Country AI Path

Give one country a working AI path game rule, one country per chat. Works for a country that has a rule to standardise and for one that has none yet.

Syntax: /country-ai-path <country or TAG> [#issue] — e.g. /country-ai-path KOS, /country-ai-path Kosovo #1234. The country is required; an issue number is optional. Requested arguments: $ARGUMENTS

Context budget

Focus trees run 8k–42k lines. Never read one end to end, and never open another country's tree to learn a shape. Everything you need about the tree comes from ai_path_report.py; grep for the specific lines it names. Templates for every artefact are in references/write.md.

1. Target

Resolve the argument to a TAG: a three-letter argument is the TAG; a name is looked up in common/country_tags/ and localisation/english/ (TAG: "Name"). No argument, or an ambiguous name — ask, don't guess. If the user names an issue, read it: its body lists the country's known defects and goes into the PR as Closes #N.

Branch <tag-lower>-ai-paths off main.

2. Facts

python tools/analysis/ai_path_report.py --tag TAG

The first line of the report decides which of three shapes you are in:

• Rule exists. Standardise it (§3–§5). The report's RULE / WIRING section lists what deviates.
• no TAG_ai_behavior rule, tree found. Build the rule from nothing: every artefact in references/write.md §1–§7, in that order. The report still audits the tree, the flags you add and the country's own mechanics, so re-run it after each artefact.
• no focus file found for tag TAG. The country has no path layer to own. If a rule exists it is removed, not converted (São Tomé #3702, Solomon Islands #4266 are the shape): delete the TAG_ai_behavior block and its TAG_AI_BEHAVIOR / RULE_OPTION_* keys from MD_game_rules_l_english.yml, then strip every has_game_rule read from the country's events — keep factor = 5 is_historical_focus_on = yes on the historical option of each fork and any situational flavour modifier, drop factor = 1 no-ops and the is_historical_focus_on = no coin-flip nudges, and delete an ai_chance block with no modifier left. Non-English files keep their orphaned keys. If no rule exists either, say so and stop — a rule needs a tree to steer. Sections 3–5 do not apply; verify with the grep set in §5 and validate_events.py / validate_localisation.py, and say in the PR that the report does not apply.

The report decides every mechanical question: rule and loc conformance, flag wiring, which focuses carry path modifiers and whether they are multiplicative, path flags that appear nowhere, killswitch orphans per rule state × historical AI on/off, mutex ties, focus_factors disagreements, dangerous completion rewards, and the country's own mechanics — the burdens its history file hands it, what relieves each, which burdens go unrelieved in some rule state, and whether each country GUI is decision-backed or player-only. Read references/audit.md after the report — it covers only the judgment the report cannot make.

3. Design

The judgment calls: which fork axis and how many options, whether each branch root is reachable by something the AI can satisfy, whether party drift smothers the ramp, what must be killswitched. Derive the fork from the prerequisite graph and each side's available, never from the option names. Where the report is ambiguous about branch structure, dispatch one Explore subagent for the taxonomy — it returns the taxonomy, not file contents.

4. Write

Every artefact from references/write.md. Focus weights are not written by hand: author the mapping and run

python tools/standardization/apply_ai_path_weights.py --tag TAG --map <mapping>

Loc drafting and _desc sentence-count fixes go to a localisation-editor subagent on haiku.

Defects you find are in scope. A broken fork, a timing race between the country's own path events, an asymmetric branch, a wrong state id, or a typo in an English string the path events show gets fixed in the same PR, in its own commit, never deferred as a follow-up. English values only — never rename a key that non-English files carry.

Rule standard. Exactly HISTORICAL + one option per alt-history path + RANDOM_PATH + NO_PATH, and NO_PATH is the default = { } block, listed last — a country the player never configures runs unscripted. No DEFAULT, no RANDOM; merge any duplicate DEFAULT/HISTORICAL. Write the options fresh — don't recycle a stub's names or bucket count. The historical option's displayed text is literally "Historical"; its _desc carries the country's history. Player-facing names, no internal jargon, no "random" in a path name. Every _desc exactly two sentences, present tense about the country, no hard dates, §8…§! on party names. Both files are ordered alphabetically by displayed country name: a new rule block goes at that position in 00_game_rules.txt, its loc block at the same position in MD_game_rules_l_english.yml, and a country sub-rule (BLR_union_state_ai_behavior is the shape) sits directly after the country's main rule in RULE_GROUP_AI_BEHAVIOR.

Wiring. Rule → set_global_flag = TAG_<PATH>_FOCUS_PATH in 999_game_rules_on_actions.txt. RANDOM_PATH's random_list includes the historical bucket; NO_PATH gets no branch. Convert country flags to global. Gate on has_global_flag, never has_game_rule, everywhere including events and strategy plans — otherwise a RANDOM roll enables the flags but not the plan. Verify NO_PATH leaves a working AI: an unconditionally-enabled strategy plan and a sane focus ai_will_do base.

Historical government. Read the report's government section. On a dated timeline verdict, write the walker (references/write.md §8) and its 00_yearly_effects.txt schedule lines, so historical AI delivers the historical head of government and not just the historical party. On an undated successor roster, write nothing and say so in the PR — the ramp decisions already deliver the party, and a walker over an undated roster installs the wrong person. Never pass change_leader_temp = 1; never inline create_country_leader.