Previously, get_or_create_key() unconditionally wrote the encryption key
to ~/.config/gws/.encryption_key on first run, even when the OS keyring
was available. This left the key material on disk as a plain file,
making credentials portable by copying the config directory.
Changes:
- Extract save_key_file() helper to deduplicate file-writing logic
- On keyring read success: delete stale .encryption_key (migration)
- On NoEntry + existing file: migrate key into keyring, then delete file
- On NoEntry + new key: try keyring first, only write file as fallback
Fixes#344
Co-authored-by: jpoehnelt-bot <jpoehnelt-bot@users.noreply.github.com>
* feat(gmail): add +reply, +reply-all, and +forward helper commands
Add first-class reply and forward support to the Gmail helpers,
addressing the gap described in #88. These commands handle the
complex RFC 2822 threading mechanics (In-Reply-To, References,
threadId) that agents and CLI users struggle with today.
New commands:
- +reply: reply to a message with automatic threading
- +reply-all: reply to all recipients with --remove/--cc support
- +forward: forward a message with quoted original content
* fix(gmail): encode message_id in URL path and fix auth signature
- Use crate::validate::encode_path_segment() on message_id in
fetch_message_metadata URL construction per AGENTS.md rules
- Update auth::get_token calls to pass None for the new account
parameter added on main
* refactor(gmail): extract send_raw_email and deduplicate handlers
- Add send_raw_email() to mod.rs: shared encode→json→auth→execute
pattern for sending raw RFC 2822 messages via users.messages.send
- Simplify handle_reply: delegate send logic to send_raw_email
- Simplify handle_forward: delegate send logic to send_raw_email
Addresses code duplication feedback from PR review.
* fix(gmail): register --dry-run flag on reply/forward commands
The handlers read matches.get_flag("dry-run") but the flag was missing
from the clap command definitions, so it always returned false. Now
dry-run works for +reply, +reply-all, and +forward.
* chore: add changeset for gmail reply/forward feature
* style: apply cargo fmt formatting
* fix(gmail): register --dry-run flag on +send command
Same class of bug fixed for +reply/+reply-all/+forward — the handler
reads matches.get_flag("dry-run") but the arg was not registered.
* fix(gmail): honor Reply-To header and use exact address matching
- Prefer Reply-To over From when selecting reply recipients, fixing
incorrect routing for mailing lists and support systems
- Use exact email address comparison instead of substring matching
for --remove filtering and sender deduplication, preventing
unintended recipient removal (e.g. ann@ no longer drops joann@)
* test(gmail): add comprehensive coverage for reply address handling
- extract_email: malformed input (no closing bracket), empty string,
whitespace-only
- build_reply_all_recipients: display-name sender exclusion,
--remove with display name, extra --cc, CC becomes None when all
filtered, case-insensitive sender exclusion
* Improves reply-all recipient deduplication
Corrects how `build_reply_all_recipients` handles multi-address `Reply-To` headers.
Previously, only the first address from `Reply-To` was used for deduplication, leading to potential redundancy by including those addresses in the `Cc` field.
The updated logic now parses all addresses in `Reply-To`, ensuring they are fully moved to the `To` field and properly excluded from `Cc`.
* style(gmail): add missing Apache 2.0 copyright headers
reply.rs and forward.rs were missing the copyright header that all
other source files in the repo include.
* fix(gmail): use try_get_one for optional --remove arg in +reply
parse_reply_args used get_one("remove") which panics when called
from +reply (which does not register --remove). Switch to
try_get_one to safely return None for unregistered args.
* feat(gmail): support --dry-run without auth for reply/forward commands
Skip auth and message fetch when --dry-run is set by using placeholder
OriginalMessage data. This lets users preview the request structure
without needing credentials.
* fix(gmail): use RFC-aware mailbox list parsing for recipient splitting
Replace naive comma-split with split_mailbox_list that respects
quoted strings, so display names containing commas like
"Doe, John" <john@example.com> are handled correctly in reply-all
recipient parsing, deduplication, and --remove filtering.
* fix(gmail): handle escaped quotes in mailbox list splitting
split_mailbox_list toggled quote state on every `"` without accounting
for backslash-escaped quotes (`\"`), causing display names like
`"Doe \"JD, Sr\""` to split incorrectly at interior commas.
Track `prev_backslash` so `\"` inside quoted strings is treated as a
literal quote character rather than a delimiter toggle. Double
backslashes (`\\`) are handled correctly as well.
* fix(gmail): address PR review feedback for reply/forward helpers
- Use reqwest .query() for metadata params per AGENTS.md convention
- Add MIME-Version and Content-Type headers to raw messages
- Add --from flag to +reply, +reply-all, +forward for send-as/alias
- Narrow ReplyConfig/ForwardConfig visibility to pub(super)
- Refactor create_reply_raw_message args into ReplyEnvelope struct
* fix(gmail): address review feedback for reply/forward helpers
- Exclude authenticated user's own email from reply-all CC by
fetching user profile via Gmail API
- Use format=full to extract full plain-text body instead of
truncated snippet for quoting and forwarding
- Deduplicate CC addresses using a HashSet
- Reuse auth token from message fetch in send_raw_email to
eliminate double auth round-trip
- Propagate auth errors in send_raw_email instead of silently
falling back to unauthenticated requests
- Use consistent CRLF line endings in quoted and forwarded
message bodies per RFC 2822
* fix(gmail): Gmail reply and forward helpers
* fix(gmail): refactor shared reply-forward helpers
* Preserve repeated Gmail address headers
* chore: regenerate skills [skip ci]
---------
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
* fix: respect account selection in MCP server and CLI --account flag (#221, #181)
MCP server now reads GOOGLE_WORKSPACE_CLI_ACCOUNT env var and passes it
to get_token instead of always using None (default account).
CLI filter_args_for_subcommand now dynamically locates the service name
instead of hardcoding skip(2), fixing --account before service name
causing unrecognized subcommand errors.
* fix: skip --api-version in first_arg detection
The first_arg loop only skipped --account but not --api-version,
so `gws --api-version v3 drive ...` would misidentify --api-version
as the service name. Now both global flags are consistently skipped.
* fix(mcp): conditionally include body/upload in full-mode tool schemas and drop empty body on execution
Full-mode tool schemas now only include `body` when the Discovery Document
method defines a request body, and `upload` when `supportsMediaUpload` is
true. This prevents LLMs from hallucinating these fields on GET-only methods.
Additionally, empty body objects (`{}`) are filtered out before execution
in both compact and full modes, and empty upload strings are ignored. LLMs
commonly send "body": {} on read-only methods, which causes Google APIs to
return HTTP 400.
* style: cargo fmt and add changeset for MCP tool schema fix
* fix(mcp): conditionally include page_all only for paginated methods
Only include the page_all property in full-mode tool schemas when the
method has a pageToken parameter, preventing LLMs from attempting
pagination on non-paginable methods.
* docs: update changeset to include page_all conditional change
Two call sites in auth_commands.rs and setup_tui.rs used
.strip_suffix(".readonly").unwrap(), which panics if a scope URL
flagged as is_readonly doesn't end with ".readonly".
Replace with .unwrap_or() to gracefully fall back to the original
URL rather than crashing on inconsistent discovery data.
* fix: select broadest scope instead of all method scopes
Discovery Documents list method scopes as alternatives (any one grants
access), but passing all of them to yup_oauth2 caused Google to include
restrictive scopes like gmail.metadata in the token. The API then
enforced that scope's restrictions, blocking query parameters like `q`.
Select only the first (broadest) scope from the method's scope list.
* fix: filter gmail.metadata from login scopes and remove token cache superset fallback
gmail.metadata restricts API behavior (blocks `q` parameter) even when
broader scopes are present in the token. Filter it out during login when
broader Gmail scopes like gmail.modify or mail.google.com are selected.
Also remove the superset fallback in token storage to prevent stale
all-scopes tokens from being reused when a narrower scope is requested.
* fix: apply select_scope to MCP server code path
The MCP server had the same bug as the CLI — passing all method scopes
to get_token. Use select_scope to pick only the broadest scope.
* style: fix cargo fmt formatting issues
---------
Co-authored-by: Justin Poehnelt <jpoehnelt@google.com>
* fix(client): send x-goog-user-project header from ADC quota project
When using Application Default Credentials with a quota_project_id set,
API requests failed with 403 because the quota project header was never
sent. Read quota_project_id from ADC and set it as a default header.
* Update src/auth.rs
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
---------
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Replaced silent `let _ =` on set_permissions calls in save_encrypted
with eprintln! warnings so users are aware if their credential files
end up with insecure permissions. Also log keyring access failures
instead of silently falling through to file storage.
load_from_disk used four nested if-let-Ok blocks that silently
returned an empty HashMap on any failure. When the encryption key
rotated or the cache file was corrupted, tokens silently stopped
loading and users were forced to re-authenticate with no explanation.
Replace with explicit match arms that log specific warnings to
stderr for each failure mode:
- Decryption failure (key changed, corrupted data)
- Invalid UTF-8 in decrypted data
- JSON deserialization failure
File-not-found is still silent since that's normal on first run.
- Update stale docstring on resolve_account to reflect fallthrough behavior
- Add breadcrumb comment on string-based error matching in main.rs
- Move identity scope injection before authenticator build for readability
Co-authored-by: jpoehnelt-bot <jpoehnelt-bot@users.noreply.github.com>
Three related bugs caused all API calls to fail with "Access denied.
No credentials provided" even after successful `gws auth login`:
1. resolve_account() rejected valid credentials.enc as "legacy" when
accounts.json was absent, instead of falling through to use them.
2. main.rs silently swallowed all auth errors (Err(_) => None),
masking real failures behind a generic "no credentials" message.
3. auth login didn't include openid/email scopes, so
fetch_userinfo_email() couldn't identify the user, causing
credentials to be saved without an accounts.json entry.
{attribution.commit: ""}
* fix: isolate flaky auth tests from host ADC credentials
Fixes#206
Both test_load_credentials_no_options and
test_get_token_env_var_empty_falls_through now override HOME to a
temp dir and clear GOOGLE_APPLICATION_CREDENTIALS, preventing the
well-known ADC path from matching on CI runners that have gcloud
credentials.
* refactor: use RAII EnvVarGuard for panic-safe env var cleanup in tests
- Introduce EnvVarGuard struct that saves/restores env vars on Drop
- Replace all manual save/restore patterns in auth tests
- Fix bug where test_get_token_env_var_empty_falls_through did not
restore GOOGLE_WORKSPACE_CLI_TOKEN
- Ensures cleanup runs even if a test panics
* fix: use var_os/OsString in EnvVarGuard for non-UTF-8 safety
---------
Co-authored-by: jpoehnelt-bot <jpoehnelt-bot@users.noreply.github.com>
* fix(executor): add Content-Length: 0 header for body-less POST/PUT/PATCH requests
Google API servers return HTTP 411 (Length Required) when a POST request
is sent without a Content-Length header, even if there is no body.
This affects all Discovery API methods where httpMethod is POST but no
requestBody is defined (e.g. gmail users.messages.trash).
Fixes#182
* chore: add changeset for content-length fix
* feat: add Application Default Credentials (ADC) support (#103)
Extends the credential chain in get_token() to include ADC as a 4th source:
1. GOOGLE_WORKSPACE_CLI_CREDENTIALS_FILE env var
2. Encrypted credentials (~/.config/gws/credentials.enc)
3. Plaintext credentials (~/.config/gws/credentials.json)
4. ADC — GOOGLE_APPLICATION_CREDENTIALS env var, then
~/.config/gcloud/application_default_credentials.json
Both authorized_user and service_account ADC formats are detected via the
'type' field and parsed accordingly. This means users can authenticate with:
gcloud auth application-default login --client-id-file=client_secret.json
and gws will automatically pick up those credentials.
Closes#103
Co-Authored-By: Claude <noreply@anthropic.com>
* fix(auth): address review feedback on ADC support
- Extract duplicated JSON credential parsing into parse_credential_file()
helper to reduce duplication between GOOGLE_WORKSPACE_CLI_CREDENTIALS_FILE
and ADC code paths; uses serde_json::from_value to avoid second string parse
- Fix well-known ADC path on macOS: dirs::config_dir() returns
~/Library/Application Support on macOS, not ~/.config; use
dirs::home_dir().join('.config/gcloud/...') instead
- Hard-error when GOOGLE_APPLICATION_CREDENTIALS points to a missing file
(was: silently fall through to 'No credentials found')
- Add test_load_credentials_adc_env_var_service_account covering service
account credentials loaded via GOOGLE_APPLICATION_CREDENTIALS
- Remove unnecessary unsafe blocks from env var tests (set_var/remove_var
are not unsafe functions; thread safety is already handled by serial_test)
- Update changeset to include GOOGLE_WORKSPACE_CLI_TOKEN at top of lookup
order and clarify ADC fallback behaviour
Addresses review feedback from jpoehnelt on #125.
Co-Authored-By: Claude <noreply@anthropic.com>
---------
Co-authored-by: Claude <noreply@anthropic.com>
* fix(setup): drain stale keypresses between TUI screen transitions
The picker's vim-style j/k navigation can leave buffered keypresses in
crossterm's event queue that leak into the next input field, corrupting
pre-populated values like client_id.
* refactor: extract drain logic into helper, add changeset
Replace bg(Color::DarkGray) with Modifier::REVERSED in highlight_style
so the selected-row highlight adapts to both light and dark terminal themes.
Co-authored-by: jpoehnelt-bot <jpoehnelt-bot@users.noreply.github.com>
Admin-only scopes (apps.*, cloud-identity.*, ediscovery, directory.readonly,
groups) require Workspace domain-admin access and cannot be granted to personal
@gmail.com accounts — Google returns 400 invalid_scope when they're included.
Changes:
- Add is_workspace_admin_scope() helper (mirrors is_app_only_scope())
to identify scopes that fail for personal Google accounts
- Exclude these scopes from the template_selects of the 'Recommended' preset
in run_discovery_scope_picker()
- Exclude them from the resolved scope list when the Recommended template
is confirmed
- Add 8 unit tests covering the new helper
Workspace admins can still access these scopes via 'Full Access' template
or by selecting them individually in the picker.
Note: this is complementary to PR #108 which filters alertcenter scopes
at the API-discovery level. This PR handles the broader set at the
recommendation layer.
Addresses #119 (Bug 1: admin scopes in Recommended preset)
Co-authored-by: Claude <noreply@anthropic.com>
Previously used dirs::config_dir() which resolves to different paths per
OS (~/Library/Application Support/gws on macOS, %APPDATA%\gws on Windows),
contradicting the documented ~/.config/gws/ path and causing users to place
config files in the wrong location (ref #119).
Now uses ~/.config/gws/ everywhere with a fallback to the legacy OS-specific
path for existing installs. Also consolidates duplicated dirs::config_dir()
calls in auth.rs and discovery.rs to use the central config_dir() helper.
When `gcloud projects list` times out (10s limit) for users with many
projects, the picker now includes a '⌨ Enter project ID manually'
option so they can type a known project ID instead of waiting or failing.
- Add '⌨ Enter project ID manually' item to project picker
- Handle the new item by prompting for input and calling set_gcloud_project
- Add EnterProjectId variant to SetupAction (tests)
- Add test_project_select_enter_manually unit test
Fixes#116
Co-authored-by: Claude <noreply@anthropic.com>
The reject_control_chars helper rejected bytes 0x00-0x1F but allowed
the DEL character (0x7F), which is also an ASCII control character.
This could allow malformed input from LLM agents to bypass validation.
* fix: use gcloud.cmd on Windows and show platform-correct config paths
On Windows, Google Cloud SDK installs `gcloud.cmd` (not `gcloud.exe`).
Rust's `Command::new("gcloud")` does not search PATHEXT, so all gcloud
invocations failed silently. Add `gcloud_bin()` helper that returns
`gcloud.cmd` on Windows and `gcloud` elsewhere.
Also replace hardcoded `~/.config/gws/` paths in error messages with
the actual platform-resolved path (`%APPDATA%\gws\` on Windows).
* chore: add changeset for Windows gcloud compat fix
* fix: filter alertcenter scopes from user OAuth login flow
The `apps.alerts` scope is restricted to service accounts with
domain-wide delegation and fails with `400 invalid_scope` when
used in the standard 3-legged OAuth consent flow. Filter it out
alongside the existing chat.app/chat.bot/keep exclusions.
Fixes#73
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: also filter apps.alerts in is_app_only_scope()
The scope filter exists in two locations: setup.rs (fetch_scopes_for_apis)
and auth_commands.rs (is_app_only_scope). Both need the apps.alerts
exclusion to prevent it from appearing in the interactive scope picker.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* chore: retrigger CLA check
* chore: retrigger CI after CLA signing
---------
Co-authored-by: Andrew Barnes <andrew.jaguars@gmail.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* fix: replace unwrap() calls with proper error handling in MCP server
Replace four serde_json::to_string().unwrap() calls in the MCP server
that could panic the process if serialization ever fails. Instead, log
the error to stderr and skip the response gracefully.
Also propagate serialization errors for params/body via the ? operator,
and log a warning when authentication silently falls back to
unauthenticated mode so users are aware of degraded state.
* fix: resolve clippy redundant_closure warnings
Replace `.map(|v| serde_json::to_string(v))` with
`.map(serde_json::to_string)` per clippy's redundant_closure lint.
* fix: harden URL and path construction across helper modules
Closes#87
- gmail/watch.rs: encode msg_id with encode_path_segment(), use
.query() for format and history params instead of format!
- modelarmor.rs: validate template with validate_resource_name() in
handle_sanitize, validate project/location/template_id in
parse_create_template_args, encode all path segments in
build_create_template_url
- discovery.rs: validate service/version with validate_api_identifier()
before use in cache filenames and discovery URLs, encode path segments
- validate.rs: add validate_api_identifier() for safe API name chars
- Add tests for all new validation and encoding paths
* refactor: pass API version as a query parameter for alternative discovery URLs.
The list_gcloud_projects() function piped stdout from `gcloud projects
list` but only read it after the child process exited. When a user has
enough GCP projects that the output exceeds the OS pipe buffer (~64KB),
gcloud blocks on write, the parent blocks waiting for exit, and neither
side makes progress — hitting the 10s timeout.
Spawn a thread to drain stdout concurrently so the pipe buffer never
fills up while the main thread polls for process completion.
Fixesgoogleworkspace/cli#96
* feat: multi-account support with --account flag, per-account credential storage
- Add --account global flag and GOOGLE_WORKSPACE_CLI_ACCOUNT env var
- Per-account encrypted credential files (credentials.<b64-email>.enc)
- Per-account token cache (token_cache.<b64-email>.json)
- accounts.json registry with default account tracking
- New auth subcommands: list, default, per-account logout
- login_hint in OAuth URL for account pre-selection
- Email validation via Google userinfo after OAuth flow
- 12 new unit tests (380 total)
BREAKING CHANGE: Existing users must run 'gws auth login' again.
Credential storage changed from single credentials.enc to per-account files.
* refactor: Improve error handling for file system operations, rename `GWS_ACCOUNT` to `GOOGLE_WORKSPACE_CLI_ACCOUNT`, and refine service account token cache path generation.
* fix: clean up per-account token caches on logout
---------
Co-authored-by: jpoehnelt-bot <jpoehnelt-bot@users.noreply.github.com>
Two bugs fixed:
1. `auth_commands.rs` — `gws auth export` masking used `s[..4]` which
panics on strings shorter than 4 characters, and `s[s.len().min(4)..]`
which evaluates to `s[4..]` on long strings — showing the entire
secret instead of masking it. Replaced with a `mask_secret()` helper
that safely shows only the first 4 and last 4 characters, or "***"
for short strings.
2. `token_storage.rs` — `save_to_disk` silently discarded the return
value of `atomic_write_async` with `let _`, causing the function to
return `Ok(())` even when the write failed. Token persistence failures
now properly propagate via `?`.
* fix: enable APIs individually and surface gcloud errors
Previously, `gws auth setup` used a single batch `gcloud services enable`
call for all 22 Workspace APIs. If any one API failed, the entire batch
was marked as failed. Additionally, stderr was piped to /dev/null, so
users never saw why APIs failed to enable.
Changes:
- Enable each API individually so one failure doesn't block the rest
- Capture stderr from gcloud and include error messages in output
- Show failure details with ⚠ warnings in the wizard status
- Include structured error objects in JSON output (apis_failed array)
Fixes#75
* address review: parallelize API enablement with tokio, remove clone
- Use tokio::process::Command + futures_util::stream::buffer_unordered(5)
for parallel, non-blocking API enablement
- Remove unnecessary failed_apis.clone() by moving into ctx.failed first
- Add changeset
* test: add coverage for enable_apis and JSON output structure
- test_enable_apis_all_already_enabled: empty input returns empty results
- test_enable_apis_with_invalid_project: verifies errors are captured (not swallowed)
- test_failed_apis_json_structure: validates {api, error} JSON shape
- test_failed_apis_json_empty: empty failures produce empty array
* Update src/setup.rs
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
---------
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
* feat: add `gws version` bare subcommand
* refactor: extract help and version flag checks into dedicated functions and add test coverage guidance to AGENTS.md
---------
Co-authored-by: jpoehnelt-bot <jpoehnelt-bot@users.noreply.github.com>
Justin Poehnelt