Profiles And Credentials

Profiles & Credentials — shared pattern reference

Reference skill describing the credential / profile / confirmation pattern used by all claude-team-toolkit skills (azure-devops, fastlane, firebase, heroku, k6, maestro, postgres, rspec, sentry, shopify, slack, trello, etc.). Implementation lives in lib/credentials.sh and lib/confirm.sh.

Overview

All credential-bearing skills share a uniform pattern:

• INI credentials at ~/.<service>/credentials (mode 600)
• Multiple [profile] sections per file (e.g., [default], [work], [prod])
• Resolution order: --profile <name> flag → <SERVICE>_PROFILE env → ~/.<service>/active_profile → [default]
• Helper sourced from lib/credentials.sh populates CTT_<KEY> env vars
• Destructive ops gated by ctt_confirm from lib/confirm.sh
• Actions audit-logged to ~/.claude-team-toolkit/audit.log (key NAMES only, never values)

When to Use

• Authoring a NEW skill that needs per-account/per-org/per-env switching
• Modifying credential loading in an existing skill
• Adding destructive operations that need typed confirmation
• Debugging why a skill picks the wrong profile
• Reviewing security posture of a skill

When NOT to Use

• One-off scripts not meant for reuse → just hardcode and move on
• Skills with no credentials (e.g., react-native, rails-security) → nothing to load
• Cross-cloud unified credential systems → use the cloud's native config (~/.aws/credentials, etc.) instead of reinventing

Profile file shape

# ~/.<service>/credentials  (mode 600)
[default]
field_a = value
field_b = secret

[other-profile]
field_a = different
require_confirm = true        # gate every mutation in this profile

After ctt_load_creds <service> [profile]:

• CTT_PROFILE=<resolved> (e.g., default)
• CTT_FIELD_A=<value>, CTT_FIELD_B=<secret> (uppercased, hyphens → underscores)

Each load clears every CTT_* variable before populating the new ones, except the control variables CTT_NONINTERACTIVE and CTT_HOME. A field present in the old profile but absent from the new one is therefore unset, not inherited — without that, switching profiles would silently carry default_team, org, project, default_app, database, read_only or require_confirm across accounts, and across services. The names come from the shell (compgen -v CTT_), not from a tracking variable: a tracking variable is state a hostile environment can pre-set.

Two boolean fields are normalised at load: read_only and require_confirm become the literal true for any of true, yes, 1, on (case-insensitive) and false otherwise, so [ "$CTT_READ_ONLY" = "true" ] is the correct test in every skill and read_only = yes cannot fail open.

ctt_save_profile refuses profile names outside [A-Za-z0-9_-], field names outside [a-zA-Z0-9_-], and any value containing a newline — each of those could otherwise write a second [section] into the INI file. ctt_load_creds applies the same name rule to the resolved profile, because the name is used in a grep pattern. ctt_audit_log replaces newlines and tabs in the action text so an id or team key cannot forge a second audit record.

Consequence for skill authors: read every field as ${CTT_FIELD:-} and treat unset as "not configured on this profile". Never assume a field survives a profile switch.

Helper API (from lib/credentials.sh)

Confirmation API (from lib/confirm.sh)

ctt_confirm "Delete app foo?" || return 1                # y/N prompt
ctt_confirm "Drop table users?" "DELETE" || return 1      # typed phrase
ctt_warn_destructive "Action description"                # noisy stderr banner

CTT_NONINTERACTIVE=1 auto-denies all confirmations (CI safety).

Common Mistakes

• Reading credentials directly with awk instead of ctt_load_creds → bypasses perm validation and the cross-profile clearing
• Assuming a field persists across a profile switch → it is cleared on every load. Read ${CTT_FIELD:-}, not $CTT_FIELD.
• Logging full credential values instead of ctt_mask → secrets leak to terminal/CI logs
• Using eval to set vars from credentials → injection risk. Helper uses printf -v for safety.
• Audit log records values not just key names → audit log itself becomes a secret store
• Confirmation without typed phrase for irreversible ops → "y" is too easy a typo
• Service name with uppercase or special chars → helper validates ^[a-z0-9-]+$

Security defaults

• Credentials directory: chmod 700. File: chmod 600. Helper refuses to load otherwise (POSIX).
• Windows: best-effort (NTFS ACLs not enforced by helper) — document in skill's safety section.
• CTT_NONINTERACTIVE=1 in CI → auto-deny destructive ops (no surprise mutations).
• Audit log at ~/.claude-team-toolkit/audit.log (mode 600). Records: timestamp, service, profile, action.
• Never write credentials into git, chat output, or any file outside ~/.<service>/credentials.