minima-cli 0.11.0__tar.gz → 0.12.0__tar.gz

minima_cli-0.12.0/CHANGELOG.md

@@ -0,0 +1,610 @@
+# Changelog
+
+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/).
+
+## [0.12.0] - 2026-07-20
+
+The Big Plan harness: the TUI's plan workflow grows verification teeth, and the
+renderer goes inline-only. A TUI/harness release — the server wire contract is
+unchanged from 0.11.0.
+
+### Changed
+- **Inline is the only renderer** — the fullscreen mode and docked sidebar are
+  deleted. Claude-Code-style inline panels replace them: hard-wrapped markdown with
+  verbatim code fences, a shared line classifier so the render, the height
+  estimate, and the panel reader can never disagree, and a stream-reseat fix so the
+  composer no longer strands mid-screen after a tall streamed reply commits.
+- **The plan workflow is enforced, not suggested** (ground truth is on by default,
+  `MINIMA_TUI_GROUND_TRUTH=0` opts out): plan steps carry `verify` shell commands;
+  a step can't be marked done while its check fails (red→green vs a captured
+  baseline); doom-loop and step-cap guards break spirals
+  (`MINIMA_TUI_STOP_STRIKES` / `MINIMA_TUI_SPIRAL_REPEATS` / `MINIMA_TUI_STEP_CAP`);
+  confidence tiers drive the UI (🟢 glide / 🟡 flag / 🔴 stop-and-ask); gate
+  verdicts are graded into outcome labels by tier (`MINIMA_TUI_GRADED_OUTCOME=0`
+  for the binary rule) while `evidence_source="gate"` stays green-tier-only.
+- **BREAKING: verify commands require consent.** Model-authored `verify` commands
+  are shown and approved before they execute — approval is session-only and a
+  mutated command re-prompts. Headless `-p` runs **fail closed** (gates go
+  unrunnable) unless `MINIMA_TUI_ALLOW_VERIFY=1` is set.
+- **Shift+Tab cycles primary agents** — Build → accept-edits → Plan; plan mode
+  asks first on write/edit/bash and blocks todowrite/task, with a universal exit
+  gate so finalized ground truth is the way out of a plan conversation.
+- The plan council convenes in parallel and only on substantive turns (follow-up
+  plan turns ~5× faster), with a live progress line and a plan-draft view.
+- Fullscreen defaults to selection-first mouse — native select/copy works out of
+  the box.
+
+### Added
+- **Named sessions and `--resume`** (by name or id), with the context status line
+  restored on resume.
+- **Git-shadow checkpoints, `/undo`, and `/rewind`** — per-run snapshots under
+  `refs/minima/ckpt/…` with byte-identical restore, rewind markers on the SQLite
+  events spine, and a turn picker with conversation/code/both restore modes.
+- **Session usage ledger** with per-turn `{model, tokens, cost}`, a
+  table-of-contents overlay (**Ctrl+T**) with per-section cost, a ground-truth plan
+  overview (**Ctrl+G**) with per-step cost and tiers, and `/why` verification views
+  backed by the same surface.
+- A mid-run tasks footer (**Ctrl+B** toggles, hide persists across sessions).
+- A PTY invariant suite (`make tui-verify`) — 18 scripted terminal scenarios
+  asserting echo latency, zero scrollback wipes, no alt-screen/mouse-capture
+  leaks, bottom-anchored composer, render budgets, and the plan/consent flows.
+
+### Fixed
+- Prompt-echo latency, scrolling and text-selection, and input-stability issues in
+  the inline TUI.
+- The permission overlay keyed preview rows by content prefix, collapsing
+  same-prefix verify commands (duplicate React keys could drop rows on the one
+  surface that must show every shell command before approval); rows are keyed by
+  position now.
+- `todowrite` accepts an unencoded array argument instead of failing on
+  `JSON.parse`.
+- The tool-body height ruler measured at the old fullscreen interior width while
+  the inline renderer paints full-width — the 1-row over-reserve floated the
+  composer off the terminal bottom when a plan-mode notice committed.
+
+## [0.11.0] - 2026-07-16
+
+The learning-loop rework: honest labels in, fabricated ones out, and a posterior
+that actually accumulates. Verified end-to-end against Mubit v0.12.0 (locally
+built and the deployed instance) before release.
+
+### Changed
+- **Truth rule across the wire** — quality is nullable everywhere; feedback carries
+  `evidence_source` (`gate` | `judge` | `human` | `none`) as the provenance of the
+  quality signal (the legacy `judged`/`verified_in_production` flags still map);
+  unlabeled turns and infra failures (`error_cause="infra"`) are cost/latency
+  telemetry only — they never touch the success aggregate, reinforcement, or
+  calibration. The fabricated 0.9-quality default is gone.
+- **Ground truth and honest labels are the default** — the GT verification spine is
+  ON (`MINIMA_TUI_GROUND_TRUTH=0` opts out); green gate verdicts are the label
+  source for gated turns, a sampled LLM judge (`MINIMA_JUDGE_SAMPLE`, default 15%)
+  labels a slice of the rest, and judge spend stays booked to the session wallet.
+- **Accumulating evidence** — the durable (cluster, model) record is a
+  read-modify-write upsert carrying outcome counters and sample rings, so organic
+  evidence no longer caps at n≈1 and one failure can't erase history. Recall embeds
+  the same gist text the write path stores, and a budgeted keyed lookup
+  (`/v2/core/lookup`, `evidence_only` recall) gives an exact-match evidence channel
+  (pairs with Mubit v0.12.0, now live).
+- **Thompson sampling is the default selection policy** — calibrated posterior
+  sampling with a capped explore share (`minima_selection_policy=argmin` or
+  `minima_argmin_orgs` opt out). The selection zoo (collapse guard, epsilon,
+  exploration bonus, shadow UCB, evidence-mass IPW) is deleted, and
+  `GET /v1/policy-value` reports doubly-robust regret-vs-oracle.
+- **Benchmark-derived catalog priors** replace hand-tuned capability numbers;
+  cross-generation seed aliasing is deleted (no more crediting 2024 outcomes to
+  2026 model ids); `minima-seed` defaults to the synthetic pack.
+- **The pre-decision LLM reasoner is deleted** — escalation is diagnostic-only
+  (`escalation_suggested:*` warnings + decision-log reasons); the harness recovery
+  ladder owns the cascade. Config shrinks 97 → 64 settings; recommendation-store
+  I/O and calibrator refits run off the event loop; a TS-mirror contract test pins
+  the wire schemas to `packages/tui/src/minima/schemas.ts`.
+- **The harness is a reference client** — phase tags, difficulty, and
+  `chosen_effort` go on the wire; cache-boundary stickiness prefers the incumbent
+  model mid-session when the tradeoff is marginal.
+
+### Added
+- **Typed SDK feedback** — `Usage(input_tokens, output_tokens, cost_usd, latency_ms)`
+  on `feedback()` in both clients, honest autocapture framing, and a loop quickstart
+  in the docs.
+- **Harness adapters** (`pip install "minima-cli[adapters]"`) — LiteLLM custom
+  routing strategy + realized-cost feedback logger, OpenHands `RouterLLM`, and a
+  `minima-route` CLI for shell-level integrations.
+
+### Fixed
+- **Reinforcement id-space discipline** — keyed-lookup hits carry numeric core-plane
+  node ids, which are never sent as reinforcement references anymore; the durable
+  record is the primary reference so one unresolvable neighbor can't drop the whole
+  reinforcement call (pairs with Mubit's cross-run `record_outcome` resolution fix).
+- Live-suite tests updated to the new contracts (diagnostic-only escalation,
+  deterministic asserts pinned to argmin) and adapter tests skip cleanly when the
+  frameworks are absent.
+
+## [0.10.0] - 2026-07-12
+
+### Fixed
+- **Ground-truth spine hardened end-to-end** (still opt-in via `MINIMA_TUI_GROUND_TRUTH=1`):
+  - Gate verdicts now carry the `rec_id` of the turn that minted them, so grounded
+    feedback can never be poisoned by a stale gate from an earlier prompt, and a
+    blocked step's red verdict is superseded by its retry (content-first flip
+    identity). Plans close when their last flip is verified, instead of haunting
+    every later turn.
+  - Step identity survives rewording (token-set matching), a step's baseline resets
+    when its `verify` command changes (no fabricated red→green), and resuming a run
+    re-adopts its active plan.
+  - Bash writes and sub-agent writes are attributed to the run; unattributed
+    ("blind") writes cap confidence at 🟡 instead of silently passing.
+  - Check runner: verify commands run in their own process group and are killed as a
+    group on timeout/abort (no orphaned children), run under a minimal env allowlist,
+    and honor `MINIMA_TUI_CHECK_TIMEOUT`.
+  - `/plan` council: injected findings are fenced as untrusted data, the whole plan
+    turn aborts cleanly on Esc (partial research kept), and each council round is
+    budget-metered ($0.25/round soft cap).
+- **Default path** (no flag required): the SQLite migration runner is race-safe across
+  concurrent sessions and self-heals a previously wedged DB; a bash tool timeout/abort
+  kills the whole process group; a throwing after-hook no longer wedges the agent loop;
+  the `task` tool is blocked in plan mode.
+- Gate-focus modal: keyboard verdict overrides (approve / reject / steer with a note)
+  reach the correct gate, and the TUI footer/overlay no longer overflow on narrow
+  terminals.
+
+### Added
+- With `MINIMA_LLM_JUDGE=1`, judge grading spend is now booked to the session wallet —
+  visible in `/cost` (as `judge overhead`), the footer, and enforced by `--budget` —
+  while staying out of feedback's `actual_cost_usd` so the cost model Minima learns
+  from stays clean.
+
+## [0.9.0] - 2026-07-10
+
+### Added
+- **Ground-Truth verification spine (experimental, opt-in via `MINIMA_TUI_GROUND_TRUTH=1`)** —
+  the agent's plan steps can carry `verify` shell commands; the harness runs them
+  (pre-work baseline → done-gate red→green), records every verdict in the SQLite ledger,
+  and derives a confidence tier that drives the UI (🟢 glide / 🟡 flag / 🔴 stop) plus
+  grounded, deterministic-over-judge feedback to Minima. Includes `/why` (inspect why the
+  harness trusts/distrusts a step), a plan footer strip, and red-gate override capture.
+  Off by default — with the flag unset the CLI behaves exactly as 0.8.x.
+- **`/plan` planning workflow (part of the same experimental gate)** — a planner persona
+  plus a design council of sub-agents that researches, drafts, critiques, and synthesizes
+  a `GROUND_TRUTH.md` design document (`/plan start · status · finalize · cancel`).
+  Without the flag, `/plan` stays the read-only toggle it always was.
+- Permission hardening for verify commands: every LLM-authored `verify` is shown verbatim
+  in the approval overlay (with an explicit "… +N more lines" marker when truncated), and
+  a stored "always allow" on todowrite never covers a verify command the user hasn't seen.
+
+### Changed
+- Transcript rendering is memoized (windowing + `memo()` message rows, render-time ref
+  mutations moved to effects) — typing no longer re-renders the whole transcript per
+  keystroke.
+- Repo-root `CLAUDE.md` added (agent-facing repo guide).
+
+## [0.8.0] - 2026-07-07
+
+### Fixed
+- **Learning loop now takes effect** — the client sends a stable `user_id` on every
+  recommend, so Minima's memory recall actually surfaces your prior outcomes and routing
+  can move off the cold-start prior. Previously no `user_id` was sent, so server-side
+  recall was always empty and `decision_basis` never left `prior`.
+- **Esc reliably aborts** — pressing Esc now cancels during the routing/recommend phase
+  too (it was a no-op there, so the model still ran), and an aborted turn no longer leaves
+  a dangling message that made the next prompt re-answer the previous one.
+- **Web search/fetch work without an Exa key** — `web_search` and `web_fetch` fall back to
+  DuckDuckGo when `EXA_API_KEY` is unset (they previously failed at call time); with a key
+  set they still prefer Exa.
+- **Rendering** — long lines (routing warnings, reasoning, tool output) are clipped inside
+  their bordered cells instead of drawing past the border; the `escalation_suggested`
+  internal hint is no longer leaked into the info line; the question and `/tree` overlays
+  reserve their height so they can't corrupt terminal scrollback.
+
+### Changed
+- Learning-loop write failures are now surfaced — a muted `ℹ learning loop: …` line in the
+  TUI and a `feedback_error` event in `--mode json` — instead of being silently swallowed.
+
+## [0.7.2] - 2026-07-06
+
+### Added (TUI)
+- **Fullscreen renderer (new default)** — alternate screen buffer with the prompt glued
+  to the bottom row, height-accurate transcript windowing, and in-app history scrolling
+  (mouse wheel / trackpad on by default; PgUp/PgDn always work). Opt out with
+  `--no-fullscreen` or `MINIMA_TUI_INLINE=1` for the classic inline renderer with native
+  terminal scrollback.
+- **Live current-action line** — while a tool runs, the footer shows what the agent is
+  doing right now (e.g. `⚙ bash: git diff --stat`), with `(+N more)` for parallel tools.
+- **New tools**: `question` (ask the user mid-run, never permission-gated),
+  `apply_patch` (multi-file add/update/delete/move), and Exa-backed `web_search` +
+  `web_fetch` (require `EXA_API_KEY`).
+- `--thinking LEVEL` is now actually applied (was parsed but ignored).
+- Busy indicator with rotating tips; thinking states renamed to `reasoning`/`running`.
+
+### Fixed (TUI)
+- **Esc / Ctrl+C abort a running turn** (was dead code behind the busy guard); a second
+  Ctrl+C within 2.5s force-quits even if a provider stream cannot be cancelled.
+- Plan mode (read-only) blocks `apply_patch` alongside write/edit/bash.
+- Live streaming region can never outgrow the viewport: over-budget final lines are
+  hard-sliced (fixes the fullscreen garble class and inline scrollback wipes).
+- `/fork` and `/clone` no longer claim fake success — they say they're not implemented.
+- glob/grep are permission-scoped as directory READS (`read from <dir>`) instead of
+  generic `run glob` prompts.
+
+### Changed (TUI)
+- **Permission denials are reframed for the model** ("the user declined … do not retry
+  the call and do not attempt the same action through other tools") — stops the
+  sandbox-spiral where models retry or work around a deliberate user decline.
+- Memory-recall block is annotated reference-only (do NOT run tools from recall).
+
+## [0.7.1] - 2026-07-05
+
+### Changed (API behavior — note for integrators)
+- **`max_cost_per_call` is now a true hard filter.** When no model fits the budget,
+  `POST /v1/recommend` returns **422** (`"no model within max_cost_per_call budget"`)
+  instead of 200 with the cheapest over-budget model plus a `no_model_within_cost_budget`
+  warning. This matches the documented "hard filter" contract and the existing
+  no-candidates behavior. Callers who want "cheapest possible regardless of ceiling"
+  should use the cost-quality slider, not `max_cost_per_call`.
+- **Auth rejects malformed bearer tokens.** A bearer token that is not in the Mubit key
+  format (`mbt_…`) now returns **401** before any work, instead of being accepted and
+  served from priors. A missing bearer with a server-configured `MUBIT_API_KEY` fallback
+  is unaffected.
+
+### Fixed
+- **Cold-start catalog prices were stale.** The vendored fallback snapshot
+  (`capability_priors.json`) is refreshed from the live catalog (e.g. `claude-opus-4-8`
+  `15/75` → `5/25`), and a scheduled workflow now keeps it current so it stops drifting
+  by hand. (The auto-refresh updates prices for the curated model list; it does not add
+  new models — expanded provider coverage such as the gpt-5 family remains a separate
+  curation.)
+- **Server version string can no longer drift.** `minima.__version__` (reported by
+  `/v1/health` and `/v1/capabilities`) now derives from installed package metadata instead
+  of a hardcoded constant that had silently lagged across releases.
+- **Feedback reconciliation types the quality value honestly** as `float | None` — unjudged
+  rows keep `NULL`, no fabricated default (preserves the M-J2 fix).
+
+### Internal
+- **PR CI added** — every PR now runs server (`ruff` + `pytest` + `mypy`) and TUI
+  (`bun test` + `tsc`) checks; the recommender is mypy-clean.
+
+## [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` is now published to PyPI automatically on every release** — `pip install minima-cli`
+  is the official install (it bundles the `minima_client` SDK). A new CI job builds + uploads the
+  sdist/wheel alongside the GitHub release and prod deploy.
+
+### Fixed
+- **Docs (API reference):** corrected the `GET /v1/health` response example (it returns `mubit`
+  not `memory`, `version 0.4.9`, plus `auth`/`reasoner` blocks and `catalog.cost_source`) and
+  clarified `status` is only `degraded` on a key-bearing probe; documented the real
+  `summary.realized` field set (it differs from `summary.estimated`); and fixed the `days`
+  parameter bound (`>0–365`) on `/v1/savings` and `/v1/calibration`. Added a PyPI install link to
+  the Client SDK page.
+
+## [0.4.9] - 2026-06-26
+
+### Fixed
+- **Multi-turn conversations with thinking enabled no longer 400 on Anthropic.** With extended
+  thinking on, Anthropic signs each thinking block and requires the signature echoed back when the
+  block is replayed — so the second turn of any thinking conversation (and any thinking + tool-use
+  turn) failed with `messages.N.content.0.thinking.signature: Field required`. The provider now
+  captures the `signature_delta` onto the thinking block and sends it back; an unsigned thinking
+  block (from another provider or an older session) is dropped rather than sent unsigned.
+- **Text selection works again in macOS Terminal.app.** Terminal.app doesn't report mouse-motion
+  events (xterm mode 1003), which Textual needs for in-app drag-select — so capturing the mouse
+  there gave neither in-app selection nor Terminal.app's native selection (only wheel-scroll). The
+  mouse default is now resolved per-terminal: ON everywhere (scroll + in-app drag-select, as in
+  iTerm2/Ghostty/WezTerm), but OFF on macOS Terminal.app so native click-drag selection + copy work
+  out of the box (scroll with PageUp/PageDown). `--mouse`/`--no-mouse` overrides; `/mouse` toggles.
+
+## [0.4.8] - 2026-06-26
+
+### Fixed
+- **A provider whose API key is invalid no longer wastes every turn routed to it.** When a model
+  call hard-fails on auth (e.g. an invalid `ANTHROPIC_API_KEY` → `401 invalid x-api-key`), that
+  provider is now blacklisted for the session and the *same* message is auto-rerouted onto a
+  provider whose key works — instead of the router re-recommending the dead provider on every
+  turn. The auth failure is also no longer fed back to Minima as a model-quality failure (it's a
+  credential problem, not a quality signal), so it can't poison the model's success estimate in
+  your namespace. Routing now also drops providers with no key configured up front, `/reconnect`
+  (and a key fixed via `/config`) clears the blacklist, and pins are never auto-rerouted.
+- **Scroll-wheel and text selection/copy both work now.** Terminal mouse-tracking is
+  all-or-nothing — capturing the mouse for scroll-wheel suppresses the terminal's native
+  click-drag selection. Mouse capture is back ON by default (wheel scroll + in-app drag-select),
+  the terminal's native selection stays reachable by holding the bypass modifier while dragging
+  (Option on macOS, Shift on Linux), and copy now also pushes to the OS clipboard
+  (`pbcopy`/`xclip`/`wl-copy`) — Textual's built-in copy emits only OSC 52, which macOS
+  Terminal.app silently ignores, so selections *looked* copied but weren't.
+
+### Added
+- **`/resume` picker shows timestamps** — each session row now shows when it was created and last
+  used (e.g. `used 2h ago · created 3d ago`), and the list is sorted most-recently-used first.
+- **`/mouse [on|off]`** command to toggle mouse capture live (scroll-wheel vs. terminal-native
+  selection) without restarting, plus a **`--no-mouse`** launch flag and an OS-aware selection
+  hint on the splash.
+
+## [0.4.7] - 2026-06-26
+
+### Fixed
+- **Gemini calls failed whenever a tool with a nested-model schema was attached** — including
+  the `/ledger` `tasks` tool (its `TaskItem` list). Pydantic emits `$ref`/`$defs` for nested
+  models, and the google-genai SDK's strict `Schema` model rejects those with a
+  `ValidationError` (`extra_forbidden` on `$ref`), failing the entire call. Because the error
+  text contains `extra_forbidden`, it was *misclassified* as a `403` "Access denied (key lacks
+  permission, or no quota)" — so it looked like a key/quota problem when it was a client-side
+  schema issue. (This is why Gemini "stopped working" once a ledger goal was active; introduced
+  with the `tasks` tool in 0.4.4.) The Google provider now sends tool schemas via
+  `parameters_json_schema` (the SDK's standard-JSON-Schema path, which inlines/converts `$ref`
+  itself per Gemini's function-declaration rules) instead of the strict `parameters` model.
+- **Client-side validation errors are no longer misread as provider auth failures.**
+  `classify_provider_error` now detects a pydantic/schema `ValidationError` first and reports it
+  as a tool-schema problem ("pin another model / report it"), so a `extra_forbidden` can never
+  again masquerade as a `403`/permission denial.
+
+## [0.4.6] - 2026-06-26
+
+### Added
+- **Raw provider errors are now surfaced and logged.** Alongside the clean classified message,
+  a failed model call now shows the provider's exact words (`└ provider said: …`) in the TUI and
+  logs them at WARNING, so an ambiguous `403/429` ("key lacks permission, or no quota") is
+  self-diagnosing — you can see whether it's `RESOURCE_EXHAUSTED`, `PERMISSION_DENIED`, a project
+  issue, or model availability, instead of guessing.
+
+### Fixed
+- **One provider hiccup wedged the entire session.** A failed model call (bad key, 403/429,
+  network) is swallowed into an *empty* assistant message with `stop_reason="error"`, which the
+  agent loop appended to history. On the *next* turn — even to a healthy provider — that empty
+  text block made the request invalid (Anthropic `400 "messages: text content blocks must be
+  non-empty"`), so a single hiccup broke every subsequent turn until the session was cleared.
+  Now (1) the loop never sends a failed call's assistant to a provider, and (2) a failed turn is
+  rolled fully out of the agent's context (assistant + the user message that triggered it), so
+  the next turn starts clean. Regression introduced in 0.4.4 (when provider errors began being
+  swallowed into an empty assistant rather than raised). Verified against the live Anthropic API.
+- **A failed model call was framed as "routing offline … /reconnect to retry Minima."** When
+  routing *succeeds* but the model *call* fails, the banner now reads e.g. `⚠ Access denied by
+  Google Gemini … — check GEMINI_API_KEY (/config) or pin another model (/model)` instead of the
+  misleading routing/reconnect framing. The provider-403 message also gained an actionable next
+  step (it was the only `classify_provider_error` branch without one).
+- **Switching models left a stale error banner up.** After a model's call failed, pinning or
+  unpinning a different model (`/model …`, `/model auto`) now clears the banner — a prior
+  model's "access denied"/offline message no longer lingers as if it were still happening.
+- **Pinning a model not in Minima's routing catalog 422'd and ran the wrong model.** Pinning
+  e.g. an OpenRouter-namespaced model (`google/gemini-2.5-flash`) sent it to Minima as a routing
+  constraint; Minima didn't recognize the id → `422 no models match the supplied constraints` →
+  routing degraded offline and ran a *different* fallback model, while the footer/banner
+  disagreed with what actually ran. A pin is now a true override: it bypasses Minima entirely
+  and runs exactly the pinned model (basis `pinned`), so any registered model — OpenRouter,
+  local, custom — can be pinned and runs as-is.
+
+## [0.4.5] - 2026-06-26
+
+### Fixed
+- **Routing 401'd for the whole session when the Mubit key wasn't resolvable at launch.**
+  The `AsyncMinimaClient`'s `Authorization` header is fixed when the client is built, so a
+  Mubit key added via the `/config` overlay (or exported after launch) never took effect —
+  `/reconnect` only cleared the banner without rebuilding the client, leaving every turn
+  routing offline with `minima error 401: pass your Mubit API key …` until a full restart.
+  Now `/reconnect` (and saving a routing key/URL in `/config`) re-reads the environment and
+  rebuilds the Minima client in place, so the fix applies immediately — no restart.
+- **Offline-fallback banner for an auth/config problem misleadingly said "/reconnect to
+  retry."** A no-key or rejected-key 401/403 is now classified separately from a transient
+  outage: the banner shows the actionable step ("no Mubit API key — add MUBIT_API_KEY via
+  /config") and drops the "/reconnect" framing (retrying alone wouldn't help). Transient
+  causes (timeout/unreachable) keep the "/reconnect to retry" banner.
+- **No-key + hosted Minima made a guaranteed-401 round-trip every turn.** With no key
+  configured against a remote endpoint, routing now short-circuits instantly instead of
+  waiting on a doomed request (local/loopback endpoints still attempt, so keyless local
+  servers are unaffected).
+
+## [0.4.4] - 2026-06-25
+
+### Added
+- **OpenRouter is now a full provider, not 4 hardcoded models.** Setting `OPENROUTER_API_KEY`
+  fetches OpenRouter's entire live model list (`GET /api/v1/models`, ~340 models) with live
+  pricing / context / modalities / reasoning, so any OpenRouter model is callable, pinnable, and
+  routable. Cached to `~/.minima-harness/cache` with a 24h TTL; degrades live → stale cache →
+  curated so startup never blocks or breaks offline.
+- **`/ledger` — cost-aware goals.** Set a budgeted objective (`/ledger set <title>`,
+  `/ledger budget <usd>`); the agent maintains a task checklist (the `tasks` tool, footer `N/M`,
+  re-anchored into the prompt each turn and persisted across `--continue`/`--resume`). The goal
+  conditions routing (its turns cluster in Minima's memory) and each turn's realized cost is
+  attributed to it — `└ ledger · spent $X · ~$Y projected · budget $B` — the cost-to-goal view no
+  other agent has. (`/goals` remains as a hidden alias.)
+- **Permission prompts before sensitive ops (default on).** write / edit / bash now ask first
+  (Enter approve · `a` always-allow this tool · Esc reject), previewing a diff or the command.
+  `/yolo` or `--dangerously-skip-permissions` disables prompting; `/edits` forces a diff review.
+- **`/thoughts`** streams the model's reasoning into a muted bubble above each answer; **`/exit`**
+  (and `/quit`) quit the TUI.
+
+### Fixed
+- **Provider failures are no longer silent.** A failed model call (bad/missing key, 401/403/404/
+  429/402, network) was swallowed into an *empty* assistant message — a blank bubble in the TUI,
+  an empty line on `--print`. The harness now classifies it and surfaces an actionable,
+  provider-aware message (e.g. "Authentication failed for Anthropic running claude-opus-4-8 — set
+  ANTHROPIC_API_KEY (/config)") in the TUI, on `--print` stderr (exit 1), and in `--mode json`.
+  Tool failures (incl. permission denials) render prominently instead of a faint line.
+- **OpenAI GPT-5 / o-series models 400'd on every call.** They reject `max_tokens` and require
+  `max_completion_tokens`; the OpenAI-compatible provider now sends the right param for the
+  `openai` provider (other OpenAI-compatible hosts keep `max_tokens`). Encoded as a small
+  per-provider request-quirks table rather than a hardcoded branch.
+- **`/confirm` could silently ignore your pick** (kept the routed model when a pick didn't
+  resolve) — now warns; the decision card marks candidates with no provider key as `⚠ no key`.
+- **`/model` had no way to unpin** — added an "auto (unpin)" entry + `/model auto` that restore
+  the full routing pool.
+- **Scary red banners for benign routing diagnostics** (`neighbor_classified`, `recall_timeout`,
+  `cold_start`, …) — these are now suppressed; the banner is reserved for actionable issues.
+- **Tool calls dumped raw JSON args** — now rendered IDE-style (diffs for edit/write, `$ cmd` for
+  bash, a clean summary otherwise) with colorized diffs.
+- **Errored turns were sometimes reported to Minima as successes** (when judging was off),
+  poisoning the routing loop — now recorded as failures.
+- **The launch splash was pinned to the left** instead of centered.
+- **`/v1/models` price overlay** — the harness now overlays Minima's authoritative live pricing
+  onto the registry at startup, so reported cost matches what the server routed against.
+
+### Performance
+- **`brew install minima` drops from ~5 min to ~3 s.** The Homebrew formula now installs
+  dependencies from prebuilt wheels instead of compiling grpcio / cryptography / pydantic-core /
+  jiter / cffi from source. (Apple Silicon + Linux compile nothing; macOS-Intel still builds only
+  `cryptography`, which publishes no x86_64 wheel.)
+
+## [0.4.3] - 2026-06-24
+
+### Fixed
+- **High CPU / fans during use.** The status bar repainted on *every* streamed token —
+  `_append_stream` called `set_state("working")` per delta and `StatusBar.set_state`
+  re-rendered unconditionally, so a 600-token reply triggered ~616 footer repaints (the
+  terminal emulator repaints on each, which spins fans). `set_state` is now idempotent
+  (no-op when the state is unchanged), the live-stream flush eased from ~33 Hz to ~16 Hz,
+  and the spinner timer is **paused while idle** (no 10 Hz wake-ups when nothing is running).
+  Measured: a 600-token stream drops from ~666 to ~40 repaints (~94% fewer); idle is quiet.
+  Memory is unaffected (steady ~70 MB RSS, no leak).
+
+## [0.4.2] - 2026-06-24
+
+### Added
+- **Multi-provider support (open & closed source).** A new provider catalog
+  (`ai/provider_catalog.py`) integrates 21 LLM providers — closed-native (OpenAI, Anthropic,
+  Gemini, DeepSeek, Mistral, xAI, Cohere, Perplexity), the OpenRouter aggregator, open-weight
+  hosts (Groq, Together, Fireworks, DeepInfra, Cerebras, Hyperbolic, Novita), and local
+  runtimes (Ollama, vLLM, LM Studio, llama.cpp, LocalAI). All speak the OpenAI
+  chat-completions protocol, so a verified `base_url` + the right API-key env var is enough.
+  Model ids + pricing were verified against each provider's official docs (June 2026).
+- **Key-gated, provider-specific routing.** Each model resolves *its own* provider's key
+  (a Groq model uses `GROQ_API_KEY`, never an OpenRouter key on `api.openai.com`); a provider's
+  models are registered only when its key is configured (so the `/model` picker stays relevant);
+  routing candidates and the offline fallback are filtered to models the user can actually run.
+  The `/model` picker now lists every registered model so any provider's model can be pinned.
+- **`minima config`** now lists the popular providers (Anthropic, OpenAI, Gemini, xAI, DeepSeek,
+  Mistral, OpenRouter, Groq, Together); more providers work by exporting their env var, and
+  local runtimes need no key.
+- **Config overlay UX:** Enter walks the fields and lands on a visible **Save** button (Enter
+  saves — Ctrl+S still works); the save hint is pinned in an always-visible footer.
+
+### Fixed
+- **`.env.example` shipped `MUBIT_ENDPOINT=http://127.0.0.1:3000`** and the docs said
+  `cp .env.example .env`; the CLI auto-loads `./.env`, silently degrading Mubit memory to a
+  dead localhost. The localhost default is now commented out, and `init_mubit` treats an empty
+  endpoint as unset (hosted default applies).
+- **OpenRouter-only / single-provider setups mis-routed.** The earlier key-aware fallback could
+  pick `gpt-4o-mini` (which hits `api.openai.com`) for an OpenRouter key → guaranteed 401. Key
+  eligibility is now provider-specific and base_url-aware; an unpriced (cost-0) custom/local
+  model is no longer mistaken for the cheapest offline fallback.
+- **`--offline` no longer dumps an httpx traceback** — routing fails fast with a clear
+  "routing disabled (offline mode)" reason, and the expected offline-fallback log drops the
+  stack trace (kept at DEBUG).
+
+## [0.4.1] - 2026-06-24
+
+### Fixed
+- **Published CLI defaulted routing to `localhost:8080`.** A freshly installed `minima` (no
+  project `.env.harness`) connected to a dev URL that isn't running, so every turn fell back to
+  OFFLINE with "Minima unreachable" — while `minima config doctor` misleadingly reported the
+  hosted endpoint. `DEFAULT_MINIMA_URL` is now `https://api.minima.sh` and is the single source
+  of truth shared by the runtime, the config store, and `config doctor` (they can no longer
+  drift). Local dev against `make run` sets `MINIMA_URL=http://localhost:8080` explicitly.
+- **Offline fallback could pick an unrunnable model.** The degraded-mode fallback chose the
+  globally cheapest model (gpt-4o-mini) regardless of configured keys, so an
+  Anthropic+Gemini-only setup hit a provider-auth error offline. It now prefers the cheapest
+  model whose provider key is actually set (e.g. `claude-haiku-4-5` / `gemini-2.5-flash`),
+  falling back to the global cheapest only when no key is present.
+
+## [0.4.0] - 2026-06-24
+
+First public, source-available release. Headline theme: the **harness** becomes a
+trustworthy, transparent cost-aware coding agent, and the **recommender** gains a
+data-grounded cost range.
+
+### Added
+- **`minima-harness config`** — per-user credential management across three surfaces
+  (CLI subcommand, `/config` TUI overlay, in-TUI editing). Secrets go to the **OS keyring**
+  when available, falling back to `~/.minima-harness/config.env` at mode `0600`; loaded into
+  the environment at lowest precedence. Sections: LLM provider keys + Mubit/Minima routing.
+- **`/prompt` layered inspector** — every system-prompt layer (base, project context, session
+  override, Mubit lessons) shown separately with per-layer token counts, editable in place.
+- **`/optimize`** — Mubit-backed system-prompt optimization (consolidates lessons + outcomes,
+  estimates token savings) with a local dedup fallback; never auto-applies.
+- **Routing decision card** — each candidate framed as **cost (with predicted range) / speed /
+  predictability**, an ROI line for pricier alternatives, and hybrid reasoning (data-grounded
+  by default, the reasoner's natural language only when evidence is thin).
+- **Data-grounded cost band** (server) — the recommend response now carries a p25–p75 cost band
+  (`est_cost_low` / `est_cost_high` / `cost_band_basis`) and `success_interval_width`, computed
+  from realized-cost history; honest "no range yet" when evidence is thin.
+- **Cost predictability in `/stats`** — estimate-vs-actual MAPE and within-band hit-rate
+  (estimated cost is now persisted per turn to the session log).
+- **MINIMA CLI welcome banner** and a centered launch splash.
+- `docs/publishing.md` — release checklist.
+
+### Changed
+- **License is now `FSL-1.1-Apache-2.0`** (Functional Source License — source-available,
+  non-compete; each version converts to Apache-2.0 two years after publication). Previously
+  `Proprietary`.
+- **Mouse capture is OFF by default** so terminal text selection + copy (drag, then Cmd/Ctrl+C)
+  works out of the box; `--mouse` opts into scroll-wheel support (otherwise PageUp/PageDown).
+- Footer renders Ctrl shortcuts as `ctrl+l` etc. instead of the `^l` caret.
+- Overlays (config, prompt, routing, optimize, and the model/session/command/tree pickers) share
+  a consistent rounded-accent card style with border titles.
+
+### Fixed
+- Copy/paste broke when mouse capture was enabled by default — restored selection-friendly default.
+- Harness Minima client timeout raised (10s → 30s) so a cold-start `recommend` that consults the
+  reasoner no longer silently degrades to offline routing.
+
+[0.4.0]: https://github.com/mubit-ai/minima/releases/tag/v0.4.0