LLMLinq v0.1.0

The version restarts here. Earlier 0.1.x builds described a viewer for CLI
history; this one drives those CLIs, so the number begins again alongside the
thing it counts.

Added

• My work — one page for whichever AI CLIs this machine has. Write a
  prompt, pick a tool, send it, and the answer is kept beside the prompt.
  Send the same prompt to a second tool and both answers sit side by side,
  which is what makes them worth comparing. The tool you pick is the tool
  that runs: a failure is reported in that tool's own words rather than
  answered by something else under a different name. Your prompt goes to the
  tool exactly as written, with nothing of ours wrapped around it. Prompts
  and answers stay on this machine; only the prompt is ever synced.
  Shaped like the chat tools people already use — a thread you read and a box
  you type in. There is no title to fill in: a prompt names itself from its
  own first line, and that name follows the prompt when it is edited.
  Answers render as what they are, so a fenced code block is a code block you
  can copy rather than a wall of backticks. History down the left is
  everything sent from here; opening one shows the prompt and every answer
  to it, each one copyable on its own.
• Generate prompt lives with the prompt box, where you are writing the
  thing it improves — before it is sent, where it can still change what comes
  back.
• Prompt generation on your own machine. A rough request becomes a precise
  prompt, written by Claude Code, Codex, Antigravity or Gemini under your own
  CLI login. No key to paste, no daily allowance, and the request never leaves
  the machine. Pick which CLI writes it, or let it use whichever answers.
• Your own API key comes first. Configuring one says which model you want,
  and you are paying for the tokens, so it leads. When it is rejected or out
  of credit you are told in the provider's own words, and a local CLI covers —
  never a silent swap.
• Flows, of two kinds, and the difference is the point.
  Code review points Claude and Codex at the same change — two branches in
  a local repository, or a GitHub pull request — and they read it themselves,
  running git or gh and opening the files around it. Nothing is pasted in: a
  diff cannot answer most of what a review asks, because what decides it
  usually lives in the code that did not change. Findings are grouped by file
  and severity, and where two reviewers independently reach the same line that
  agreement is marked.
  Release readiness runs your project's own tests, linter and audit: every
  gate is a fact rather than an opinion, which is why only this one is allowed
  to decide what ships. Review a diff against a branch, or the whole project.
• Flows of your own. Pick the kind, pick which CLIs review or which gates
  run, and paste a review prompt you already trust. A review flow may carry
  any instruction, because prompt text is inert; a gate flow may only choose
  gates that already exist, because a gate carries a command.
• Ollama, listed beside the CLIs, with the model it is actually running.
  On a machine with no CLI signed in it is the thing that works. Nothing is
  installed on your behalf, and when there is nothing at all the guidance
  leads with the free route.
• Share prompts you generate with the resource gallery, off until you turn
  it on.

Changed

• One roster of local tools. Every place that offers a tool — the picker,
  prompt generation, My work, Flows — reads the same list from one place, and
  the runner honours exactly what that list offered.

• Navigation follows the work: My work, then Flows, then Timeline. My
  work is the page LLMLinq opens on.

• Generate prompt moved to the prompt box on My work, where you are
  writing the thing it improves — rather than after the fact.

• The Compose library became My work. Prompts saved there are carried over on
  first run, and the file they came from is left untouched.

• Nothing is metered any more. Generation runs under your own login, so
  the daily allowance, its model categories and the usage page are gone.

• The mark now matches llmlinq.com.

• Sync chat history sends the whole history — every prompt and the
  answer each tool gave — to your account in one press, down the same path
  the timeline already used. It lands beside your prompt timeline on the
  portal, under its own page, and one search covers both and matches on
  answers as well as prompts.

Removed

• The share-with-the-gallery checkbox. A prompt an AI CLI writes for you
  is now always offered as a candidate — never what you typed, never your
  history, never your answers — and the pipeline it goes through was made
  much stricter to earn that: names, dates of birth, addresses, postcodes,
  national IDs, IBANs, payment cards (Luhn-checked), credentials in
  connection strings, JWTs and API_KEY= assignments are all stripped by
  fixed rules; key material rejects the candidate outright rather than being
  redacted; an entropy check rejects long tokens no rule recognises; text
  still reading as being about a particular person is rejected; and a second
  model reviews the rewrite, asked only whether anything personal survived.
  Every gate fails closed, including one that could not run.

• Per-prompt Sync in My work. Syncing is something you do to the
  history; a button on every entry made a readable history a chore to
  assemble.

• Revise. Rewriting a prompt after the fact edited the record of
  something you had already asked. Improving a prompt now happens on My work,
  before it is sent, where it can still change the answer. Existing revisions
  are left on disk untouched.

Fixed

• The AI CLI that works is the one used. Discovery stopped at the first
  copy on PATH, which found a Homebrew codex 117 releases out of date that
  failed every request with an error blaming the model. Every copy is now
  found — including builds bundled inside editor extensions, which is where
  the working one lives — and the newest is used.
• A CLI that cannot answer says why. "Not signed in" is reported as itself
  rather than as a timeout, and a CLI that just failed is stepped over for a
  few minutes so every generation does not wait on it.
• Choosing Ollama ran a CLI instead. The picker offered Ollama, but the
  list the runner sorted by preference held CLIs only — so the choice was
  dropped silently and a CLI answered under its own name. Being served a
  different tool than the one you picked, without being told, is the one
  thing this must never do.
• Only a tool that can go and look may review. The flow builder offered
  every tool it found, including a local model served over HTTP — which can
  answer a prompt but cannot run git or open a file, so it would have
  reported confident findings about code it never saw. Picking one that
  cannot review also fell back to an arbitrary member of a set, so the same
  review could run as a different reviewer each time; it now falls back in
  preference order.
• A stale name in the review module's exports broke ruff on main.
• Release builds no longer depend on Actions artifact storage. That is a
  fixed account-wide quota which release builds fill on their own, and when
  it ran out every build failed at the upload step — binary built, smoke
  test passed, nothing published. The platform builds now hand their
  archives to the publish step through an invisible draft release instead,
  which has no quota. Everything still lands on the live release in one
  step, so it never holds new archives beside an old checksums file.

Install — macOS / Linux:

curl -fsSL https://llmlinq.com/install.sh | sh

Install — Windows (PowerShell):

irm https://llmlinq.com/install.ps1 | iex