euron-code 0.2.0__tar.gz → 0.4.1__tar.gz

euron_code-0.4.1/CHANGELOG.md

@@ -0,0 +1,247 @@
+# Changelog
+
+## 0.4.1 (2026-11-05)
+
+### Fixed
+- Pasting text with emoji into a Windows terminal crashed the prompt ("surrogates not allowed") and
+  then failed every request. Text is now repaired before it is saved or sent.
+
+### Changed
+- Clearer privacy wording; the limits section shows the daily allowance and where to track usage.
+
+## 0.4.0 (2026-11-05)
+
+### Run the app, not just the tests
+- `start_background_process` (waits for a port or a ready line), `read_process_output`, `stop_process`:
+  at most 3, 64 KB output each, whole process trees stopped on exit, `/clear` and `/rewind`; `/ps`.
+- `check_app`: local URLs only; status, timing, title, `contains`, `wait_seconds`; `browser: true` uses
+  Playwright when installed (new extra `euron-code[browser]`) for console errors, failed requests and a
+  screenshot.
+- The system prompt now says to use these for servers instead of `bash`.
+
+### Reuse Euron's open-source code (V4 contract section 3)
+- `euron_reuse_search` and `euron_reuse_read` over the gateway's index of Euron's public MIT repos; the
+  system prompt asks to search before writing a substantial common component and to keep the licence
+  notice; `/reuse <query>`.
+
+### Context optimisation with Euron Open ContextForge (MIT, vendored in `euron_code/contextforge/`)
+- Before any model-written summary: older duplicate tool results are dropped, bulky logs, test runs and
+  JSON are condensed to their verbatim signal lines (errors, failures, stack frames, head and tail), and
+  stubs keep the facts they drop. Often that is enough and no model call is spent.
+- Every summary now carries a verbatim fact ledger (ids, paths, versions, env names, URLs, ports, dates,
+  error classes) and the user's stated constraints, plus permission grants, git branch and state, and the
+  last command with its exit code.
+
+### More tools and gaps closed (audit increment 6)
+- `find_symbol` (definitions and references), `delete_file` and `move_file` (both undone by `/rewind`).
+- `AGENTS.md` and `CLAUDE.md` are read alongside `EURON.md` (trust-gated like it).
+- Inline `<tool_call>` JSON is repaired (trailing commas, single quotes, Python literals); garbage
+  becomes an error result for the model, never a crash.
+- `/rewind` works after `euron --resume`: checkpoints are kept in `~/.euron/checkpoints/<session>/`
+  (user-only, at most 50 MB per session).
+- The context window and output cap come from `/v1/models` unless you set them.
+- `/export` warns that the file contains tool output.
+
+### Euron-only client and privacy (V4 contract section 1, audit increment 3)
+- A 403 `client_not_supported` shows the gateway's message ("works through the euron app only").
+- README and the first image attach say where data goes: prompts on GPUs rented from Vast.ai (not
+  logged), images described by Google Gemini (not stored).
+
+### CI
+- `.github/workflows/ci.yml`: pytest on ubuntu, windows and macos x Python 3.9 and 3.12 (ripgrep on one
+  job), then build the wheel and sdist and check the wheel contents.
+
+## 0.3.0 (2026-11-05)
+
+### Never stuck at the output limit
+- An answer cut by the output limit (`finish_reason: length`) is continued automatically (up to 3 times)
+  and joined into one answer. A tool call cut mid-arguments (typically a huge `write_file`) is never run;
+  the model is asked to make the change in smaller steps. When reasoning alone used the budget, the
+  request is retried once with `reasoning_effort: low`. Each case shows one dim note, never an error.
+- `reasoning_effort` is sent on every call: `medium` by default, `low` for simple follow-ups
+  ("thanks", "continue") and continuation retries. `/effort low|medium|high`, setting `reasoningEffort`,
+  shown in the status bar.
+- `max_tokens` is sized per request: min(server max, context window - prompt estimate - 512), at least
+  1024. A request that would not fit is compacted first. The prompt estimate starts from the last
+  server-reported `prompt_tokens` and adds only the new messages.
+
+### Context management for a 32k window
+- An output budget is reserved (min(maxOutputTokens, window/4)); percentages are of the usable rest.
+- From 50%, tool outputs are cleared biggest and oldest first until under 40% (the latest step is never
+  touched). Each stub names the call (for reads: path and line range) so the model can read it again.
+  The note reports real numbers: "freed 9,400 tokens, now 38%".
+- The conversation is summarised at 70%, or whenever the next request would not fit. The summary keeps
+  the task, decisions, files changed, commands and results, the last error and open todos (files
+  changed and the last error are also added verbatim).
+- `read_file` reads 200 lines by default and stops at a line boundary before ~6k tokens. Files over
+  2,000 lines return an outline (definitions with line numbers) unless a range is given. The system
+  prompt says: search first, read ranges, edit with `edit_file`, never rewrite big files.
+
+### A live session like Claude Code
+- **Activity line** while the agent works: an animated glyph, a rotating verb, elapsed time, tokens
+  sent and received (counted live from the stream), what is happening (`thinking`, `writing app.py`,
+  `running pytest 12s`, `waiting for a GPU slot (2 ahead)`, `retrying in 3s`) and `esc to interrupt`.
+  On a terminal it redraws 10 times a second; without one it prints a progress line every 15 s; in `-p`
+  mode it appears on stderr only when stderr is a terminal. The gateway's `: queued, N ahead` comments
+  are shown instead of being ignored.
+- **Status bar**: model and effort, mode, session tokens and their value, context %, today's and this
+  week's remaining quota, session uptime. Refreshed continuously.
+- **Type while it works.** The turn runs in a worker thread and the prompt stays live. `/usage`, `/cost`,
+  `/context`, `/status`, `/model`, `/effort`, `/help`, `/todos`, `/permissions` run at once (a model switch
+  applies from the next model call, a mode change from the next tool call); anything else is queued as
+  your next message. Permission questions are answered in the same prompt. Esc or Ctrl-C stops the
+  turn (a running command is killed) and puts queued messages back into the input line.
+- **`/` menu** pops up as soon as you type `/`, with every command, custom command and skill. `/model`
+  opens a picker (arrow keys + Enter) with context size and status. `@word` searches the project's files.
+- `/status`, `/whoami`, `/config`, `/bug`, `/release-notes`, `/users` (alias of `/usage`).
+
+### Accounts and modes
+- Plain `euron` with no saved key logs you in in the browser and goes straight into the session.
+- `euron login` / `/login` when logged in: shows who, offers to log in as a different account
+  (`--force` skips the question). The old key is replaced only after the new login is approved. Logout
+  stays local (there is no CLI revoke API); revoke keys at euron.one/code-agent.
+- `--permission-mode bypassPermissions` and `--dangerously-skip-permissions` (alias of `--yolo`) with a
+  red warning; Shift+Tab reaches bypass only when started with it. `acceptEdits` is shown as "auto-apply
+  edits". `--add-dir DIR` adds folders the agent may work in.
+
+### One login at a time
+- A key revoked because the account logged in on another device (401 `signed_in_elsewhere`) shows that
+  message, drops the local key and offers "Log in here? (this signs out the other device)"; after the
+  login the interrupted message is carried on. `-p` exits 4 with the message.
+- When another session holds the account (409 `session_in_use`), Euron Code asks "Take over here?"; yes
+  retries once with `X-Euron-Takeover: 1`. `-p` exits 4 unless `--takeover` is given.
+- After `euron login`: "Signed in as ... (any other device using this account was signed out)".
+
+### Plan quotas and value (V3 contract)
+- `/usage` shows the plan tier, all four periods (day, week, month, year) with bars and refill times,
+  `quota.blocked` prominently, and value lines at the reference API price ("≈ $1.07 (₹94) at Claude
+  Sonnet 5.5 API list prices · included in your plan"), plus what the allowance is worth. `/cost` and the
+  status bar show this session's value. Always labelled as a reference, never a bill.
+- New 429 codes `monthly_token_limit`, `yearly_token_limit` and `cooldown`, each with its refill time in
+  local time; exit code 5 in `-p` mode.
+
+### Automatic project setup and local cache
+- First `euron` in a folder without `EURON.md` or `.euron/` (not your home folder or a drive root) offers
+  to set it up; `--yes` or `euron init` does it without asking. Locally, with no model call: `EURON.md`
+  (stack, commands, layout, entry points), `.euron/rules/` (testing, git, security and per-stack
+  rules), `.euron/settings.json` (allow rules for the detected test/lint commands and read-only git),
+  `.euron/hooks.example.json` (off; `/hooks enable` turns them on, pre-trusted because Euron Code wrote
+  them), `.euron/commands/` (`/fix-tests`, `/commit`, `/pr`), and `.euron/cache/` in `.gitignore`.
+- `.euron/cache/` keeps the repo map and file index, keyed by git HEAD and folder mtimes. Safe to delete,
+  no secrets, only written when `.euron/` exists.
+
+### Attach anything
+- `@file`, `/attach`, or drag files into the terminal (Windows `"C:\..."` and PowerShell `& '...'`
+  quoting, Unix escaped spaces and `file://` are recognised). Word, Excel, PowerPoint and PDF become
+  compact text locally (stdlib zip+XML; `pypdf` for PDF), as do CSV/TSV (sample + column summary),
+  JSON, YAML and notebooks, capped at about 6k tokens per file with a note; unknown binaries are only
+  described. The new `read_document` tool reads more pages, sheets or rows on demand.
+- `@folder` attaches a digest within a budget: tree, file types and sizes, heads of key files.
+- Images (png/jpg/gif/webp/bmp, up to 5 per message, 5 MB each) are sent as `image_url` parts, resized
+  to 1568 px with the optional `euron-code[images]` extra (Pillow), which also enables Ctrl+V and
+  `/paste-image`. A 400 `images_unavailable` drops the images with a clear note and retries.
+- Session transcripts never store image bytes.
+
+### Security (independent audit, before release)
+- **Project trust.** A repo's `.euron/settings.json` / `settings.local.json` can no longer loosen anything
+  silently: allow rules, `defaultMode` other than default/plan, `shell`, `db.*`, `webFetch: true` and
+  raised limits apply only after the user trusts the repo (one prompt listing exactly what it ships,
+  with view / trust / ignore; fingerprinted in `~/.euron/trusted-projects.json`, so a change asks again).
+  Restrict-only keys (deny rules, plan mode, lowered limits) always apply. `-p` ignores untrusted keys
+  with a warning; `euron trust` / `/trust` trust a repo. Rules the user saves with `s` or
+  `/permissions allow` now live in `~/.euron/projects/<hash>.json`, outside the repo.
+- **Repo instructions are trust-gated too:** project and parent `EURON.md`, `.euron/rules`, `.euron/skills`
+  and `.euron/commands` load only for a trusted repo. The system prompt no longer says project memory
+  "overrides" anything: project-provided guidance follows `# Tools and safety`, which always wins.
+- `grep` skips secret-looking files (`.env*`, keys, credentials) in both the ripgrep and the Python path
+  and says which were skipped; naming one explicitly asks, like `read_file`.
+- Commands, `!` commands and hooks run without `EURON_API_KEY` / `EURON_CODE_API` in their environment.
+- Bypass mode still asks for commands that touch existing paths outside the project and for a list of
+  destructive commands (that list asks even with an allow rule); `web_fetch` to localhost, private
+  ranges, link-local and metadata addresses asks in every mode unless that exact host was allowed.
+- Edits to `EURON.md` and `.euron/**` always ask in acceptEdits mode.
+- Session transcripts are created user-only (0600, folder 0700) on POSIX.
+
+### Runaway guard and honest completion
+- 3 identical failing calls add a "change your approach" note; 5 stop the request (`stop_reason: loop`).
+  Identical successful calls get a note at 4 and stop at 8.
+- Step budget per request, `maxStepsPerPrompt` (default 60): the REPL asks to continue; `-p` exits 6.
+- A local report after a change: `Changed: ... | Ran: ... (exit N) | Not verified: ...`; `-p --output-format
+  json` gains `files_changed` and `commands`. One automatic nudge when files changed and no command
+  succeeded afterwards (`verifyNudge`, off in plan mode); the detected test command is in the
+  environment block.
+- After `write_file` / `edit_file`: a syntax check for `.py` (compile), `.json` and, when node is installed,
+  `.js` files.
+
+### Fixed
+- Windows Terminal, VS Code and other ConPTY hosts: a lone Esc stayed inside prompt_toolkit's VT input
+  parser until the next key, so Esc did not interrupt. The timed flush now delivers it (found by
+  driving the CLI in a real ConPTY).
+- Safe commands such as `/usage` typed while a permission question is open now run, and the question
+  stays open, instead of being rejected as an invalid answer.
+
+### Changed
+- Defaults: `compactThreshold` 0.7, `microCompactThreshold` 0.5, new `microCompactTarget` 0.4,
+  `microCompactKeepSteps` 1, `outputReserve`, `reasoningEffort` (all fractions of the usable window).
+- The CHANGELOG ships inside the wheel (for `/release-notes`).
+
+## 0.2.0 (2026-11-05)
+
+### Added
+- **Usage periods.** `/usage` and `euron usage` show Today / This week / This month with a progress bar
+  against each limit (the week only when the plan has a weekly limit), reset times in local time,
+  requests, sessions, tasks and the share of prompt tokens served from the server's cache, plus the
+  30-day table with cached / sessions / tasks columns. The status bar shows the tighter of the day and
+  week allowance.
+- **Weekly limit.** A `weekly_token_limit` 429 stops cleanly with the local reset time; exit code 5 in
+  `-p` mode, the same as the daily limit. Both limits read `resets_at` from the error when present.
+- **Session and task headers.** Every model call sends `X-Euron-Session` (the saved session id, kept on
+  `--continue` / `--resume` / `/resume`) and `X-Euron-Task` (new for every message you send).
+- **Rules files.** `.euron/rules/*.md` and `~/.euron/rules/*.md`, loaded after `EURON.md`. Rules with
+  `paths:` front matter load only once a matching file is read or edited. `/rules`, `/rules new`.
+- **Skills.** `.euron/skills/<name>/SKILL.md` and `~/.euron/skills/<name>/SKILL.md`. Only names and
+  descriptions are in the system prompt; the new `skill` tool loads a body on demand; `/<name> [args]`
+  runs one. Built in: `review`, `test-writer`, `explain-codebase`. `/skills`, `/skills new`.
+- **Custom commands.** `.euron/commands/<name>.md` and `~/.euron/commands/<name>.md` become `/<name>`
+  with `$ARGUMENTS`; shown in `/help` and Tab completion.
+- **Hooks.** `SessionStart`, `UserPromptSubmit`, `PreToolUse`, `PostToolUse`, `Stop` in settings, with a
+  tool-name `matcher`, JSON on stdin and exit code 2 to block or give feedback. Project hooks must be
+  trusted once per project (re-asked when they change); skipped with a warning in `-p` mode.
+  `/hooks`, `/hooks new`.
+- **Plan mode.** Read-only mode in which the model proposes a plan; leaving it is the user's choice.
+  Shift+Tab cycles default -> acceptEdits -> plan; also `/plan` and `--permission-mode plan`.
+- **Shortcuts.** `!command` runs a command and adds its output to the next message; `@path` attaches a
+  file or folder (Tab completes paths); Esc Esc / `/rewind` undoes the last turn's file edits and drops
+  the turn; Ctrl-L clears the screen; `/resume` and `/export` inside a session.
+- **`/context`.** A breakdown with a bar: system prompt, memory, rules, skills index, tool definitions,
+  messages, tool results, free space.
+- **`/init` repo map.** A locally built map (stack, manifest commands, layout, entry points) is handed to
+  the model, which writes a short `EURON.md` with a repo map section; offers `.euron/rules/` and
+  `.euron/settings.json` skeletons.
+
+### Token saving (automatic)
+- The system prompt and tool list stay byte-identical for the whole session; plan mode, hook output and
+  `!` output go at the end of the user message.
+- Re-reading an unchanged file (mtime + size, or content hash) returns "unchanged since your last read
+  at turn N"; a read that reached the end of the file covers later partial reads.
+- Micro-compaction at 60% of the context: tool outputs older than the last 3 steps become one-line stubs
+  (no model call); the summary still happens at 75%.
+- `/cost` and `/context` report prompt tokens sent, tokens served from the server's prefix cache
+  (`usage.prompt_tokens_details.cached_tokens`) and tokens saved by each measure. `-p --output-format
+  json` adds `cached_tokens` and `saved_tokens`.
+
+### Changed
+- `/help` lists your commands, skills and the shortcuts.
+- `--permission-mode` and `/permissions mode` accept `plan`.
+- New settings: `microCompactThreshold` (0.6), `microCompactKeepSteps` (3), `hooks`.
+
+### Fixed
+- On Windows PowerShell 5.1, a command that wrote to stderr (for example `pytest ... 2>&1` with a
+  warning) was reported as exit code 1 with CLIXML noise even when it passed. The real exit code is
+  reported now and stderr is plain text.
+
+## 0.1.0
+
+First release: agent loop with file, search, shell, database, web, todo and sub-agent tools; device
+login; `EURON.md` memory and `/init`; `/compact` and auto-compaction; saved sessions; permission rules
+and modes; `-p` headless mode.