heartcode 0.1.15 → 0.1.24

package/CHANGELOG.md

@@ -2,6 +2,192 @@
 
 What changed in each version of HeartCode, newest first.
 
+## 0.1.24 (2026-10-31)
+
+- A pinned model (one picked with /select_model, or -m) is now used by the helpers it spawns too:
+  with a model pinned, HeartCode won't hand a helper's job to another of your models, and a
+  helper uses the pinned model even if a different one is named. Auto mode still spreads work
+  across your models as before.
+- A built page is now actually driven like a user before it's called done, not just loaded: HeartCode
+  fills the inputs, clicks every button, and toggles every checkbox in the headless browser, and
+  any error a control throws (a dead button, a popup that fails) is a must-fix, fed back to the
+  model to fix and re-test — like Replit's test-and-fix loop. (Navigation is blocked so one click
+  can't end the test.) These failures also become lessons.
+- HeartCode now learns from its mistakes in any language, not just web: when something it ran
+  failed and then passed after a fix (a Python traceback, a failing test, a build error, a
+  broken button in the browser…), it saves the fix as a one-line rule in memory, tagged by
+  technology, and recalls it on later tasks of the same kind so the mistake isn't repeated.
+  Near-duplicate lessons are skipped. A rerun only counts as a fix if a file was edited between.
+
+- When the model uses any external tool, library, API or framework, it now goes to the official
+  source first — always, even when it's sure it knows — assumes its memory is stale, relearns the
+  current way from that official page, then acts on it. Official docs/sites/repos come first;
+  blogs, listicles and memory are a last resort. (A change to the project's own code, or a question
+  the code answers, still needs no lookup.)
+  This reminder is repeated across the prompt (research, how it works, on a failure, using a tool,
+  every message, and for small models) so a model is far more likely to actually do it.
+
+## 0.1.23 (2026-10-31)
+
+- The "retrying in 0 seconds" message no longer lingers on the spinner while the retried request is
+  in flight: the countdown is cleared the moment the wait ends and the new request starts.
+- "Make sure it works" now applies to everything the model builds, not just web pages: before it
+  finishes it must run what it made and see it work — a CLI with real input, a server by calling
+  its endpoints, a library by a small usage snippet, tests if there are any — not just "no error".
+- A page the model builds is now checked that it actually works, not just how it looks: HeartCode
+  reads the browser console when it opens the page, and any JavaScript error (an unclosed brace, a
+  wrong name, a blocked resource) is a must-fix blocker fed back to the model before it can finish.
+- While retrying, the retry message and the "responding / waiting for model" timer are on separate
+  lines instead of one long line that got cut off.
+- A page the model builds is now checked that it actually works even when no browser is available,
+  and verification never passes silently: HeartCode statically catches JavaScript that doesn't
+  parse and the Tailwind-CDN @apply trap (applying a class you defined yourself, which throws
+  "Cannot apply unknown utility class"); and if it couldn't open the page in a browser at all,
+  the model is told so and must verify it itself before finishing, instead of assuming it's fine.
+- A model pinned in one CLI (/select_model) is now truly session-local: changing the model order
+  on the settings page (or in another instance) no longer unpins it or switches its model — that
+  only affects instances on auto. The pin isn't written to disk either, so it can't become
+  another instance's #1. (Before, a model-order change silently switched every running instance.)
+- /select_model now picks the model(s) for this session: tick one to pin it (no rotation), several
+  to rotate among just those, or all/none for auto (your whole list). /select_model <model> pins,
+  /select_model a, b, c chooses a set, /select_model auto goes back to the list. It's session-local
+  — your saved ranking (settings page) is untouched and stays the default for new sessions. The
+  status bar shows it: "auto (model) (N)", a pinned model's name alone, or "model (K chosen)".
+  A single pinned/chosen set is also what spawned helpers use (no handing off outside the set).
+  In the picker, Ctrl+A ticks all and Ctrl+N unticks all, so you can clear then tick what you want.
+- Command and @file suggestions now show while the model is working, not only at an empty prompt:
+  type `/` or `@`, use ↑/↓ and Tab to complete; Enter queues the command to run when the task is done.
+- The browser (chrome-devtools) no longer litters the project: it runs in its own scratch directory,
+  so a screenshot or snapshot the model saves by a bare name (screenshot.png, snapshot.txt) lands
+  there instead of the project folder. Pages are still opened by absolute path, so nothing breaks.
+- The built-in Chrome (chrome-devtools) runs headless (no window) with its own throwaway profile, so
+  it doesn't pop up a window, works over SSH, and runs don't fight over one profile. Override `args`
+  in ~/.heart/mcp.json for a visible browser.
+
+## 0.1.22 (2026-10-31)
+
+- Times are written out: "responding 5 hours 33 minutes", "took 1 minute 25 seconds", "retrying in
+  10 seconds", "thought for 12 seconds", "resumed … 3 hours ago" (the two biggest units).
+- Ctrl+R no longer starts voice: /voice does (keybindings.json can still give it a key).
+- Click a tab to go to it.
+- /rewind puts a message of several lines back in the input box whole (only its last line came back).
+- The key hints say what a key does: "Shift+Tab change permissions", "Ctrl+T open a new tab".
+- Finding its way in a project, in this order: the model sees every entry of the project root
+  (names only; what .gitignore leaves out is marked), then goes folder by folder (list_files with
+  a path) and searches; a language server only when the task needs code followed across files.
+- Language servers are the model's choice, for any language: it finds out which one fits, checks
+  it's installed, and uses it with the new lsp tool (definition, references, hover, errors).
+  HeartCode remembers it for the project, and checks each edit of such files with it. No built-in
+  list any more.
+- find_command: the model checks whether a command is installed before it relies on it (instant,
+  runs nothing).
+- Building a UI: the model researches the best real examples of that exact kind of screen and
+  the current UI library for the stack, writes a design brief and its own checklist from them,
+  builds with the library, then looks at screenshots (desktop and phone) and fixes what's off. A
+  model that can't see has one of your models that can check them.
+- The status bar says how many providers are on and how many models are in your list:
+  "local (2) · qwen3 (3)".
+- A queue for messages typed while it works: Enter adds one (shown above the input, numbered).
+  When the answer is done they go all together, as one message, or one by one: Ctrl+_ switches
+  (saved). Ctrl+] stops the answer and sends the first one now, Ctrl+\ sends them all now, as one
+  message (one request). The model no longer gets them in the middle of an answer.
+- Web search asks several engines at once and puts their results together: public SearXNG
+  instances (from searx.space; five at a time, the next five when none answers well, pages 2–3
+  when the first has few good results), Mwmbl, DuckDuckGo, and for coding questions Stack Overflow,
+  npm and GitHub. Each result is ranked by the query's words in it (rare words count more) and by
+  how many engines found it, and says where it came from. Searching fails only when every engine
+  does; then the model is told search can't be reached, instead of "try other words" again and again.
+- What the model makes: stability and UI/UX come first.
+- A page it builds or restyles is opened in the built-in browser when it says it's done: a model
+  that can see reviews the screenshots (desktop and phone width) in a request of its own and lists
+  what's wrong, and the model fixes it (twice per message at most).
+- A reply that goes into a loop (the same text over and over) is cut, and the model asked again.
+- A path the model copied wrong (the project's long path without its first "/", or with folder
+  names run together) is read as the project path it meant; a command that went outside the
+  project says where the project is.
+- Searches, page reads and file reads asked for in one reply run at the same time, and servers are
+  asked for several calls per reply (parallel_tool_calls). The model is told to look things up
+  together, and to get everything a task needs before doing it, as much as it takes.
+- A search, page, read or command made again in the same message, with nothing changed since,
+  isn't run again: the model gets the earlier result. Ten such repeats stop the message; four
+  failing commands in a row get a "rethink this" note.
+- The model researches the kind of thing it's asked to make, not the names in the request, and
+  installs only what its plan needs, after making sure it's the real one.
+
+## 0.1.21 (2026-10-31)
+
+- How long HeartCode took to answer: the spinner counts "responding 45s" while it works, and when
+  it's done a line stays under the answer ("✻ took 1m 12s"). Time spent waiting for your answer
+  to its questions isn't counted, and the line doesn't keep counting after it's done.
+- The mouse wheel scrolls just the conversation: the tabs, the input and the status bar stay where
+  they are.
+- Select text in the conversation by dragging with the mouse: it's highlighted and copied when you
+  let go ("copied 42 characters"). A click or a key clears it.
+- Keys to scroll the conversation: Shift+↑/↓ (3 lines), Alt+↑/↓ or PgUp/PgDn (a page).
+- Closing a tab (Ctrl+Q, or the new /close_tab) asks first.
+- The model's answers are drawn as Markdown as they stream in: headings, **bold**, *italic*,
+  `code`, lists, quotes, code blocks and rules (/copy and /export keep the Markdown as written).
+  The thinking too, and answers in sessions and tabs from before, when they're shown again.
+- The model is told to give choices as a numbered list, and to keep to simple Markdown (no tables).
+
+## 0.1.20 (2026-10-31)
+
+- The settings page's accent is green.
+- A System / Light / Dark switch at the bottom of the settings page's sidebar (System follows your
+  computer's setting).
+- Voice works like Claude Code's: /voice (or Ctrl+R), Space to talk (again to pause, again to go
+  on), Enter to finish, Esc to cancel; what you say shows in the input box as you say it.
+
+## 0.1.19 (2026-10-30)
+
+- The settings page, redone: one accent color, calmer lists (a row's buttons
+  show when you point at it), clearer type, hover, press and focus states, dialogs that animate in,
+  and a layout that works on a phone too.
+- A provider switched off is gone from everywhere it was still showing: the status bar (it showed
+  the off provider's model when only one other model was on), `heart providers`, /select_model,
+  and on the settings page Your models, its count, Test all and the provider ranking. Its models
+  keep their places in your list for when it's on again.
+- /rewind numbers your messages, newest first.
+- The settings page says "Disconnected" and can't be used once it's turned off (in HeartCode or
+  with Ctrl+C), instead of failing quietly.
+- The compaction switch on the settings page says 85%, as it is now.
+
+## 0.1.18 (2026-10-30)
+
+- Before building anything, the model researches it first (the current, most recommended way and
+  the latest stable versions, from the web and the official docs), plans the whole job, and asks
+  what it still isn't sure of: free models' training data is old.
+- Tab names are at most 20 characters, "..." included.
+- Settings changed while HeartCode works now reach it at once, every time: before every request and
+  before each tool's result is cut to the context window, HeartCode takes in what changed (it no longer
+  depends on the file watcher being quick enough). A model's context window, set up on the settings
+  page mid-task, is used from the next step, also by helpers (the agent tool) already at work, which
+  used to keep the old one until they finished.
+- Auto-compact when the conversation takes 85% of the window of the model answering (room is kept
+  for the reply), no longer every 100 messages: a long session is no longer summarized at a few
+  percent of a big window. When the model list moves on to a model with a smaller window, the
+  conversation is compacted for that model before it's sent, instead of being refused as too big.
+  The spinner says how full it is ("context 86% full (~660k of 768k)").
+- A running HeartCode no longer overwrites what you set up for a model: saving what it learned
+  about the model (token counts, limits) used to write its old copy of your settings back over them.
+
+## 0.1.17 (2026-10-30)
+
+- Stronger system prompts (tool mode and guided mode for small models): act instead of describing,
+  don't guess paths or APIs, search the web when unsure or knowledge is dated, never stop until the
+  request is done, verify by output, and answer short and to the point. Small local models follow
+  these far better.
+- The system prompt now tells the model the OS and shell it's running on (e.g. "macOS (darwin, arm64),
+  shell zsh"), and how to find and install a tool it needs: use what's already there, else install it
+  the right way for this OS (asking first unless commands are already allowed).
+
+## 0.1.16 (2026-10-30)
+
+- Switch a provider off (and on again) on the settings page, without deleting it: its key and its
+  models in your list stay, but `auto` skips them, also in a HeartCode running already. `-m
+  provider/model` still runs one of its models. `heart providers` says which are off; `heart models
+  --all --json` gives them `off: true`.
+
 ## 0.1.15 (2026-10-30)
 
 - Settings page, Providers & models: Test on each model sends it "Hello" and shows what was sent

package/SKILL.md

@@ -50,7 +50,8 @@ No model set up yet: `models` is empty (and `heart -p` fails with "no model set
 `heart models [--all] <filter>` keeps only the ids that contain `<filter>`.
 
 `providers` (only with `--all`): every provider's models. `list: false`: the user switched that
-provider's list off, and `models` are the ones they added by name. A provider that can't be
+provider's list off, and `models` are the ones they added by name. `off: true`: the user switched
+the provider off; `auto` skips its models (they aren't in `models`), but `-m provider/model` still runs one. A provider that can't be
 reached has an `error` (in text: a line on stderr) and is skipped; the command still exits `0`.
 
 **Which model to pass:**
@@ -70,8 +71,9 @@ all refused the request outright, else after `--max-retries` (10) whole rounds o
 model (`-m <model>`) waits out rate limits and outages, up to `--max-retries` tries, and stops at
 any other error. A provider that refuses a part of the request HeartCode can do without (the
 model's thinking sent back, keys in the tools' schemas, images) gets it again without that part,
-and that is remembered for the model for a day. A long conversation is summarized every 100
-messages.
+and that is remembered for the model for a day. A conversation is summarized when it takes 85% of
+the window of the model answering (room is kept for the reply); when the list moves on to a model
+with a smaller window, it's summarized for that model before it's sent.
 Each request is fitted to the model it goes to: one that can't see gets a note in place of each
 image, one that can't call tools gets the tool calls as text.