synartesis 0.6.2 → 0.6.11

package/CHANGELOG.md

@@ -2,6 +2,225 @@
 
 What changed, and why it mattered. Dates are release dates.
 
+## 0.6.11 — 2026-09-13
+
+### Fixed
+
+- **A conversation with Gemini died on the round after its first tool call.**
+  Gemini 3 signs the reasoning behind every function call it makes, and
+  refuses the next request outright if that signature does not come back on
+  the part it arrived on: `Function call is missing a thought_signature in
+  functionCall parts`. Nothing here was keeping it. The signature is now
+  carried through the loop and handed back untouched -- opaque, unread, and
+  never shown to anybody.
+- **The same failure was waiting in the Claude adapter.** Passing thinking
+  blocks back is required when tools are in play, and the adapter was
+  rebuilding each assistant turn out of its text and its tool calls alone --
+  so the first tool call would have worked and the round after it would have
+  been a 400. The thinking blocks are kept in the order they were produced,
+  signature intact, and go back at the front of the turn they belong to.
+- The loop between the two is what actually lost them, and now has a test of
+  its own: the adapters were each correct in isolation.
+
+Nothing changes for OpenAI-compatible endpoints -- Mistral, Ollama, LM Studio,
+vLLM, OpenAI itself. That protocol asks for nothing back, and inventing
+something to send would be worse than sending nothing.
+
+## 0.6.10 — 2026-09-13
+
+### Added
+
+- **Every model is told where it is before anybody says anything.** A model
+  arrives knowing how to call tools and nothing about why these particular
+  ones behave the way they do -- so a held call reads as a malfunction and it
+  goes looking for a way around it. That is the one behaviour this product
+  cannot tolerate, and it was not the model's fault: nobody had told it. Now
+  it is told, in the same terms for Claude, Gemini, and anything on an
+  OpenAI-compatible endpoint, local models included. What Synartesis is, what
+  the recording is for, what the four classes mean, and the one rule -- never
+  reach for a different tool that does the same thing without being held.
+- **Each tool carries its own class, in its own description.** `write_file`
+  now says it is reversible and that what it replaces is copied first;
+  `move_file` says it cannot be undone and will be held for approval; a tool
+  the policy does not describe says it is treated as the worst case. On the
+  tool rather than in the prompt, because that is where a model looks when it
+  is choosing between two tools that do nearly the same thing -- a note at the
+  top of a conversation is read once, this is read every time the list is.
+- **The briefing is particular, not general.** It names the servers actually
+  connected, the session this conversation is recorded as, today's date, and
+  the calls that will stop and wait -- by name. "Some calls are held" cannot
+  be planned around; "fs.move_file is held" can.
+
+## 0.6.9 — 2026-09-13
+
+### Added
+
+- **A rate limit is waited out rather than handed over.** Every hosted
+  provider refuses when requests arrive too fast, and every one of them says
+  how long to wait. A turn that fails on that has not gone wrong, it has
+  arrived early, and reading a paragraph of JSON and pressing send again is
+  the application failing to do something it knows exactly how to do. It now
+  waits the stated time and tries again, at most twice, and says so on its own
+  line while it waits -- a window that silently stalls for a minute is
+  indistinguishable from one that has hung. It will not retry after the model
+  has started speaking: a retry replays the request from the beginning, and
+  words already on screen would be said twice.
+
+### Changed
+
+- **A quota of zero is not a rate limit, and the difference is the whole
+  point.** A free Google key asking for a Pro model is told `limit: 0` and
+  "please retry in 23s" in the same breath. Both are true and only one is
+  useful: there is no allowance to come back to, so that retry succeeds on no
+  schedule at all. The window now separates them -- a spent quota says how
+  long to wait, a quota that was never there says to turn on billing or pick
+  another model, and nothing waits for it.
+- **The presets start on models a new key can actually use.** Gemini pointed
+  at a Pro preview and Mistral at `mistral-large-latest`; a free Gemini key is
+  allowed none of the former and an entry Mistral key is refused the latter
+  outright. A preset is a first impression, and greeting somebody with a quota
+  refusal on their first message is a poor one. Gemini starts at
+  `gemini-3.8-flash` and Mistral at `mistral-small-latest`, with a note saying
+  what billing buys. Anyone already set up keeps what they chose.
+- **Refusals name the model, not the adapter.** They read
+  "gemini-3.1-pro-preview", not "gemini:gemini-3.1-pro-preview".
+
+### Fixed
+
+- The Gemini adapter's fallback model was `gemini-3-pro-preview`, which Google
+  retired. A default nobody set has to be one that answers.
+
+## 0.6.8 — 2026-09-13
+
+### Added
+
+- **The model name can be changed in the window.** Every provider retires
+  models on its own schedule, and when one goes the request fails with a 404
+  naming its replacement. Until now the only way out was editing a JSON file
+  with the application closed, to keep using a service already paid for. The
+  sheet -- **Models and keys** now, since it does two jobs -- shows the name
+  each model points at and lets it be retyped. The key stays: it belongs to
+  the account, not to the model name.
+
+### Changed
+
+- **A provider's refusal arrives as a sentence.** Each of them says no in its
+  own dialect wrapped in its own JSON, and the window showed the wrapper. The
+  three that happen in practice now read plainly: a retired model says what
+  replaced it, a model the plan does not include says it is the plan, and a
+  refused key points at where keys are kept. The original text is kept
+  underneath, because a guess dressed as an explanation is worse than both.
+
+  Anything unrecognised is passed through exactly as it arrived.
+
+- **Gemini's default is `gemini-3.1-pro-preview`**, which is what Google
+  replaced `gemini-3-pro-preview` with.
+
+## 0.6.7 — 2026-09-13
+
+### Fixed
+
+- **A card no longer says a refused call cannot be undone.** A call that was
+  denied, or is still waiting for a person, changed nothing -- and the card
+  said "no way back recorded" about it, which reads as damage nobody can put
+  back. It now says nothing was applied, and why: refused before it reached
+  the system, or held and waiting for a decision. On this product, of all
+  products, that sentence has to be right.
+
+- **The site could render blank where nothing is looking at it.** Sections
+  arrive as they are scrolled to, which an IntersectionObserver decides -- and
+  an observer only fires while the page is actually being rendered. A tab
+  loaded in the background is not: the browser suspends the lifecycle, the
+  callback never runs, and every hidden state stays hidden. Harmless when
+  somebody switches to the tab, and not harmless for anything that renders a
+  page without showing it. Three seconds with nothing arrived now shows the
+  page whole, finishing the states rather than starting them, since a
+  suspended tab freezes a transition where it stands.
+
+### Changed
+
+- **The logo in the empty room is the pale one**, so it sits on the page
+  rather than on a dark tile stuck to it. Same artwork, same generator.
+
+- **A thousand lines of generated coordinates are gone.** Nothing had drawn
+  the mark from them since the logo became a file; the window, the icon and
+  the cards all show the picture now.
+
+## 0.6.6 — 2026-09-13
+
+### Changed
+
+- **The logo is a file now, and every surface shows that file.** 0.6.5 made
+  the drawings agree; this stops there being drawings. `brand/synartesis-mark-1080.png`
+  is the logo -- the framed mark with its fringe, on its ground -- and the
+  window, the site masthead, the favicon, the touch icon and the cards all
+  display it rather than rendering their own version at their own weight.
+
+  The window's mark stops spinning while a turn runs and breathes instead: a
+  frame going round and round reads as a fault rather than as work.
+
+## 0.6.5 — 2026-09-13
+
+### Changed
+
+- **One logo, everywhere.** The mark is generated -- one set of numbers behind
+  the window, the application icon, the site and the cards on the README --
+  but three surfaces had been drawing their own version of it instead: the
+  rail beside the wordmark, the site masthead and favicon, and the brand PNGs
+  each had a hand-drawn circle with an arrow that resembled the mark without
+  being it.
+
+  They now all show the generated one. Where the fan of five hundred
+  hairlines cannot survive -- a favicon, a 22px rail -- the generator emits
+  the identical geometry filled rather than combed, so it is the same glyph at
+  a different weight rather than a different drawing. `brand/mark.html` draws
+  nothing of its own any more; it displays `synartesis-logo.svg`, which the
+  generator writes.
+
+## 0.6.4 — 2026-09-13
+
+### Changed
+
+- **The API keys sheet gives the key a line of its own.** Somebody opens that
+  sheet to do exactly one thing, and the field for it was squeezed into a
+  column beside a paragraph of prose, sharing a row with the model's name and
+  its note -- while a key is sixty characters or more. The row now explains
+  itself on top and the field runs the full width underneath, with a Cancel
+  that did not exist before, Escape to leave, a wider sheet and room between
+  the rows.
+
+### Added
+
+- **A test for the half the other key tests did not cover.** They prove a key
+  never reaches disk, a log, or the window -- all of which would still be true
+  of a key quietly dropped on the way to the provider. This one follows it in
+  through the sheet, through the keychain seam, and back out as the value the
+  request is built with.
+
+### Fixed
+
+- **A transient API error no longer sinks a release.** Nine installers built
+  and none were published, because the single call that creates the release
+  answered 500 once. Creating and uploading are retried.
+
+## 0.6.3 — 2026-09-13
+
+### Fixed
+
+- **The packaged desktop application could not start.** It opened and died on
+  its first import with `Cannot find package '@modelcontextprotocol/sdk'`.
+  The bundler keeps a package.json's `dependencies` out of the bundle, which
+  is right for a library -- whoever installs it gets them -- and wrong for an
+  application that ships as one file plus one native binding. The MCP SDK,
+  `yaml` and `zod` were left as bare imports and nothing supplied them. Every
+  0.6.2 installer has this; this release is the fix.
+
+  The packaging script now refuses to pack a bundle whose top-level imports
+  are not shipped with it. Electron, the native binding and node builtins are
+  supplied; anything else has to be inside. Only top-level imports are
+  checked, deliberately: bundled libraries carry guarded `require`s for
+  optional native accelerators, and those are allowed to be absent.
+
 ## 0.6.2 — 2026-09-12
 
 ### Fixed

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "synartesis",
-  "version": "0.6.2",
+  "version": "0.6.11",
   "description": "An undo layer for AI agents.",
   "type": "module",
   "private": false,