · eval (-m eval)

{minima_cli-0.5.0 → minima_cli-0.7.0}/README.md

@@ -114,7 +114,6 @@ Sync + async clients and zero-code `autocapture`: **[Python Client SDK](docs/cli
 | [Multi-Tenancy](docs/multi-tenancy.md) | One deployment, many orgs, per-org Mubit instances. |
 | [Operations](docs/operations.md) | Deployment, health, degradation, monitoring, secrets. |
 | [Examples](docs/examples.md) | Guided tour of the runnable examples. |
-| [Agent Harness](docs/harness.md) | `minima_harness`: a Minima-routing port of PI's agent toolkit. |
 
 ## Examples
 
@@ -129,33 +128,21 @@ Runnable, progressively advanced — in **[`examples/`](examples/)**:
 | 5 | [`05_autocapture.py`](examples/05_autocapture.py) | Zero-code intake via `mubit.learn`. |
 | 6 | [`06_routed_llm_call.py`](examples/06_routed_llm_call.py) | Routing a real Claude call + feedback. |
 | 7 | [`07_multitenant_admin.py`](examples/07_multitenant_admin.py) | Provision an org, call as that tenant. |
-| 8 | [`harness_warmup.py`](examples/harness_warmup.py) | The `minima_harness` agent loop (demo mode needs no keys). |
 
-## Agent harness
+## The `minima` CLI
 
-[`minima_harness/`](src/minima_harness) is a lean Python port of
-[`@earendil-works/pi`](https://github.com/earendil-works/pi)'s agent toolkit, made
-Minima-native: an `Agent` runtime with tool calling **plus** a `MinimaAgent` that routes
-every prompt through Minima and feeds the realized tokens/cost/quality back. It is the
-"run the model yourself" half of the Minima loop, packaged.
-
-```python
-from minima_harness.minima import MinimaAgent, HarnessConfig
-
-agent = MinimaAgent(HarnessConfig.from_env())   # MINIMA_URL, candidates, judge policy
-await agent.prompt("Summarize this incident.", task_type="summarization", slider=3)
-# -> Minima picked the model, the agent ran it, judged quality, fed the outcome back
-```
-
-Try it with no keys via the in-process demo:
+The `minima` CLI/TUI — a coding agent that routes every prompt through Minima and feeds
+realized tokens/cost/quality back — is a TypeScript/Bun binary that lives in
+[`packages/tui/`](packages/tui/) and installs via Homebrew:
 
 ```bash
-uv run python examples/harness_warmup.py          # demo (in-process Minima + fake provider)
-uv run python examples/harness_warmup.py --live   # real Minima + real providers
+brew tap mubit-ai/minima
+brew install minima
 ```
 
-Full architecture, the loop mapping, and extension guide:
-**[Agent Harness](docs/harness.md)**.
+The PyPI package (`pip install minima-cli`) is **not** the CLI: it ships the Python
+client SDK (`minima_client`) and the server + operator tooling (`minima-seed`,
+`minima-calibration-report`). The old Python harness/TUI was removed in v0.7.0.
 
 ## Configuration
 
@@ -192,7 +179,7 @@ src/minima/
   tenancy/          runtime · registry · context · keys · secrets
   seeding/          routerbench · synthetic · run_seed (minima-seed CLI)
   schemas/          common · recommend · workflow · feedback · models_catalog · strategies · admin
-src/minima_harness/   ported pi-ai (ai/) + pi-agent-core (agent/) + Minima integration (minima/) — see docs/harness.md
+packages/tui/         the `minima` CLI/TUI (TypeScript/Bun, ships via Homebrew)
 client_sdk/minima_client/   client (sync+async) · autocapture · errors
 docs/               full documentation       examples/   runnable examples
 tests/              unit · integration (FakeMemory) · live (-m live) · eval (-m eval)