Alluvia 0.3.0 — the first five minutes

The first five minutes, fixed — and MCP writes become opt-in.

• Newest sessions first — distillation now always processes your most
  recent sessions first, so the map is freshest where it matters.
• First-run cap — the very first refresh distills your ~50 most recent
  sessions and says so ("the remaining N backfill automatically on the next
  refresh"). You see labeled themes, connections, and unfinished threads in
  minutes instead of waiting out a full-history distill — and the label/
  status stages keep budget headroom on day one. Tune or disable with
  ALLUVIA_FIRST_RUN_CAP (0 turns it off). (#11-class first-run failures,
  designed out.)
• MCP is read-only by default (#12) — rate_proposal and propose_next
  (the tools that write your judgments or spend your LLM budget) now return
  a clear "disabled" message unless you opt in on your machine:
  [mcp] writes = true in config.toml or ALLUVIA_MCP_WRITES=1. Surprise
  writes are impossible, not merely detectable.
• Windows CI (#13) — the platform detection that already claimed Windows
  is now tested there.
• Docs: SECURITY.md (private vulnerability reporting), CONTRIBUTING.md,
  a ChatGPT-history import guide, and a README that says the quiet part
  loudly: the whole product works end-to-end on a free-tier key.

Install: uv tool install alluvia · PyPI · Closes #12, #13

Alluvia 0.2.2 — sediment & gold

Visual identity — sediment & gold. alluvia now looks like what it does: a
patient field survey of your own thinking, with one scarce point of gold at the
find.

• New mark & wordmark — three tributaries converge at a gold node and
  continue as one stem (it diagrams the product, and reads as a merge graph).
  Lockup, wide/square marks, favicon, and a social-preview card live in
  assets/; the README carries a dark/light-aware lockup.
• Dashboard reskin — the alluvia serve dashboard moves onto the brand
  palette (Ink/Basin/Paper/Wash/Silt/Gold), now light- and dark-aware instead
  of dark-only, with IBM Plex Mono station labels and an inlined favicon. Zero
  external requests, as before.
• Accessible by construction — every theme status carries a labelled dot
  (never colour alone), status and source palettes are contrast- and
  colour-vision-validated on both grounds, and gold is a filled mark reserved
  for the find/active state — never text on a light surface.
• Brand guide — docs/BRAND.md documents the palette, the scarcity rule, the
  ground-scoped token roles, and the voice.
• No engine, CLI, MCP, or storage behaviour changed.

Install: uv tool install alluvia · PyPI

Alluvia 0.2.1 — live resource usage

• alluvia top (#10) — live resource usage of every running alluvia
  process: sampled CPU %, RAM, cumulative disk read/write, uptime, with a
  machine context line. --watch N for a refreshing view; status shows
  the same process list.
• Network accounting at the source (#10) — per-process network bytes
  aren't visible to unprivileged userland on any OS, and alluvia's only
  traffic is its LLM calls — so the governor now counts them itself:
  calls and bytes sent/received per model, persisted with the breaker
  state and reported by top.
• New dependency: psutil (cross-platform process metrics).

Install: uv tool install alluvia · PyPI · Closes #10

Alluvia 0.2.0 — the good citizen

The good-citizen release: alluvia now runs concurrently, dies cleanly, shows
you everything it keeps, and repairs itself (#5, #6, #7, #8).

• Concurrent sessions (#5) — the store runs in WAL mode (readers never
  block a running refresh; legacy stores upgrade transparently), refresh
  takes a single-writer lock released by the OS on any process death
  ("already running (pid N)" instead of double-spending your LLM budget),
  and serve reuses an already-running dashboard or walks to a free port.
• Kill-anytime contract (#5) — any alluvia process can be killed at any
  instant with zero corruption and zero cleanup debt. SIGTERM behaves like
  Ctrl-C; interrupting a refresh prints "paused — everything done so far is
  saved" and resumes on the next run.
• alluvia status (#6) — every path alluvia touches with sizes, the
  store broken down by data class (raw = source of truth · derived =
  rebuildable · judgments = yours), and what's live right now. --json for
  scripting.
• alluvia doctor (#7) — diagnoses the whole installation and applies
  every safe repair automatically: WAL on legacy stores, schema migrations,
  pruning orphaned derived rows, config permissions, stale digest flags,
  impossible governor cooldowns. Raw data and judgments are never touched.
  --check reports without repairing (exit 1 if repairs are needed),
  --live proves your provider key with one tiny call, and
  --rebuild-derived is the confirmed recovery lever that discards derived
  data while raw sessions and your ratings survive.
• --verbose and refresh --plan (#8) — see the pipeline and governor
  think; preview exactly what a refresh would do (sessions pending, theme
  work, cooldowns in effect) without spending a single LLM call.

Install: uv tool install alluvia · PyPI · Closes #5, #6, #7, #8

Alluvia 0.1.2 — see what's happening

You can now see what alluvia is doing (#4). Long stages used to run silently and
look hung — worst of all a first refresh, which downloads the local
embedding model with no output at all.

• Live progress everywhere it was silent — refresh shows each stage
  (distilling N sessions, embedding, mapping themes, linking) with rich
  progress bars on a terminal and plain line output when piped or in CI.
  ingest shows a running session count.
• Deliberate waits are narrated — when the governor waits out a rate
  limit it says so (⏳ rate-limited: waiting 30s (model)) instead of
  freezing; the first-run embedding-model download announces itself.
• Embeddings run in batches so progress moves (and memory stays flat).
• alluvia --version.
• Issue templates (bug report asks exactly for the refresh summary and
  version; feature template fits both small asks and full RFCs) and a PR
  template with a test-first checklist.
• MCP tools, library callers, and the test suite see zero new output —
  progress is CLI-only by construction.

Install: uv tool install alluvia · PyPI · Closes #4

Alluvia 0.1.1 — ride out rate limits

Resilience release: provider rate limits no longer darken the map (#1, #2, #3).

• LLM governor — every provider call now runs behind a provider-agnostic
  governor: escalating cooldown ladder, per-model circuit breakers, and
  role→model fallthrough chains, all driven by abstract call outcomes (HTTP
  status and the standard Retry-After header are optional fast-paths, never
  dependencies). Breaker state persists in SQLite, so short CLI runs respect
  cooldowns learned by earlier runs instead of re-hammering an exhausted model.
• Fallthrough chains — on Groq's free tier each model has its own daily
  budget; distill/label/status/why calls fall through to sibling models when
  the head model hits a wall. propose never falls through — generation fails
  loud rather than silently downgrading. Override per role with
  ALLUVIA_LLM_CHAIN_<ROLE> (comma-separated).
• unfinished never goes dark — when the status classifier is
  unavailable, recurring themes get a heuristic status (recently touched →
  open, stale → dormant) instead of unknown. Heuristics are never cached;
  the classifier upgrades them on a later refresh.
• Degradation is visible — refresh prints a per-stage summary and a
  warning with the provider retry time when a stage degraded; themes /
  unfinished, the MCP tools, and the dashboard surface the same signal. An
  all-unknown map now says "the classifier hasn't completed" instead of
  pretending nothing is unfinished.
• Distillation pauses immediately (and resumably) when every model is cooling;
  fallback labels cut at word boundaries; adaptive pacing (AIMD) discovers a
  provider's sustainable request rate empirically.
• Breaking: every environment variable now uses the ALLUVIA_ prefix
  (ALLUVIA_DB, ALLUVIA_LLM_MODEL_<ROLE>, ALLUVIA_LLM_CHAIN_<ROLE>, …) — 0.1.0
  shipped them under an inconsistent prefix. If you exported variables for
  0.1.0, re-export them with the new prefix; config.toml users are
  unaffected.

Install: uv tool install alluvia · PyPI · Closes #1, #2, #3

Alluvia 0.1.0 — first release

First installable release. Everything to date: five source adapters
(Claude Code, Cursor, Windsurf*, Antigravity*, ChatGPT export), raw-first
SQLite store with a swappable vector index (sqlite-vec/numpy), the
distill→embed→cluster→label→status→link engine, four lenses
(themes / connections / unfinished / propose) with a human ratings loop,
a proactive weekly digest with mute + dismissal-learning, eight MCP tools,
multi-provider role-mapped LLM support (Groq/OpenAI/Anthropic), config.toml

• alluvia init onboarding, and cross-platform source detection.
  (*Windsurf/Antigravity ship log-and-skip: their stores are schema-less
  protobuf — see docs/DEBT.md.)

Every capability was validated against a real 400+-session corpus through
live gates; see docs/validation/.

Install: uv tool install alluvia · PyPI