Deliver vault_op_service_account_claude to the agent hosts and install a
role-owned ~multica/.local/bin/op shim that exports OP_SERVICE_ACCOUNT_TOKEN
to /usr/bin/op only. op run / op plugin run get env -u spliced after "--"
so the child never sees the token. The daemon unit is untouched (WDWL-18).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Agents get `op` so op:// references resolve at command time and secret
values stay out of prompts and transcripts. Installed per 1Password's
apt instructions (signed-by keyring, debsig policy + keyring,
1password-cli), with the key fingerprint-checked before it is trusted;
EL via their yum repo. multica_op_version pins, "" tracks the repo.
Tag: multica_op.
Optional, default-off hook writes OP_SERVICE_ACCOUNT_TOKEN from the
playbook (ansible-vault) to a 0600 file owned by the daemon user. It is
deliberately not wired into the daemon environment: the user unit
belongs to the multica CLI, and the delivery design is pending.
Rolled out with --tags multica_op to haunted-mansion and spaceship-earth;
second run changed=0, op 2.41.1 as multica, debsig verifies the .deb.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The npm-installed Claude Code on spaceship-earth updated itself from the
pinned 2.1.304 to 2.1.308 right after its first login, so every play would
downgrade it and it would update again. New multica_claude_settings is
merged (recursively, not templated, since Claude writes the file too) into
~multica/.claude/settings.json.
provision-agent-host.yml sets DISABLE_AUTOUPDATER=1 and mirrors the model
mix already live on haunted-mansion: model sonnet, subagents haiku,
MAX_THINKING_TOKENS 4000, BASH_MAX_OUTPUT_LENGTH 20000, thinking off.
spaceship-earth: back on 2.1.304 and staying there; second run changed=0.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
rblundon