contextzip/CHANGELOG.md

Changelog

All notable changes to contextzip are documented here.

The format follows Keep a Changelog. This project uses Semantic Versioning.

---

[0.1.1] — 2026-05-15

Fixed

• LICENSE link in README now points to GitHub instead of a relative path that broke on PyPI
• License metadata updated to SPDX string format ("MIT") to comply with setuptools ≥ 77

---

[0.1.2] — 2026-05-15

Added

• Smart framework and language detection (Next.js, Node.js, Python, Django, FastAPI, Rust, Go, Ruby)
• Automatic exclusion rules per detected ecosystem
• .gitignore patterns respected automatically (--no-gitignore to bypass)
• --include flag for scoping to specific directories (exact path-boundary matching)
• --exclude flag for additional gitignore-style exclusion patterns
• --dry-run flag to preview without creating any files
• --verbose flag showing every file with its size
• --output flag for custom ZIP destination
• --no-clipboard flag to skip clipboard/folder-open step
• Tiered clipboard strategy: file-on-clipboard (macOS/Linux), Explorer with file selected (Windows), path fallback
• Warnings for large files (≥ 1 MB), binary files, and skipped files (dangling symlinks, permission errors)
• Rich terminal output with panels, progress bars, and colour-coded detection results
• Compression stats (ratio, size before/after) shown after ZIP creation

---

[0.2.0] — 2026-05-15

Added

• --git-changes mode for packaging only files reported as changed by git
• Automatic detection of:
  • staged files
  • unstaged modifications
  • untracked files
• Dedicated git-aware packaging pipeline optimised for AI/code review workflows
• Rich git change summary panel with staged/unstaged/untracked breakdown
• Verbose git mode output showing categorized changed files
• Graceful handling for:
  • repositories with no changes
  • deleted files
  • submodules
  • repositories outside the selected project tree
  • missing git installations
  • non-git directories
• Support for monorepo-aware repository root resolution
• Binary and large-file checks now also apply in git-changes mode

Improved

• File resolution architecture refactored to support multiple packaging strategies cleanly
• Internal file scanning pipeline separated from git-based file selection logic
• CLI output consistency improved between standard scan mode and git-changes mode

Developer Notes

• Introduced new contextzip.git module for git integration and porcelain parsing
• Added resolve_files_from_git() pipeline for pre-selected file resolution
• Added structured GitChanges and GitError models for cleaner error handling and extensibility

---

[0.2.1] — 2026-05-16

Fixed

• --exclude / -e patterns containing Windows-style paths (e.g. .\CHANGELOG.md) were silently ignored due to backslash and leading .\ — patterns are now normalized to forward-slash posix paths before matching
• Files passed to -e were not being excluded from the ZIP despite appearing in the exclusion list

Improved

• -e now accepts multiple space-separated patterns in a single flag invocation: -e file1 file2 folder/ in addition to the existing repeatable form -e file1 -e file2
• Folder exclusion via -e folder/* is now supported and normalized to folder/ automatically

[0.2.2] - 2026-05-16

Added

• contextzip exclude PATTERN… subcommand — exclude multiple files and folders in a single invocation without repeating -e (e.g. contextzip exclude CHANGELOG.md LICENSE .github/)
• contextzip include PATH… subcommand — symmetric counterpart to exclude; packages only the specified paths (e.g. contextzip include src/ app/)
• All modifier flags (--dry-run, --verbose, --output, --no-clipboard, --git-changes, --no-gitignore) work on both subcommands

Changed

• -e / --exclude flag on the main command restored and kept for full backwards compatibility alongside the new subcommand
• Warning messages for large files now mention both -e PATTERN and contextzip exclude PATTERN as remediation options
• _normalize_pattern promoted to a named module-level function (testable, documented); behaviour is unchanged

Fixed

• Removed the previous positional-argument approach to exclusions on the main command, which was ambiguous and undiscoverable

---

[0.2.3] — 2026-05-19

Added

• Dedicated Ruby ecosystem exclusion rules via new contextzip.rules.ruby module
• Automatic exclusion of common Ruby / Rails artifacts:
  • Bundler directories and lockfiles
  • Rails logs, temp files, assets, storage, and Spring cache
  • Built gems and compiled Ruby extensions
  • Coverage and generated documentation directories
• Additional package discovery keywords for AI tooling and code packaging:
  • claude
  • chatgpt
  • code-packaging

Changed

• Ruby project detection now uses the dedicated Ruby rule module instead of the generic base rules
• Git-based packaging now enforces base exclusion safety rules even for tracked files
• Repository governance and metadata files are now excluded by default:
  • CHANGELOG*
  • LICENSE*
  • CONTRIBUTING*
  • SECURITY.md
  • .github issue / PR templates
• .gitignore is no longer excluded automatically from packaged context archives

Fixed

• Prevented git-tracked sensitive files (e.g. .env, secrets, binaries) from being included in --git-changes mode
• Improved ZIP packaging resilience when file stat operations fail during archive creation
• Fixed progress tracking during ZIP generation to avoid repeated filesystem stat calls
• Fixed Ruby framework detection routing so Ruby projects receive ecosystem-specific exclusion behaviour

---

[0.2.4] — 2026-05-20

Added

• Persistent .contextzip/ workspace directory for generated context packages

• Automatic output naming based on packaging mode:

  • codebase.zip for standard packaging
  • changes.zip for --git-changes

• Automatic .contextzip/ exclusion from packaged archives and file scanning

• Automatic .gitignore integration:

  • .contextzip/ is appended when a git repository is detected
  • missing entries are added only once

• Repository-root-aware workspace placement:

  • .contextzip/ is created at the git repository root when .git/ is detected
  • falls back to the current working directory outside git repositories

Changed

• ZIP outputs are now written to the persistent .contextzip/ workspace by default instead of temporary system directories
• Generated archives now overwrite previous outputs with the same logical name instead of creating temporary randomized filenames
• --output now bypasses all .contextzip/ workspace logic and writes directly to the user-specified destination

Improved

• Generated context packages are now easier to rediscover, reuse, and inspect across AI-assisted development workflows

• Workspace output structure now provides a foundation for future features such as:

  • timestamped package history
  • metadata indexing
  • incremental packaging
  • AI debug sessions
  • package caching

Fixed

• Prevented .contextzip/ contents from recursively packaging previous generated archives
• Improved fallback behaviour when workspace creation fails due to filesystem permissions or read-only directories

[0.3.0] — 2026-05-24

Added

• --prompt / -p flag for AI-powered file selection — describe your task in plain English and contextzip automatically selects the minimum relevant files (e.g. contextzip --prompt "Change toast color on failed login")
• Gemini 2.5 Flash Lite integration via a thin httpx client — no Google SDK required; a single lightweight POST call to the Gemini API
• First-run onboarding flow for users without an API key:
  • Displays a guided panel with a link to Google AI Studio (free, no credit card)
  • Optionally opens the browser automatically
  • Saves the key persistently on confirmation
• Persistent API key storage in the platform-appropriate config directory:
  • ~/.config/contextzip/config.json on Linux / macOS
  • %APPDATA%\contextzip\config.json on Windows
  • GEMINI_API_KEY environment variable always takes precedence
• contextzip config subcommand for key management:
  • contextzip config — show current key status and source
  • contextzip config --reset-key — clear stored key and re-run onboarding
  • contextzip config --show-key-path — print config file location
• prompt.txt is written as the first entry in every AI-generated ZIP, containing the task description, detected framework, and selected file list — so any AI tool receiving the archive immediately knows the context
• Keyword-based heuristic file scorer as a rate-limit fallback:
  • Activates only on confirmed HTTP 429 from Gemini
  • Scores files by token overlap with the prompt, directory semantics, and extension type
  • User is warned clearly that heuristic was used and Gemini results are preferred
• --prompt --dry-run combination — preview AI-selected files without creating a ZIP
• API key validation on load — keys not starting with AIza are treated as unconfigured and trigger onboarding rather than a cryptic API error
• diagnose_api_key() surfaces a human-readable explanation when a key exists but fails validation, before any API call is attempted

Changed

• create_zip() in packager.py now accepts an optional prompt_txt parameter; when provided, prompt.txt is written as the first ZIP entry
• xdg-open on Linux now runs with stdout and stderr suppressed to prevent KDE / DBus / MIME warnings from appearing over the API key prompt

Developer Notes

• Introduced contextzip.config module for cross-platform config path resolution and key read / write / delete operations
• Introduced contextzip.ai package with three modules:
  • gemini.py — Gemini API client with typed exception hierarchy (GeminiError, GeminiRateLimitError, GeminiUnavailable)
  • heuristic.py — keyword + directory scoring fallback scorer
  • selector.py — project map builder, AI orchestration, prompt.txt generation; returns (paths, prompt_txt, method) so callers know which selection path was taken
• Added httpx as a required runtime dependency

[0.3.1] — 2026-05-25

Added

• contextzip watch -- <command> subcommand — wraps any dev server or build process, buffers its output, detects errors automatically, and packages a debug-ready ZIP in a single keypress
  • Supports npm run dev, python manage.py runserver, and any process that writes errors to stdout/stderr
  • On error detection, renders an inline prompt beneath the error output: [D] package debug context [S] skip
  • Pressing D writes .contextzip/debug-context.zip immediately; the wrapped process continues running uninterrupted
  • On Ctrl+C, if no errors were packaged during the session, a final prompt offers to capture the full session output as a fallback
• debug-context.zip — flat ZIP structure produced by watch:
  • prompt.txt — auto-generated from detected framework, error type, and referenced files; no --prompt flag required from the user
  • terminal-error.txt — cleaned, noise-stripped error block and stack trace
  • source-files.zip — inner ZIP containing source files extracted from stack frames, with relative paths preserved
• contextzip/error_parser.py — full terminal output processing pipeline:
  • strip_ansi() — removes all CSI, OSC, and Fe escape sequences and bare carriage returns from raw process output
  • detect_error_block() — backward scan through the rolling buffer to find the most recent error; walks back from terminal exception lines to capture the full block including the Traceback header
  • extract_paths() — runs per-framework path regexes on the error block, resolves relative and bare filenames, and filters out stdlib, venv, and node_modules paths
  • strip_noise() — removes known noise lines (request logs, startup banners, HMR chatter) and collapses blank line runs
  • build_prompt_txt() — generates structured, AI-ready prompt.txt content
  • process_buffer() — convenience function running the full pipeline in a single call
• contextzip/watcher.py — process management and packaging:
  • Spawns child via subprocess.Popen with piped stdout/stderr; two daemon threads drain pipes concurrently into a deque(maxlen=2000) rolling buffer
  • Detection thread polls the buffer every 300ms using detect_error_block()
  • Error deduplication: MD5 of the first three non-blank lines of an error block; same signature within a session suppresses repeat prompts
  • Single-keypress D/S prompt via termios/tty raw mode on Unix and msvcrt on Windows, with an input() fallback for unusual environments
  • Child terminated cleanly on Ctrl+C: SIGTERM → 3-second grace period → SIGKILL
• contextzip/rules/errors/python.py — error detection pattern sets for Python, Django, and FastAPI:
  • 18 error start patterns covering tracebacks, syntax errors, Django/DRF exceptions, uvicorn/gunicorn ERROR logs, and pytest failures
  • 4 stack frame path patterns
  • 21 noise patterns for Django startup, request logs, pytest collection output, and pip install chatter
• contextzip/rules/errors/node.py — error detection pattern sets for Node.js, Next.js, and React:
  • 33 error start patterns covering JS runtime errors, Next.js error symbols (✗, ×, ⨯), webpack/bundler errors, npm errors, and unhandled promise rejections
  • 7 stack frame and import path patterns
  • 32 noise patterns for Next.js/Vite/CRA startup, HMR chatter, and request logs

Changed

• cli.py refactored from a single 1,100-line file into four focused modules with no behaviour changes:
  • cli.py — Click group, four subcommands, and _run(); routing only
  • cli_display.py — all Rich rendering helpers (print_detection, print_scan_summary, print_file_warnings, print_package_result, etc.)
  • cli_ai.py — AI selection orchestration and normalize_pattern
  • cli_onboard.py — Gemini API key onboarding flow and browser launch
• pyproject.toml updated to include contextzip.rules.errors in package discovery

Developer Notes

• contextzip/rules/errors/ follows the same pattern as contextzip/rules/: one file per ecosystem, each exporting ERROR_START_PATTERNS, PATH_PATTERNS, and NOISE_PATTERNS as lists of compiled re.Pattern objects — adding support for a new framework means adding one file
• _load_patterns() in error_parser.py maps ecosystem names from DetectionResult.ecosystems to rule modules; unrecognised ecosystems fall back to loading both Python and Node patterns as a best-effort default
• cli_display.py functions accept an optional con: Console parameter (defaulting to the module-level console) making every display function independently testable without monkey-patching
• No new runtime dependencies introduced — watch uses only stdlib (subprocess, threading, queue, termios/msvcrt, zipfile)

[0.3.2] — 2026-05-25

Added

• Python API — contextzip is now usable as a library, not just a CLI tool:

  • get_git_changes(path?) — returns a FileCollection of git-modified, added, and untracked files; applies base safety rules (no secrets, no binaries) automatically
  • get_files(path?, include?, exclude?, use_gitignore?) — returns a FileCollection of all project files after applying ecosystem rules, .gitignore, and any extra exclusion patterns
  • create_zip(collection, output?) — writes a FileCollection to a ZIP archive and returns a PackageResult; output path is caller-controlled
  • detect_ecosystem(path?) — returns a DetectionResult with ecosystem names, display string, and confidence level
  • FileCollection — unified return type for get_git_changes and get_files; supports iteration, len(), and bool() directly; .files is always a plain list[pathlib.Path] usable without zipping
  • PackageResult re-exported from the top-level package
  • Exception hierarchy: ContextzipError (base), NotARepositoryError, GitNotFoundError, GitCommandError, NoFilesError
  • All API functions default path to Path.cwd(), matching CLI behaviour

• contextzip/api.py — new module implementing the public API; no Click, no Rich output, no SystemExit; raises typed exceptions on failure

• contextzip/packager.py — added create_zip_silent() and _workspace_output_path_silent(): same ZIP logic as the CLI path but without a Console dependency, used internally by the API

Changed

• contextzip/__init__.py — now exports the full public API (get_git_changes, get_files, create_zip, detect_ecosystem, FileCollection, PackageResult, and all exception types) so from contextzip import ... works without knowing internal module paths

Fixed

• filters.py — symlink guard inner try block repeated the identical relative_to call that already failed in the outer block; the fallback now correctly uses the pre-resolve (original_path) to compute the relative path for symlinks that point outside the project tree
• watcher.py — _drain_output_queue silently consumed sentinel values (raw=None) with continue, making them invisible to _check_reader_done; sentinels are now put back into the queue so done flags are set correctly
• watcher.py — _read_key_windows had no timeout and would spin forever in non-interactive terminals; now uses a 60-second monotonic deadline matching the existing Unix select timeout
• packager.py — removed unused _safe_name helper
• ai/gemini.py — corrected stale docstring (gemini-2.0-flash-lite → gemini-2.5-flash-lite) to match the actual DEFAULT_MODEL constant

[0.3.3] — 2026-05-29

Fixed

• contextzip.ai and all subpackages were missing from the published wheel due to an incomplete include allowlist in pyproject.toml — replaced with auto-discovery so all current and future subpackages are captured automatically

[0.3.4] — 2026-06-21

Added

• contextzip eod — builds a paste-ready end-of-day prompt from today's Claude/ChatGPT conversation plus whatever code changed, copied straight to the clipboard. Does no summarizing itself — that's left to whichever AI tool the prompt gets pasted into

• contextzip handoff — same pipeline, aimed at continuing a project in a fresh chat after hitting a usage limit, rather than at an end-of-day report

• Per-branch commit checkpoints (.contextzip/markers/eod.json, .contextzip/markers/handoff.json) so every run only reports what's new since that command last ran on that branch — no manual bookkeeping, no date-based guessing

• Three-case code-change resolution per file, checked in priority order:

  1. Not pushed — diff against the upstream branch (covers committed- but-unpushed commits and uncommitted edits in one comparison) + the complete current file
  2. Diverged from Claude — diff against the version Claude last produced, matched by filename against a local artifacts folder + the complete current codebase file
  3. Since last run — diff against the stored marker commit, or the branch's merge-base with the default branch on a branch never tracked before + the complete current file

  Brand-new untracked files skip the diff (no baseline exists) and are included as full content only

• Best-effort fetching of Claude's artifact files for case 2, using the user's own Claude.ai session key (contextzip config --set-session-key). Explicitly degrades gracefully — a missing key, an expired cookie, or the endpoint changing shape just skips case 2 with a warning, never fails the whole eod/handoff run

• contextzip config --set-session-key / --reset-session-key — manage the Claude session key alongside the existing Gemini key; default contextzip config now shows both keys' status

• Text-clipboard support (clipboard.handle_text()) — eod/handoff produce a prompt to paste, not a file to upload, so the existing file-clipboard tiers gained a text-mode sibling (pbcopy / xclip+xsel / clip.exe)

• exports/ added to the base exclusion ruleset (next to .contextzip/) — conversation exports and fetched Claude artifacts are contextzip's own working files, never code to package or report on

Changed

• packager._ensure_gitignore() generalised to accept an arbitrary entry instead of being hardcoded to .contextzip/, so exports/ can register itself in .gitignore the same way on first use

Developer Notes

• New modules: markers.py (per-branch checkpoint storage), code_changes.py (the three-case decision tree), claude_export.py (locating the latest export, inline-vs-attach sizing), claude_artifacts.py (best-effort artifact fetching, rewritten on httpx for consistency with ai/gemini.py rather than introducing a second HTTP stack), brief.py (orchestrates eod/handoff end to end)
• git.py gained branch-aware primitives: get_current_branch(), get_default_branch() (origin/HEAD detection with fallback chain), get_merge_base(), is_ancestor(), get_upstream_branch(), diff_against() (working-tree-aware — captures committed-since-baseline and uncommitted changes in one diff), changed_files_against()
• A stored marker that's no longer an ancestor of HEAD (e.g. after a rebase) is detected via is_ancestor() and silently recomputed from the branch's merge-base, with a warning surfaced to the user rather than producing a nonsense diff
• claude_artifacts.py matches Claude's flat artifact filenames against the codebase by basename only, since artifacts carry no knowledge of the repo's directory structure; ambiguous matches (same filename in multiple directories) pick the most recently modified candidate and surface a warning rather than guessing silently

[0.3.5] — 2026-08-09

Added

• Monorepo-aware framework detection. Detection now scans project_dir plus a shallow, bounded subdirectory tree (2 levels deep), instead of only checking the project root. A layout like root/frontend/package.json + root/backend/requirements.txt now correctly detects both Node.js and Python and stacks both exclusion rule sets — previously this fell through to only the universal base rules, leaving node_modules/, .venv/, and similar unexcluded except by whatever .gitignore happened to catch. The scan skips dependency/build directories (node_modules/, .venv/, .git/, etc.) so it never walks into a large dependency tree. Detection output now shows which subdirectory each ecosystem was found in, e.g. Next.js (frontend/) + FastAPI (backend/).
• Configurable .contextzip/ workspace location, resolved in priority order: CONTEXTZIP_WORKSPACE_LOCATION env var → project config (.contextzip.json, committed, team-shared) → personal config (per-machine, via contextzip config --set-workspace) → default (git-root, unchanged behavior). Accepts "git-root", "cwd", or an explicit path.
• contextzip config --set-workspace <location> / --reset-workspace — set or clear your personal workspace location default.
• contextzip config (no flags) now shows the resolved workspace location and which config tier it came from.
• New .contextzip.json project config file (optional, git-committed) for pinning workspace_location for every contributor.

Fixed

• .gitignore registration for the .contextzip/ entry now correctly derives the git root independently, rather than assuming it's always the workspace directory's parent — a latent bug this release's configurable workspace location would otherwise have triggered (it only skips writing when the workspace is genuinely outside the repo, e.g. an absolute custom path).

Removed

• eod and handoff commands, along with the Claude.ai session-key config option they depended on. Both relied on an undocumented Claude.ai endpoint to fetch previously-produced artifacts, which was never a stable foundation for a shipped feature.
  • Removed modules: brief.py, claude_export.py, claude_artifacts.py, code_changes.py, markers.py
  • Removed CLI options: contextzip config --set-session-key / --reset-session-key
  • Removed config keys: claude_session_key
  • Also removed now-unused branch-diffing internals from git.py (get_current_branch, get_upstream_branch, diff_against, and related helpers) that existed solely to support eod/handoff
  • If you relied on eod/handoff, there's no replacement in this release — conversation exports and per-file diff prompts will need to be built by hand, or via your own script against contextzip --git-changes.

[0.3.6] — 2026-08-23

Added

• cz CLI alias. cz can now be used as a shorthand for contextzip, with identical commands and behavior.
• .contextzip/ project workspace. Contextzip now uses a dedicated workspace containing config.json and an output/ directory for generated context archives.
• Project configuration via .contextzip/config.json. Added persistent project-level preferences for always_include, always_exclude, and AI-assisted context selection.
• AI configuration. Added configurable AI preferences including whether AI selection is enabled, the provider to use, and the maximum number of files AI can select.
• Configured file selection. always_include and always_exclude preferences are now applied during context generation.
• Configurable AI file limits. The configured ai.max_files value is now respected by the AI selection pipeline.
• Legacy .contextzip.json deprecation. Project configuration is moving from .contextzip.json to .contextzip/config.json.

Changed

• Generated context archives are now organized under .contextzip/output/.
• Contextzip configuration is now resolved through the new project configuration system while preserving the existing personal configuration for machine-specific settings and credentials.

[0.3.7] — 2026-08-23

Added

• Visual config UI. contextzip config --ui opens a local browser tab for setting always_include/always_exclude by clicking through your actual file tree, with live file counts and packed size as you go.
• Exclude suggestions. The config UI automatically flags common non-code file types not already covered by the default exclusion rules — PDFs, office docs, fonts, media, design files, and anything over 1MB — as one-click "exclude all" chips.
• First-run offer. A project with no config at all offers to launch the visual config UI on its first run. Declining is remembered — it won't ask again unless you run contextzip config --ui yourself.

Changed

• Declining the visual config UI's first-run offer is now stored in the personal config (~/.config/contextzip/config.json) so it's asked at most once per machine.

Security

• The local config UI binds to 127.0.0.1 only and requires a random per-session token on every request (matching Jupyter Notebook's approach) — nothing about the project's file structure leaves the machine.

[0.3.8] — 2026-08-24

Added

• apply-zip command. Closes the other half of the round trip: takes a ZIP handed back by an AI tool and writes it into the project. Every file is classified — new, modified, unchanged, drifted (changed or removed locally since it was sent), or untracked (no baseline to compare against) — before anything is written.
• Local sidecar manifest. Every ZIP contextzip creates now also gets a <name>.manifest.json written next to it in .contextzip/output/ — a hash of each included file, taken at zip-time. It's written to disk only, never inside the ZIP, so it's never uploaded and never visible to whatever AI tool the ZIP is pasted into.
• .contextzip/inbox/. Drop an AI-returned ZIP here and apply-zip finds it automatically; passing a path explicitly always overrides auto-detection. Multiple ZIPs in the inbox are listed and you're asked to be explicit rather than guessed at.
• Automatic backups. Every file apply-zip is about to overwrite is copied first into .contextzip/backups/<timestamp>/.
• apply_zip() Python API function, mirroring create_zip()'s plain-function, no-Click, no-Rich style.
• New --dry-run, --verbose, --yes, and --manifest flags on apply-zip.

Changed

• Consumed ZIPs are moved to .contextzip/inbox/applied/<timestamp>-name.zip rather than deleted, so they double as an audit trail alongside the backups.

Security

• apply-zip validates every ZIP entry's resolved path before extracting anything, refusing archives with path-traversal entries (../) that would write outside the project tree — no override flag, no exceptions.

Notes

• v1 of apply-zip only adds and modifies files — deletions are never inferred from a ZIP's contents. A file missing from a returned ZIP is left alone.

[0.3.9] - 2026-08-26

Security

• Config file permissions (~/.config/contextzip/config.json) are now locked to 0600 on every write, not just on creation. Previously a pre-existing config file from an older install (or one created under a permissive umask) could stay group/world-readable indefinitely — the file can hold a Gemini API key in plaintext, so this closes a real local exposure on shared machines.
• The local config UI (contextzip config --ui) now compares the request token with hmac.compare_digest instead of ==, removing a timing-based side channel on the single-use session token.
• Substantially expanded the list of secret/credential files that are always excluded from packaged zips, regardless of framework: SSH private keys (id_rsa, id_ed25519, id_ecdsa, id_dsa), keystores (.jks, .keystore, .p12, .pkcs12, .ppk), CLI/package-manager credential files (.npmrc, .netrc, .pypirc, .pgpass, .dockercfg, Docker config.json), cloud provider credentials (.aws/credentials, .aws/config, *serviceaccount*.json, kubeconfig), and Terraform state (*.tfstate*, .terraform/), which routinely contains plaintext secrets even for "just infra" resources.

Fixed

• apply-zip now detects when every entry in an AI-returned zip is nested under one incidental top-level folder — the shape produced by zip -r out.zip myfolder or GitHub's "Download ZIP" — and strips it automatically when doing so clearly improves the match against the project's manifest, so files land at their real project-relative paths instead of recreating a wrapper folder.
• apply-zip now warns before applying a zip whose structure barely matches the project at all (under 10% of paths found in the manifest, even after wrapper stripping), instead of silently treating every file as brand new.

Changed

• .contextzip/inbox/applied/ now keeps only the most recently applied zip instead of accumulating one per session.

Internal

• Removed an unused GitChanges import in api.py.
• Clarified the .gitignore comment describing what lives under .contextzip/ and confirming the Gemini API key is never stored there.

[0.4.0] - 2026-08-26

Added

• Config UI redesign. contextzip config --ui is now a five-tab interface (Files, AI selection, Workspace, Advanced, Raw JSON) instead of a single always_include/always_exclude pane. The file tree and pattern list now render as a diff — a solid gutter and +/- marks on every row and pattern, matching the git-diff mental model the rest of the tool already uses. Still a single dependency-free HTML string with no build step, no Node.js, and no network calls from the page itself.
• New, persisted project preferences (.contextzip/config.json), editable from the new UI tabs or by hand:
  • ai.prompt_template — house conventions prepended to every generated prompt.txt, ahead of the task description.
  • limits.max_file_size_mb — replaces the fixed 1 MB "large file" threshold with a per-project value.
  • limits.redact_secrets — reserved for a future scrub of secret-shaped values inside otherwise-included files. Persisted and editable now; not yet enforced during packaging (see Known limitations below).
  • applied_zip_retention — how many past apply-zip archives to keep in .contextzip/inbox/applied/ before pruning (was hardcoded to 1).
  • webui.auto_open / webui.port — skip auto-launching a browser tab and/or pin the config UI to a fixed local port instead of a random free one.

Changed

• contextzip config --ui's local server now exposes the full project config through /api/state and accepts the full config through /api/save, instead of only always_include/always_exclude.
• Large-file warnings in normal CLI output now reflect the project's configured threshold instead of a fixed 1 MB.

Known limitations

• limits.redact_secrets can be toggled on and is saved correctly, but nothing reads it yet — no redaction currently happens during packaging. Flagging this explicitly so it isn't relied on before the enforcement lands.

[0.4.1] — 2026-08-30

Added

• Per-mode output folders. .contextzip/output/ now splits by run mode instead of one flat folder: output/codebase/ (plain contextzip), output/git-changes/ (--git-changes), output/prompt/ (--prompt, filename vibe.zip), and output/watch/ (debug-context.zip). The folder name matches the flag that produced it, so it's unambiguous which zip is which without having to remember which command you last ran.
• Automatic workspace cleanup. After every successful command (main run and apply-zip), .contextzip/ is pruned automatically: old zip/manifest/report sets per mode folder, old backups/ folders, and old archived zips in inbox/applied/ are all deleted down to the most recent one — no confirmation, no separate command. New project settings: cleanup.enabled (default true) and cleanup.keep_recent (default 1).
• limits.redact_secrets is now enforced. Previously persisted in config and editable in the config UI but silently ignored by packaging (flagged as a known limitation in 0.4.0). Now, text files already going into the archive are scanned for secret-shaped values — AWS/Google/GitHub/Slack/Stripe/Anthropic/OpenAI keys, JWTs, private-key blocks, and a conservative generic key/secret/token/password assignment pattern — and matched values are replaced with [REDACTED] before writing. Only scans files already known to be text and within limits.max_file_size_mb; binary and oversized files are never touched. Redactions are surfaced in a terminal warning and itemized in the .report.txt file, never applied silently.
• .report.txt files. Every packaging run (and apply-zip run) writes a full plain-text report alongside its ZIP — codebase.report.txt / codebase.apply-report.txt — containing everything that used to fill the terminal: the excluded-directory breakdown, the complete included-file list, per-file warnings, and (when applicable) which files had secrets redacted.

Changed

• Terminal output redesigned. Every command now prints a compact, one-line-per-step log (✓ / ! / ✗) instead of boxed Rich panels and tables — matching the register of tools like pnpm, vite, and ruff. A full-width terminal is no longer needed to read a run.
• The "Scanned" and "Packed" steps are now combined into a single line, e.g. Scanned 1518 files & Packed 42 files · 42 included (394 KB), 1476 excluded, 394 KB → 126 KB, ↓68% smaller.
• Removed the contextzip vX.X.X / project-path banner from every command — output now starts directly at the first real step.
• apply-zip's manifest auto-detection now searches recursively across the new per-mode output subfolders, so it still finds the right manifest under the new layout.
• Full file listings and excluded-directory breakdowns moved out of default terminal output — shown under -v/--verbose, or in the full report file.

Removed

• .gitignore patterns applied is no longer printed (no signal, dead weight).
• Opening the output folder no longer prints a confirmation line.

[0.4.2] — 2026-08-31

Added

• limits.redact_secrets is now enforced. Previously persisted in config and editable in the config UI but silently ignored by packaging (flagged as a known limitation in 0.4.0). Now, text files already going into the archive are scanned for secret-shaped values — AWS/Google/GitHub/Slack/Stripe/Anthropic/OpenAI keys, JWTs, private-key blocks, and a conservative generic key/secret/token/password assignment pattern — and matched values are replaced with [REDACTED] before writing. Only scans files already known to be text and within limits.max_file_size_mb; binary and oversized files are never touched. Surfaced in a terminal warning and itemized in the .report.txt file, never applied silently.

Changed

• .contextzip/ is now fully local by default, no exceptions. config.json was previously carved out of the workspace's .gitignore so it stayed trackable automatically. It's now ignored like everything else in the workspace — nothing under .contextzip/ is ever pushed unless you explicitly opt in with git add -f .contextzip/config.json.

[0.4.3] — 2026-09-20

Added

• contextzip update / cz update — checks PyPI for a newer release and updates in place, so you no longer have to remember pip install --upgrade contextzip after every one of your own releases. Detects whether this install came from pip, pipx, or uv tool and re-invokes the matching one against the currently running interpreter — never a possibly-different pip off PATH. --check reports the current/latest version without installing; --yes skips the confirmation prompt.

Fixed

• apply-zip no longer dumps a nested copy of the project when the returned zip wraps everything in a folder named after the project itself (e.g. project/, or a folder matching the package's own name). The wrapper-stripping logic added in an earlier release had a guard that skipped stripping whenever a real folder with that name already existed in the project — which is exactly what happens when an AI tool names the wrapper after the project. Now, whenever a manifest is available, the decision is made by comparing how well paths actually match the manifest with vs. without stripping, not by guessing from what's on disk.
• contextzip config --ui now closes its browser tab automatically after a successful save and hands control back to the terminal, instead of only telling you to close it yourself.

[0.4.4] — 2026-09-21

Added

• contextzip update / cz update — checks PyPI for a newer release and updates in place, so you no longer have to remember pip install --upgrade contextzip after every one of your own releases. Detects whether this install came from pip, pipx, or uv tool and re-invokes the matching one against the currently running interpreter — never a possibly-different pip off PATH. --check reports the current/latest version without installing; --yes skips the confirmation prompt.
• .contextzip/inbox/README.txt is now created automatically the moment the .contextzip/ workspace is first set up (and self-heals if it's ever missing) — a plain-language explainer of what the inbox is for, how to drop a zip into it, and what happens afterward, so the folder isn't a silent, unexplained thing you have to already know about.

Fixed

• apply-zip no longer dumps a nested copy of the project when the returned zip wraps everything in a folder named after the project itself (e.g. project/, or a folder matching the package's own name). The wrapper-stripping logic had a guard that skipped stripping whenever a real folder with that name already existed in the project — exactly what happens when an AI tool names the wrapper after the project. The decision is now made by comparing manifest match-rate with vs. without stripping instead of guessing from what's on disk.
• apply-zip no longer leaves a zip's wrapper folder in place when the freshest manifest is scoped to a small subset of the project (most commonly .contextzip/output/git-changes/changes.manifest.json, which only ever tracked files touched in one prior git-diff run). A zip of otherwise-legitimate new files matched that manifest at 0% whether or not the wrapper was stripped, so the fix above's manifest-comparison saw a tie and played it safe by leaving the wrapper in place. It now falls back to comparing the zip's top-level folders against the project's actual current directory layout (app/, components/, lib/, etc.) whenever the manifest signal ties, and strips confidently when that layout lines up.
• apply-zip <path> no longer drops its .apply-report.txt next to wherever the zip happens to be (e.g. the project root) when given an explicit path outside .contextzip/inbox/. The report now always writes to .contextzip/inbox/, matching where it's always shown up when a zip is auto-detected from there. The zip itself is still left in place, unmoved, when given an explicit path.
• Auto-cleanup now actually removes old *.apply-report.txt files sitting loose in .contextzip/inbox/ instead of leaving every one ever written to pile up indefinitely. Keeps only the most recent cleanup.keep_recent (default 1), same rule as everything else it prunes.

Removed

• The per-apply backup mechanism (.contextzip/backups/<timestamp>/) is gone entirely. apply-zip no longer copies overwritten files there before writing; the manifest-diff confirmation prompt and the fact that the applied zip is always archived (not deleted) remain the safety net. Auto-cleanup now also wipes any leftover .contextzip/backups/ folder outright on projects upgrading from an older contextzip version, rather than just pruning it down over time.

[0.4.5] — 2026-09-27

Fixed

• --git-changes was silently ignoring .contextzip/config.json. _run() only built the exclude/force-include rules from always_include/always_exclude — and folded in the current run's --include/--exclude — inside the standard-scan branch. resolve_files_from_git() had no parameters to receive any of that, so --git-changes packaged exactly what git status reported, no more and no less, regardless of what the config or the flags for that run said. A project could have *.snap in always_exclude and still see it land in every --git-changes zip. resolve_files_from_git() now accepts extra_exclude, include_only, and force_include and applies them with the same precedence resolve_files() already used — force-include first, then the safety floor plus excludes, then include-only — so both modes resolve a project's files identically.
• contextzip watch could package a file straight out of always_exclude if a stack trace happened to name it — the debug-zip path built its file list from extract_paths() alone, with no awareness that a project had already asked for that file to be left out everywhere else. process_buffer() now filters extracted paths through the project's always_exclude/always_include before they reach prompt.txt or source-files.zip.
• The Python API (get_files(), get_git_changes()) never read .contextzip/config.json at all — library callers got only the built-in rules plus whatever include/exclude they passed in explicitly, even on a project with a committed config the CLI already honored. Both functions now load project config by default and apply always_include/always_exclude/limits.max_file_size_mb the same way the CLI does. This is a behavior change for existing callers relying on the old, config-blind result — pass the new use_project_config=False to keep it.

Changed

• Empty-result messages now read "Nothing to compress in this one" instead of "Nothing to package" — shown both when --git-changes finds a clean working tree and when a standard scan excludes every file.
• Unknown-ecosystem detection now points you to request your language. When contextzip can't identify a project's stack, the detection line is followed by Don't see your language? Add it here: <link> instead of leaving that one line as the only output.

[0.4.6] — 2026-10-07

Fixed

• The ZIP was never actually copied to the clipboard on Windows or Linux. On Windows there was no clipboard step at all: contextzip opened Explorer and left you to copy the file yourself. On Linux it wrote the ZIP's raw bytes to the clipboard as application/zip, which browsers ignore on paste, and only when xclip happened to be installed. Both platforms now put a real file reference on the clipboard, the same thing Ctrl+C on a file does, so you can paste the ZIP straight into Claude, ChatGPT or any other upload box.
  • Windows: sets CF_HDROP directly through the Win32 API (plus "Preferred DropEffect" so Explorer pastes a copy), then reads it back to confirm. If that fails it falls back to PowerShell Set-Clipboard -LiteralPath, and only then to opening Explorer.
  • Linux: sets text/uri-list, the target browsers read when you paste a file. Order of attempts: wl-copy (only if already installed, on Wayland), a new built-in X11 clipboard owner, then xclip (only if already installed). Every attempt is read back before contextzip reports success.
• A failed clipboard copy was silent. When contextzip fell back to opening the folder it printed nothing, so you couldn't tell the copy had failed. It now says so and, when it knows, how to fix it.
• Explorer fallback lost the selection on paths with spaces or commas. explorer /select, is now passed as one correctly quoted command line.
• xclip / wl-copy are run with their output discarded rather than captured, so they can no longer block the CLI.

Added

• Built-in X11 clipboard owner — nothing to install on Linux. pip install contextzip is now enough: contextzip speaks the small part of the X11 protocol needed to own the clipboard itself (standard library only, no new dependency). It offers text/uri-list, x-special/gnome-copied-files and application/x-kde-cutselection, so browsers and GNOME/KDE file managers can all paste the file. A tiny background process holds the clipboard and exits on its own as soon as something else is copied.

Changed

• Headless and SSH Linux sessions (no display) now print the ZIP's path instead of trying to launch xdg-open.
• ClipboardResult has a new optional hint field explaining why a copy fell back and how to fix it.
• macOS: quotes and backslashes in the ZIP's path are now escaped in the Finder script. Behaviour is otherwise unchanged.

Known limitations

• Pasting on Wayland-native browsers relies on XWayland passing the clipboard through, which GNOME and KDE do. A Wayland-only setup without XWayland needs wl-clipboard (wl-copy) installed; otherwise contextzip opens the folder and tells you what to install.