agentleFS
therapystherapys/dotagents/skills/gog/SKILL.md

gog

Skill8 starsChanged 2 months agoMITView on GitHub

Use the gog CLI for safe Google Workspace automation: Gmail and email, Calendar and events, Drive files, Docs, Sheets, Contacts, Tasks, Chat, and other Google services. Covers auth, stable JSON, scoped reads, and approved writes.

We split this skill into the rules it gives an agent

23 rules. 0 are common, 7 are distinctive, and 16 are very specific to this project.

  • 0

    Common

    Most of the 49 edited versions we read keep these.

  • 7

    Distinctive

    This author's own choices. Worth a look.

  • 16

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Use gog when built-in Google connectors are missing a feature, when shell automation needs stable JSON, or when you need to inspect local …

    line 8
  2. Very specific:

    Prefer --json --wrap-untrusted for agent parsing when reading Google content.

    line 43
  3. Distinctive:

    Do not print access tokens, refresh tokens, OAuth client secrets, or keyring passwords.

    line 48
  4. Very specific:

    If GOG_KEYRING_PASSWORD is provided by a shell startup file or service environment, use the matching shell/entrypoint so gog can unlock the …

    line 50
  5. Very specific:

    Use --no-input in automation so auth/keyring prompts fail clearly.

    line 56
  6. Very specific:

    Use --dry-run first where commands support it.

    line 57
  7. Very specific:

    Use --readonly for tasks that must not mutate Google data.

    line 58
  8. Very specific:

    Use --gmail-no-send or GOG_GMAIL_NO_SEND=1 unless sending mail is the requested task.

    line 62
  9. Very specific:

    Default for existing human/user OAuth reauth: preserve broad service access.

    line 88
  10. Very specific:

    Before reauth, run gog auth list --check --json --no-input and inspect the account's existing services.

    line 89
  11. Distinctive:

    Use narrow services only for throwaway/test accounts, service-specific bot accounts, explicit user requests, or scoped security experiments.

    line 94
  12. Very specific:

    If this fails with keyring.password while the same gog auth doctor works in the shell, fix the service or agent environment before …

    line 111
  13. Distinctive:

    Run the CLI leg in a detached tmux session so it survives command boundaries:

    line 119
  14. Very specific:

    Write the URL to a mode-0600 file and hand it to the browser by file reference (see $browser-use).

    line 140
  15. Distinctive:

    If tmux asks for the file-keyring passphrase, source it from that host's login environment via the login shell and paste it in without …

    line 166
  16. Distinctive:

    Verify, and require both the account and its scope breadth:

    line 169
  17. Very specific:

    Before writes, identify the account, object id, and exact mutation.

    line 197
  18. Very specific:

    Use docs list-tabs <documentId> --json to discover tab titles/IDs before targeting a tab.

    line 213
  19. Very specific:

    Use docs write --markdown --replace --tab <tab> for whole-tab formatted replacement.

    line 214
  20. Very specific:

    Use docs update --markdown --tab <tab> for formatted insertion/append without replacing the whole tab.

    line 215
  21. Very specific:

    Use docs update --replace-range START:END for precise plain-text replacement.

    line 216
  22. Distinctive:

    When testing creation commands, name artifacts with a clear temporary prefix and delete or trash them after verification.

    line 220
  23. Distinctive:

    Use generated command docs and schema instead of guessing flags:

    line 246

“Common” is measured: kept by at least 6 in 10 of the 49 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

gog

Use gog when built-in Google connectors are missing a feature, when shell automation needs stable JSON, or when you need to inspect local Google auth state before acting.

Fast Path

gog --version
gog auth list --check --json --no-input
gog auth doctor --check --json --no-input
GOG_HELP=agent gog --help
gog schema --json

GOG_HELP=agent makes root help emit a compact automation contract and common read-only recipes; commands and behavior stay unchanged. Machine output, non-interactive behavior, stable exit codes, command guards, and untrusted-content wrapping apply across the CLI. schema exposes command syntax, stable exit codes, and effective safety state for automation.

For JSON output projection, --fields is accepted as an alias for --select on commands that do not define their own API field-mask --fields; commands with a local field-mask flag keep that command-specific meaning.

--results-only unwraps the primary result before --select projects it. For lists, select item-relative fields: --results-only --select id. Dot paths do not broadcast through nested arrays (--select items.id selects nothing). Unmatched object fields are omitted.

Pick the account explicitly for API work:

gog --readonly --account user@example.com gmail search 'newer_than:7d' --json --wrap-untrusted

Prefer --json --wrap-untrusted for agent parsing when reading Google content. Human hints and progress should stay on stderr; stdout is for data.

Safety Rules

  • Do not print access tokens, refresh tokens, OAuth client secrets, or keyring passwords.
  • If GOG_KEYRING_PASSWORD is provided by a shell startup file or service environment, use the matching shell/entrypoint so gog can unlock the file keyring non-interactively. Do not print the value.
  • In headless/service agents, verify the service environment, not just the login shell. GOG_KEYRING_BACKEND=file, GOG_KEYRING_PASSWORD, and HOME must be present in the process that launches gog.
  • Use --no-input in automation so auth/keyring prompts fail clearly.
  • Use --dry-run first where commands support it.
  • Use --readonly for tasks that must not mutate Google data; remove it only for the exact write the user approved.
  • Destructive commands require --force; do not add it unless the user asked for that exact mutation.
  • Use --gmail-no-send or GOG_GMAIL_NO_SEND=1 unless sending mail is the requested task.
  • For shared agent environments, prefer a baked readonly or agent-safe binary from docs/safety-profiles.md.

Runtime command guards:

gog --readonly --enable-commands gmail.search,gmail.get --gmail-no-send \
  --account user@example.com gmail search 'from:example@example.com' --json

gog --enable-commands drive.ls,docs.cat --disable-commands drive.delete \
  --account user@example.com drive ls --max 10 --json

Auth

OAuth setup is partly interactive. An agent can inspect and diagnose it, but a human normally completes browser consent:

gog auth credentials list
gog auth add user@example.com --services all-user --force-consent
gog auth remove user@example.com

Default for existing human/user OAuth reauth: preserve broad service access. Before reauth, run gog auth list --check --json --no-input and inspect the account's existing services. When replacing an expired or revoked token, do not silently reduce scope; prefer --services all-user --force-consent unless the user explicitly asks for narrower scopes.

Use narrow services only for throwaway/test accounts, service-specific bot accounts, explicit user requests, or scoped security experiments. Safety should normally be enforced at command time with --enable-commands, --disable-commands, --gmail-no-send, dry-runs, and account selection, not by under-scoping durable user auth.

Service accounts are Workspace-only and mainly fit Admin, Groups, Keep, and domain-wide delegation flows; they do not solve consumer @gmail.com OAuth.

For OpenClaw/systemd setups, run the diagnostic through the actual agent entrypoint after restarting the service:

openclaw agent --agent main --message \
  'Run: gog auth doctor --check --no-input && gog gmail search "newer_than:1d" --max 1 --json'

If this fails with keyring.password while the same gog auth doctor works in the shell, fix the service or agent environment before reauthenticating.

Browser-Driven Reauth

An agent can complete this flow end to end when it can drive a signed-in browser. Consent still happens in a real browser; nothing here bypasses it.

Run the CLI leg in a detached tmux session so it survives command boundaries:

tmux -L gog-auth new-session -d -s auth -x 200 -y 50
tmux -L gog-auth send-keys -t auth \
  "gog auth add user@example.com --services all-user --force-consent --timeout 15m" Enter

Capture the consent URL with -J:

tmux -L gog-auth capture-pane -t auth -p -J -S - \
  | grep -oE 'https://accounts\.google\.com[^ ]+' | tail -1 > "$url_file"

-J is mandatory. capture-pane otherwise returns the URL hard-wrapped at the pane width. A truncated consent URL does not fail loudly: Google renders Invalid OAuth Request / Invalid response_type: missing, which reads like a client misconfiguration and sends you debugging the wrong thing. Verify the captured URL contains response_type before using it.

Write the URL to a mode-0600 file and hand it to the browser by file reference (see $browser-use); never echo it. Appending &login_hint=user@example.com skips the account chooser and removes a whole class of wrong-account risk.

Expect up to two interstitials when the OAuth client is unverified or in testing:

  1. "Google hasn't verified this app." Continue is a low-emphasis link on one side; Back to safety is the prominent button. Activating the visually obvious control aborts the flow. A developer-info control sits in the tab order between them, so count focus stops deliberately instead of guessing.
  2. "You're signing back in to <app>" — confirm the displayed account is the intended one, then Continue.

The listener enforces --timeout. When it expires the tmux pane simply returns to a shell prompt, so a flow that "did nothing" is often an expired listener rather than a browser problem. Read the pane before re-driving the browser, and restart the CLI leg rather than reusing a stale URL. Complete the browser leg promptly; batch the navigate-and-activate steps instead of round-tripping.

The browser does not have to run on the CLI's host. The callback targets http://localhost:<port>, so when they are separate machines, forward that port from the browser host to the host running the listener before opening the URL, and confirm the forward is live first. Keep the browser on the host whose profile holds the intended Google session.

If tmux asks for the file-keyring passphrase, source it from that host's login environment via the login shell and paste it in without printing it.

Verify, and require both the account and its scope breadth:

gog auth list --check --json --no-input

Confirm the target account reports valid and retains the expected service list; a successful login that silently narrowed scopes is a failed reauth.

Common Reads

gog --readonly --account user@example.com gmail search 'newer_than:3d' --max 10 --json --wrap-untrusted
gog --readonly --account user@example.com gmail get <messageId> --sanitize-content --json --wrap-untrusted
gog --readonly --account user@example.com gmail thread get <threadId> --sanitize-content --json --wrap-untrusted

gog --readonly --account user@example.com calendar events --today --json --wrap-untrusted
gog --readonly --account user@example.com drive ls --max 20 --json --wrap-untrusted
gog --readonly --account user@example.com docs cat <documentId> --json --wrap-untrusted
gog --readonly --account user@example.com sheets get <spreadsheetId> Sheet1!A1:D20 --json --wrap-untrusted
gog --readonly --account user@example.com contacts list --max 20 --json --wrap-untrusted

For Gmail body inspection, prefer --sanitize-content unless the user explicitly needs raw payloads.

Writes

Before writes, identify the account, object id, and exact mutation. Prefer commands that support --dry-run, and clean up disposable live-test objects.

gog --account user@example.com docs write <documentId> --append --text '...'
gog --account user@example.com docs write <documentId> --tab "Data" --markdown --replace --file data.md
gog --account user@example.com docs update <documentId> --tab "Data" --markdown --file block.md
gog --account user@example.com docs update <documentId> --tab "Data" --replace-range START:END --text 'replacement'
gog --account user@example.com docs update <documentId> --tab "Data" --markdown --replace-range START:END --file block.md
gog --account user@example.com sheets update <spreadsheetId> Sheet1!A1 --values-json '[["hello"]]'
gog --account user@example.com sheets batch-update <spreadsheetId> --data-json @updates.json
gog --account user@example.com drive upload ./file.txt --parent <folderId> --json

For Google Docs tab work:

  • Use docs list-tabs <documentId> --json to discover tab titles/IDs before targeting a tab.
  • Use docs write --markdown --replace --tab <tab> for whole-tab formatted replacement.
  • Use docs update --markdown --tab <tab> for formatted insertion/append without replacing the whole tab.
  • Use docs update --replace-range START:END for precise plain-text replacement; add --markdown to replace that exact range with formatted markdown.
  • START:END is a Google Docs UTF-16 API range. Resolve it from docs cat --raw, docs raw, or another documents.get readback; do not guess indexes.
  • --replace-range and --index are mutually exclusive.

When testing creation commands, name artifacts with a clear temporary prefix and delete or trash them after verification.

gmail batch delete permanently deletes messages and requires the broader https://mail.google.com/ OAuth scope. Prefer gmail trash; when permanent deletion is required, follow the exact reauthorization command printed by gog.

For larger Sheets writes, prefer sheets batch-update over loops of sheets update; it sends multiple value ranges in one Sheets API request and accepts inline JSON or @file input.

For normal Gmail replies, use the first-class commands instead of rebuilding reply MIME through gmail send:

gog --account user@example.com gmail reply <messageId> --body-file reply.txt
gog --account user@example.com gmail reply-all <messageId> --body-file reply.txt \
  --bcc introducer@example.com --remove former-participant@example.com

They inherit the subject, quote by default, preserve display names and inline images, and treat --to/--cc/--bcc as additive placement or moves. Use --no-quote to omit the original.

Discovery

Use generated command docs and schema instead of guessing flags:

gog <service> --help
gog <service> <command> --help
gog schema <service> <command> --json

Upstream docs:

Upstream repo paths:

  • CLI entrypoint: cmd/gog/
  • Command implementations: internal/cmd/
  • OAuth/keyring: internal/googleauth/, internal/authclient/, internal/secrets/
  • Generated command docs: docs/commands/

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

Reviews can't be read right now.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.