Ansible

Three biases decide most calls:

• The ansible-core line the project pins decides what may be written, not the binary on the machine. The 2.19 templating rewrite is the boundary: content that runs on 2.18 can fail, warn, or render literally on 2.19 and later with no YAML change.
• ok and changed are claims made by the module, not measurements. A command task claims a change on every run, an async launch always claims one, and a --check run skips whatever it cannot simulate. Idempotence is proven by a second run reporting changed=0, and by nothing else.
• Where a value comes from is not guessable. Inventory layout, group_vars adjacency, ansible.cfg discovery, and collection versions each change what a play does. Read them before the first edit.

Core Version

Read the pin before writing: requires_ansible in meta/runtime.yml, the ansible-core line in a requirements or execution-environment file, and the CI matrix. ansible --version reports one machine. The ansible community package pulls one ansible-core line: 11 → 2.18, 12 → 2.19, 13 → 2.20, 14 → 2.21.

Support windows, which decide whether a pin is defensible:

• 2.18 — end of life May 2026. No fixes of any kind. Control Python 3.11 to 3.13, target 3.8 to 3.13.
• 2.19 — security fixes only from 2026-05-18, end of life November 2026. Control 3.11 to 3.13, target 3.8 to 3.13.
• 2.20 — critical fixes only from 2026-05-18, security only from 2026-11-02, end of life May 2027. Control 3.12 to 3.14, target 3.9 to 3.14.
• 2.21 — GA 2026-05-18, critical only from November 2026, end of life November 2027. Control 3.12 to 3.14, target 3.9 to 3.14.

Floor per feature — the changes the rules below depend on:

• 2.18 — loop_control.break_when; meta: end_role; timedout and vaulted_file tests; raw over SSH to Windows runs unwrapped, without powershell.exe.
• 2.19 — trust-tagged templating, boolean-only conditionals, no multi-pass rendering, lazy evaluation, native types only; ALLOW_BROKEN_CONDITIONALS; DISPLAY_TRACEBACK; SSH_ASKPASS replaces sshpass; a become timeout is unreachable; omit dropped from loop items; bool filter narrowed; plural collections_paths removed; paramiko and auto_legacy deprecated.
• 2.20 — control Python 3.12+, target 3.9+; transport = smart removed; Galaxy v2 API removed; vaultid filter parameter removed; include_vars rejects a non-list extensions; INJECT_FACTS_AS_VARS deprecated, default flips in 2.24.
• 2.21 — register projections and the _task object; ansible-galaxy refuses a collection whose requires_ansible excludes the running core; paramiko and auto_legacy removed; invocation dropped from results; failure inference from rc deprecated, warnings from 2.22; to_yaml decrypts vaulted values again.

Read [${CLAUDE_SKILL_DIR}/references/versions/core-2.NN.md] — core-2.18.md through core-2.21.md — when writing against a feature near the floor, when raising the floor, and before touching content that must run on 2.18 and 2.19 both. Each carries what its line added, changed, removed, and deprecated with the removal target, and the traps.

Templating

The 2.19 engine holds on every maintained line. Write for it; do not write for 2.18 unless the project pins 2.18.

• Never put {{ }} inside when, failed_when, changed_when, until, break_when, assert.that, or a register projection. They are expressions already. Delimiters there are the error Template delimiters are not supported in expressions.
• A conditional must evaluate to a boolean. when: some_string fails with Conditional result was 'x' of type 'str' ... Conditionals must have a boolean result. Write some_string | length > 0, is truthy, is defined, or | bool. A YAML list under when is and; write or in one string.
• A string that arrives with {{ }} in it is data. Module results, facts, command output, and file contents are never rendered, and no setting restores rendering. A custom plugin that returns template text must tag it with trust_as_template.
• Types survive. None is not ""; a list through a string filter becomes a string; range() needs | list before it is stored; "127" == 127 is false. Convert with int, string, bool, list.
• default() binds to the immediate value only. A container with an undefined leaf is defined; put default at the leaf. default(x, true) also replaces false, 0, "", and empty containers.
• default(omit) belongs on the module argument. A loop item value that resolves to omit is dropped from the item, so item.key is undefined, not a placeholder.
• set_fact evaluates on assignment; a vars entry is re-evaluated on every access. Use set_fact to freeze a value, vars to derive one.
• {% if %}text{% endif %} in a value yields None when the branch is not taken. Add an explicit {% else %}{{ "" }}{% endif %}.
• set_fact does not turn "yes" or "true" strings into booleans. Use YAML booleans.

Read [${CLAUDE_SKILL_DIR}/references/templating.md] when writing a filter chain, a boolean conversion, a combine merge, a serialization, or when an error names a type — it carries the expression contexts, the type rules, undefined handling, omit, dictionary and list filters, and the trust model.

YAML

• Quote a value that starts with {{. Unquoted, { opens a flow mapping and the error is a YAML syntax error.
• Quote file modes: mode: "0644". Unquoted 0644 is 420. Symbolic u=rw,g=r,o=r also works.
• Write true and false. yes, no, on, off are booleans to the parser; version: yes is not a string.
• Quote an expression that contains : inside assert.that or when; otherwise it parses as a mapping and fails with Conditional expressions must be strings.
• !unsafe marks a value as data so it is never rendered even from a trusted file. {% raw %} keeps delimiters literal inside a template that is otherwise rendered.
• INI [group:vars] values are strings, while key=value on a host line is a Python literal. Apply | bool or | int when consuming, or use YAML inventory.

Idempotence

• command and shell report changed on every run. Give them creates, removes, or changed_when; changed_when: false on a read. Prefer the module that owns the state (apt, service, file, lineinfile).
• command runs no shell. No |, >, ;, &, globbing; $VAR is expanded by Python and left alone when unset. Use argv for arguments with spaces and quote for interpolated values.
• A shell pipeline hides the first failure. Start it with set -o pipefail and set executable: /bin/bash.
• An async task always reports changed. Put changed_when, creates, and friends on the async_status task.
• lineinfile regexp must match the line after the change too, or every run rewrites it. blockinfile needs a distinct marker per logical block, or one block replaces another.
• A loop is changed if any item changed, and every notified handler fires once for the whole task. Per-item truth is in results[].
• state: latest and an unpinned git checkout give a different result on a different day. Pin the version.
• Check mode skips a module that cannot simulate, and command without creates or removes is one of them. check_mode: false forces a task to run under --check; ansible_check_mode is the variable to branch on. --diff reveals file contents; diff: false on a secret.

Failure

• ignore_errors swallows only a failed result. An undefined variable, a syntax error, a missing interpreter, or an unreachable host is not covered; ignore_unreachable handles the last.
• A failed_when list is and. failed_when: rc != 0 or 'ERROR' in stdout in one string fails on either.
• A become timeout is unreachable, not failed, so ignore_errors does not cover it.
• A rescue that succeeds reverts the failure for play control: the host continues, and neither max_fail_percentage nor any_errors_fatal counts it, while the recap counts it under rescued. An invalid task definition or an unreachable host never enters rescue. ansible_failed_task and ansible_failed_result are set inside it.
• max_fail_percentage is per serial batch and must be exceeded, not met: with serial: 4, aborting on two failures needs 49. any_errors_fatal finishes the failing task on the current batch, then stops every host.
• A module that returns rc: 1 and no failed is failed by inference on 2.21 and earlier, and the inference is deprecated from 2.21. Set failed or call fail_json in every module.
• _task.result replaces register inside failed_when, changed_when, and until on 2.21; in a loop it needs default({}) on the first iteration.

Handlers

• A handler does not run on a host where a later task failed unless force_handlers is set. The configuration is changed and the service is not restarted. meta: flush_handlers runs the handlers notified so far, and works inside rescue.
• Handlers run in definition order, once per play section (pre_tasks, roles and tasks, post_tasks), no matter how many tasks notified them or in what order.
• Two handlers with one name shadow: the last loaded wins, in the order roles:, handlers:, import_role, include_role. Notify a role's handler as role_name : handler_name; use listen topics to decouple.
• Never template a handler's name. Names are templated early; a name with an undefined variable, a syntax error, or a value resolving to omit is skipped on 2.19 and later and fails the play on 2.18, and in no case runs. Put the variable in the handler's arguments. listen topics cannot be templated at all.
• A handler inside an include_tasks file cannot be notified; a handler that is an import_tasks is rewritten into its tasks. A handler cannot include_role or import_role, and handlers ignore tags.

Delegation and Execution

• Under delegate_to, everything except inventory_hostname is the delegate's: ansible_host, the interpreter, connection, become, and shell options, even inventory_hostname_short. Reach the original host with hostvars[inventory_hostname]. Facts gathered by a delegated task land on the original host unless delegate_facts: true.
• Delegation does not serialize. Five forks writing one file on one delegate overwrite each other. Use throttle: 1, or run_once: true with a loop over ansible_play_hosts_all.
• run_once is once per serial batch, evaluated with the first host's variables. Truly once is when: inventory_hostname == ansible_play_hosts_all[0].
• throttle can only lower parallelism below forks (default 5) and the batch size. serial takes a number, a percentage, or a list of either; the last batch holds the remainder and a batch is never smaller than one.
• order: inventory is the compiled selection order, not the file order. Use sorted when the order matters.
• connection: local runs modules under /usr/bin/python unless ansible_python_interpreter is "{{ ansible_playbook_python }}". delegate_to: localhost has no such trap.
• add_host and group_by change inventory for later plays, subject to --limit; the current play's host list is already fixed.
• Third-party strategy plugins are deprecated. linear, free, host_pinned, and debug are the set.

Variables

Lowest to highest, the rungs that decide most disputes: role defaults/, inventory group vars, inventory host vars, facts and cached set_fact, play vars, role vars/, block and task vars, include_vars, registered results and set_fact, role parameters, extra vars.

• defaults/ is for what a caller overrides; vars/ is for what the role forces. A value in vars/ beats inventory, facts, and play vars; block and task vars, include_vars, set_fact, role parameters, and -e beat it.
• roles: and import_role publish a role's variables to the whole play at parse time, visible to earlier tasks; include_role publishes nothing unless public: true, and then only to later tasks. include_role also runs the role every time it is called, where roles: entries with equal parameters run once per play.
• A registered variable exists on every host, including where the task failed or was skipped. Test is skipped or is failed, not is defined.
• set_fact with cacheable: true makes two copies, a host variable at set_fact precedence for this run and a fact seven rungs lower that the cache persists. meta: clear_facts removes only the fact.
• Groups at one level merge alphabetically; ansible_group_priority reorders them and is honored only in an inventory source, never in group_vars/.
• group_vars/ next to the playbook beats group_vars/ next to the inventory, and ansible-inventory without --playbook-dir never reads the playbook-adjacent copy, so it can print a value a play does not use.
• -e key=value is a string; -e '{"key": 1}' and -e @file.yml keep types.
• Ansible loads the first ansible.cfg found — ANSIBLE_CONFIG, ./ansible.cfg, ~/.ansible.cfg, /etc/ansible/ansible.cfg — and refuses one in a world-writable current directory.
• Write ansible_facts['os_family'], not ansible_os_family. The injected form warns on 2.20 and later and is off by default from 2.24.
• Facts are a snapshot from the gathering task, once per play by default; a cached fact is from a previous run within CACHE_PLUGIN_TIMEOUT (86400 s). gather_subset and --limit decide which hosts have which facts.

Read [${CLAUDE_SKILL_DIR}/references/precedence.md] when two definitions of one name disagree, when a value differs between hosts or commands, or when designing a group_vars layout — it carries the full 22-rung ladder, the inventory merge order, directory adjacency, INI and extra-vars typing, scope, and role publication.

Roles and Collections

• Always the FQCN: ansible.builtin.copy. ansible.legacy.copy only where a local library/ override must win. Short names resolve at run time and fail ansible-lint --profile production.
• The collections: keyword is not inherited by roles, and never applies to lookups, filters, or tests. A role sets its own list in meta/main.yml, and a role inside a collection searches its own collection first.
• meta/argument_specs.yml validates coercively and passes the original value through: 1 satisfies type: str and the role receives the integer. The validation task carries the always tag and runs dependencies' specs first.
• A role listed twice with equal parameters runs once per play. Different role parameters (not vars:), tags, or when make it distinct; allow_duplicates: true in the role's own meta/main.yml forces it. Deduplication runs before variable evaluation.
• dependencies run before the role and are deduplicated the same way; vars: inside a dependency entry is play scoped and reaches roles that ran earlier. Galaxy import requires dependencies in namespace.role form, never a URL.
• requirements.yml takes roles: and collections: sections in one file. Roles take a single version with no ranges; collections take >=, ==, != (treated as unpinned), and *. ansible-galaxy install -r installs both sections; role install and collection install each install their own.
• A collection's requires_ansible is enforced at install from 2.21: a mismatch is skipped, not installed and failed later. COLLECTIONS_ON_ANSIBLE_VERSION_MISMATCH governs both the install and the load-time warning.
• plugin_routing in meta/runtime.yml redirects, deprecates, and tombstones names. Read it before declaring a plugin missing.
• A Galaxy server must speak the v3 API from 2.20. collections_path is singular from 2.19.
• Roles inside collections cannot embed plugins in library/ or filter_plugins/; standalone roles can, and the plugins stay visible to roles that run after them.

Connection and Become

• Pipelining is off by default and is the largest cheap speedup. It needs requiretty absent from sudoers and does not apply to file-transfer modules (copy, template, fetch).
• ssh_args replaces the default -C -o ControlMaster=auto -o ControlPersist=60s. Add options through ssh_common_args or ssh_extra_args; a custom -o ControlPath disables the managed control path.
• SSH passwords use SSH_ASKPASS from 2.19; sshpass is opt-in through ansible_ssh_password_mechanism. A key with a passphrase needs an agent; Ansible never proxies the passphrase prompt.
• become_user does not imply become. -K supplies one become password for every host; per-host passwords go in vaulted ansible_become_password.
• Privilege escalation cannot be limited to specific commands. Modules run from a temporary file with a changing name, so a sudoers command allowlist breaks every module.
• An unprivileged connection user becoming another unprivileged user needs POSIX ACLs (setfacl) on the target, a shared group named in ansible_common_remote_group, or pipelining. Without one of them Ansible errors rather than fall back to world_readable_temp.
• Interpreter discovery is auto. Pinning ansible_python_interpreter silences the warning and freezes the path.
• ansible_host must never depend on inventory_hostname; use inventory_hostname_short so delegation works.
• become does not open a systemd session: XDG_RUNTIME_DIR is empty and systemctl --user fails; become_method: machinectl opens one.
• paramiko is gone from core in 2.21. ansible_connection: paramiko fails to load; network platforms use ansible.netcommon connections.

Vault

• Vault protects data at rest only. Decrypted values reach modules, diffs, callbacks, and errors. Put no_log: true on the task that handles the secret; it does not protect a later debug of the same variable.
• ansible-vault encrypt_string produces an inline !vault scalar; encrypt encrypts a whole file. The vault_ prefix convention (db_password: "{{ vault_db_password }}") keeps variable names greppable at the cost of two variables in precedence.
• A vault-id label is a hint. Every supplied password is tried, matching label first; DEFAULT_VAULT_ID_MATCH restricts each password to its label. Two files can carry the same label and different passwords; rekey fixes it.
• A password source with the executable bit set is run, not read. A client script must accept --vault-id, print the password, and be named *-client or *-client.EXT. A non-executable file is read as the password.
• ANSIBLE_VAULT_PASSWORD_FILE or vault_password_file in ansible.cfg is a silent default source that changes which secret a run uses from machine to machine; ansible-lint honors both.
• An encrypted file passed as src to copy, template, unarchive, script, or assemble lands decrypted on the target.
• to_yaml on a vaulted value writes !vault ciphertext on 2.19 and 2.20, and the decrypted value on 2.21 (vault_behavior selects). vault_encrypted tests a variable, vaulted_file a path.

Verification

• ansible-lint from the project root, under the ansible-core line the project pins. With no --profile every rule applies; --profile production is the bar for content that ships. The rules that fire most on generated code: no-changed-when, risky-shell-pipe, fqcn, name[casing], name[template], var-naming[no-role-prefix], yaml[octal-values], yaml[truthy], no-free-form, partial-become. # noqa: rule-id at the end of the line suppresses one rule; skip_list hides it from output.
• --syntax-check parses without connecting; --check --diff on one host simulates what modules can simulate and skips the rest.
• Run the play twice; the second recap is the idempotence result. Molecule's idempotence action is exactly that comparison and nothing deeper: changed_when: false passes it, a bare command fails it. A clean second recap proves the play converged, not that the target runs the result: a run against a real target is a change, and the effect check in the devops skill's references/effect.md is what closes it.
• molecule test runs a configurable test_sequence; a scenario that dropped idempotence or verify passes all the same. The built-in driver provisions nothing; ansible-native scenarios create instances in create.yml and write their addresses into an inventory source. Drivers come from molecule-plugins.
• ansible-inventory --list -i <source> --playbook-dir . shows the merged variables a play will see; ansible-config dump --only-changed shows which ansible.cfg won; ansible-doc -t <type> <fqcn> is the argument spec.
• ansible-lint 26.8.0 and Molecule 26.8.0 support the two newest ansible-core lines at their release. Testing 2.18 content needs a pinned older toolchain in its own environment.

Read [${CLAUDE_SKILL_DIR}/references/lint.md] when a rule fires that the list above does not name, when configuring .ansible-lint, or when a lint result disagrees with a run — it carries the profiles, every rule that fires on generated code, suppression, and the coupling to the ansible-core line. Read [${CLAUDE_SKILL_DIR}/references/molecule.md] when writing or repairing a scenario — it carries the ansible-native and legacy layouts, the sequences, the idempotence criterion, dependency paths, and the MOLECULE_ variables.

Boundary

The automation controller and its job templates and surveys (AWX and the Automation Platform), automation mesh, Automation Hub certification, the platform's Event-Driven Ansible service, and Lightspeed are out of scope; a play never configures them. ansible-rulebook, ansible-builder, ansible-navigator, and the Ansible development tools MCP server are open source and usable without the platform.

Application

When writing Ansible, apply these rules silently — do not narrate a rule while following it. Where the existing content contradicts one, follow the content and flag the divergence once. State the ansible-core line an edit assumes when it is not 2.19-compatible.

When reviewing Ansible, cite the violation and give the corrected line inline. Do not lecture.

Integration

The devops skill governs the change: what it must carry before apply, how it is applied, and the evidence that closes it. This skill governs how Ansible behaves and how Ansible is written, and wins on any question of what a play, role, or inventory does. The coding skill governs the shape of the work; the containers skill owns the images an execution environment is built from. All are active at once.