Twenty-two community pull requests and the reports behind them. Where
v1.6.7 chased a token budget spent on hidden thinking, this one answers
a different complaint: I did a thing and the app quietly did nothing.
New surfaces: a recycle bin that makes a deleted chat recoverable,
search across the full text of your conversation history, and
`knowledge_frontier` — a tool that reads a knowledge base and goes
looking for what is not in it.
Fixes worth naming: a reasoning model that stops mid-thought now fails
with the counters that prove it rather than an opaque error; deep book
URLs open the book instead of the library; the reading outline tracks
the page again; an EPUB that macOS re-zipped reads as the book it is; a
file whose extraction produced nothing is listed with the reason instead
of skipped in silence; the launcher stops killing a backend that is
merely still starting.
Also in this commit: README prose plus all eleven translations, the
v1.6.8 release notes with v1.6.5 archived past the one-week line, three
duplicate keys removed from the locale files (`Delete permanently` had
two different zh values, and the wrong one was winning), and the mypy
annotation that made every `conn.execute` in the new file library look
like a call on None.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Ship the breaking front- and back-end runtime and route refactor,
learner and guardian flows, grounded reading, knowledge and parser
updates, and recoverable long-running streams.
Release notes: assets/releases/ver1-6-3.md
Model configuration was one layer replicated seven times: every service
stored its own profile with its own credentials, so a single OpenRouter
key had to be pasted into five separate pages. This release splits that
in two — a **connection** holds one vendor credential and links it down
to every service that vendor can serve, and a **task model** takes the
calls nobody asks for (naming a conversation, writing the composer's
starting points) off the model picked for actual reasoning. Both are
opt-in; existing per-service profiles resolve exactly as before.
Around that: Settings drops the card grid for a persistent searchable
navigator whose landing page answers "what state am I in" instead of
repeating the directory; Mastery Path files study conversations under
their topic and opens them in the tutor rather than dropping them on
/home; the sidebar folds features into More and reorders by drag; the
reading companion is rebuilt from /home's own chat components; and the
LightRAG engine talks to the SDK directly, so its extra no longer drags
MinerU in transitively.
Also back-merges PR #1041 (Anthropic SDK 1.x temperature via extra_body)
and PR #741 (owner-only auth secret/token files), which had been merged
straight into main and never flowed back to dev.
Release notes: assets/releases/ver1-6-1.md
Gates: pre-commit 15/15, ruff clean, pytest 5802 passed, web 450 node
tests, eslint 0 errors, docs build clean.
The hook was on v0.14.7 while `.github/workflows/tests.yml` installs
ruff==0.16.0. The two disagree on real files — 0.14.7 wanted a blank
line removed that 0.16.0 considers already formatted — so the local gate
and the CI gate were judging the same code by different rules, and which
one you satisfied depended on which ran last.
Verified clean on this tree: `pre-commit run --all-files` passes at
v0.16.0 with no rewrites, and `ruff check .` + `ruff format --check .`
at 0.16.0 agree.
* feat: add comprehensive pre-commit CI hooks
- Add safety for dependency vulnerability scanning
- Add bandit for Python security linting with appropriate skips
- Add interrogate for docstring coverage checking (60% threshold)
- Add mypy for type checking with permissive CI configuration
- Update requirements.txt with safety and bandit dependencies
- Configure all tools for CI compatibility (non-blocking for initial setup)
* security: add hardening measures based on upstream issues
- Add .gitattributes rules to force LF line endings for critical files (fixes CRLF security issues)
- Enhance document validator with MIME type validation and stricter PDF size limits
- Fix subprocess security issues by adding shell=False to prevent command injection
- Add resource exhaustion protection with separate PDF size limits (50MB vs 100MB general limit)
Addresses upstream issues:
- #97: Cross-platform deployment failures due to CRLF line endings
- #31: Large PDF processing failures and resource exhaustion
- Security hardening for subprocess calls to prevent injection attacks
* docs: streamline contribution process and pre-commit checks
- Update and format CONTRIBUTING.md for clarity
- Simplify pull request template
- Adjust pre-commit config to be more lenient locally
* chore: update .gitignore, .secrets.baseline, pyproject.toml, and requirements.txt for improved configuration and dependency management
Squashing some bugs and cleaning up the linting mess:
- Fixed a bug in research.py where 'config' was referenced before it was loaded in the WebSocket handler.
- Cleaned up redundant and unused imports across several agent utility files.
- Re-disabled ESLint in the pre-commit config for now since the /web directory is missing a valid configuration file.
- Normalized end-of-file formatting across the repo.
Cleaning up some critical vulnerabilities found during the last audit. Specifically:
- Upgraded jspdf in web/ to fix a critical Path Traversal vulnerability.
- Removed --exit-zero from ruff pre-commit hooks so we actually block broken code instead of just auto-fixing and moving on.
- Re-enabled ESLint in the pre-commit pipeline to maintain frontend code standards.
- Updated vitepress in docs/ to clear out some moderate CVEs in the dev server dependencies.
