Full RIMAPI integration: fix data gaps, alerts, incident control, scenario pipeline

[jkbennitt]

Summary

Wires every useful RIMAPI endpoint into the RLE client and fixes the data gaps that were hurting agent decisions. Makes partial progress on #7's save creation pipeline by rewriting it to use RIMAPI (replacing XML surgery), but does NOT fully close #7 — see "What's still needed for #7" below.

• Phase 1 — Fix data gaps: steel, wood, components, power_net were hardcoded to 0 in get_resources(). Agents skipped research benches because they thought they had no wood. Now wired to /api/v1/resources/stored and /api/v1/map/power/info. Also wired faction goodwill and fixed the equip/repair_rect/destroy_rect actions that were in the catalog but silently failing.
• Phase 2 — Agent awareness + deterministic scenarios: alerts endpoint (/api/v1/ui/alerts) gives agents structured problem signals like "Need Beds" / "Colonist Starving" with target IDs. Incident trigger lets scenarios fire ToxicFallout/RaidEnemy/Plague deterministically instead of hoping the storyteller cooperates. Screenshots opt-in via screenshots_enabled.
• Phase 3 — Spawn + setup: spawn_pawn, spawn_item, send_drop_pod, change_weather plus SetupCommand in scenario YAMLs. Rewrote create_scenario_saves.py — 200+ lines of XML regex surgery replaced with declarative RIMAPI calls.
• Phase 4 — Pawn edit helpers: edit_pawn_skills, edit_pawn_traits, edit_pawn_health, edit_pawn_needs for scenario customization.

What scenario saves got built (and what they still lack)

All 5 advanced saves were rebuilt via RIMAPI and committed to docker/saves/. They have the item/pawn/incident state from setup commands, but they're still day-0 snapshots because RIMAPI has no endpoint to fast-forward the game clock:

Save	✅ Got	❌ Still missing vs #7 spec
first_winter	Medium difficulty	Not advanced to day 30; no shelter/food/research buildup
toxic_fallout	+1 pawn, wood/meals/steel, ToxicFallout active	Not advanced to day 10 stable colony first
raid_defense	+2 pawns, weapons, flak vests, resources	Not advanced to day 15; no walls/sandbags built
plague_response	+2 pawns, medicine, meals (Plague fires at scenario runtime)	Not advanced to day 10; plague hediffs not pre-applied
ship_launch	+2 pawns, high-tier resources (plasteel/gold/uranium)	Not advanced to day 60; research still mostly incomplete

What's still needed for #7

This PR delivers the client-side infrastructure for save creation but not the advanced colony states the issue spec originally called for. The remaining work needs either:

1. A RIMAPI tick-advance endpoint — fast-forward N ticks, then save. Would unblock "day 30" / "day 60" scenarios in one API call. File as feature request upstream.
2. Orchestrated in-game play — run an agent colony for N real-time ticks, then save. Slow but uses existing APIs. Could be a new script.
3. Manual dev-mode setup — keep the original manual workflow for these 5 saves as a one-time thing. Not sustainable.

set_research_target(force=True) only queues a project, it doesn't instantly complete it — noted in the create_scenario_saves.py comments so future sessions don't try to reuse that approach for advancement.

How this moves the needle on #6

Problem (from #6 comment-4159780616)	Fixed by
Mood 0.40 — agents don't prioritize beds/cooking	Phase 2a alerts expose "Need Beds" / starvation with target IDs
Research 0.226 — bench rarely placed	Phase 1a real wood/steel counts unblock construction decisions
Need harder scenarios where baseline struggles	Phase 2b deterministic incident triggers + incident prediction

Notable discoveries during live testing

• RIMAPI's save_game() returns before Unity flushes the file to disk — without polling we got 67 KB truncated writes. Script now polls file size until it stabilizes above 5 MB.
• RIMAPI's spawn_item can't auto-split amounts above max_stack — it null-refs and destabilizes the game. Script splits into MAX_STACK[def_name] chunks.
• Back-to-back load_game() calls fail — previous map needs time to tear down. Script adds 8s between scenarios and 10s after load before first write command.
• equip was in the agent system prompt + WRITE_CATALOG but had no client method or executor handler. Agents proposed it and it silently fell through to generic dispatch with wrong params.

Test plan

• 376 tests pass (added 28 new: Phase 1-4 coverage + review-gap tests + perf regression for the power_info double-fetch)
• mypy src/ strict clean
• ruff check src/ tests/ scripts/ clean
• Smoke test runs all 6 scenarios without error
• End-to-end live verification: all 5 scenario saves built via RIMAPI, reload cleanly, have expected populations (4/5/5/5/3) and wealth increases matching spawned items
• Integration tests for _fire_scheduled_incidents with MockClient assertions

🤖 Generated with Claude Code

[jkbennitt]

Self-review + fixes

Ran a review pass on this PR (since I wrote it all, wanted critical eyes). Found 7 issues ranging from a real perf bug to convention nits. All addressed in commit 851bc4c (last commit on the branch). 7 new tests added as regression coverage.

High/medium issues — fixed

Severity	Issue	Fix
HIGH	get_power_info() called twice per tick (once inside get_resources(), once in get_game_state()) — doubles HTTP latency for power data	Refactored: get_resources(power_info=None) accepts an optional pre-fetched PowerData. get_game_state fetches once, passes through. New test test_power_info_called_once_per_tick asserts call count = 1.
MED	.get(k) or default anti-pattern — would erase legitimate 0 / empty values if upstream ever sent them	Added _pick() helper using explicit k in data membership. Replaced pattern in get_power_info, get_factions, get_alerts, get_resources_stored.
MED	change_weather used unescaped query string (?name={weather_def}) — inconsistent with body-based convention elsewhere, unsafe if weather_def ever has special chars	Uses JSON body now. test_change_weather_sends_json_body verifies shape.
MED	first_winter scenario was a day-0 copy of base save, not the day-30 state the plan called for	Documented in SCENARIOS dict with explanation: RIMAPI has no tick-advance endpoint, set_research_target(force=True) queues but doesn't complete research. Not silently shipping as unchanged.
MED	create_scenario_saves.py and run_scenario.py reached into client._get(...) (private method)	Added public client.ping() -> bool. run_scenario.py now uses client.get_colony() for the load-ready poll.

Low-severity — fixed

Issue	Fix
trigger_incident silently sent map_id as string	Docstring now explains it's because TriggerIncidentRequestDto.MapId is typed as C# string.
Hardcoded COLONY_X, COLONY_Z = 132, 137 in create_scenario_saves.py would silently break if base save is regenerated	_compute_colony_center() reads live colonist positions (centroid), falls back to constants.

Not addressed (deliberate)

• take_screenshot manual envelope unwrap — _get auto-unwraps {"data": ...} but _post doesn't, so take_screenshot does result.get("data", result) manually. Fixing this requires changing _post semantics across ~25 write methods that rely on the raw envelope (result["success"] is True). Tracked as a separate followup — too risky to bundle here.
• .rws files committed to git — deliberate trade-off for Docker benchmark reproducibility. Noted in PR description.

Test coverage added (7 new, total 376)

• test_change_weather_sends_json_body
• test_trigger_incident_nested_parms (raid_strategy, arrival_mode)
• test_flat_list_response (get_resources_stored handling both response shapes)
• test_ping_returns_true_when_alive
• test_ping_returns_false_on_connection_error
• test_power_info_called_once_per_tick (perf regression)
• test_fire_scheduled_incidents_swallows_exceptions (game loop robustness)

Verification

• 376 tests pass
• mypy src/ strict clean
• ruff check src/ tests/ scripts/ clean

🤖 Generated with Claude Code

[jkbennitt]

Pushed f377487 — unified the RIMAPI envelope unwrap between _get and _post.

The trap: _get unwrapped {"success": bool, "data": ...} → data, but _post returned the raw envelope. take_screenshot worked around this with a manual result.get("data", result). Future write endpoints that return payloads (e.g. spawn_pawn → {"pawn_id": 999, "name": "Val"}) would hit the same gotcha.

Why changing _post in place was safe (verified before picking this over a parallel _post_unwrapped helper):

1. Grepped ["success"] / .get("success") repo-wide. Only hits: 30 test assertions and one unrelated event_log.py line. Zero production callers inspect success on a write method — action_executor.py:94, run_scenario.py, run_benchmark.py, create_scenario_saves.py all discard the return.
2. The unwrap is conditional on "data" in body. {"success": true}-only responses pass through unchanged, so the 29 existing result["success"] is True assertions stayed green without edits. Only test_spawn_pawn (the one write test that mocks an actual data payload) needed its assertion updated.

Changes

• src/rle/rimapi/client.py: extracted _unwrap_envelope() helper, applied in both _get and _post. call() now returns Any and routes both verbs through unwrapping primitives. take_screenshot loses the manual unwrap. Write methods' return types widened dict[str, Any] → Any to match.
• tests/unit/test_rimapi_client.py: new TestEnvelopeUnwrap class (cross-verb consistency + no-op-when-no-data-key); updated test_spawn_pawn to assert on the unwrapped payload.

Verification: pytest (378 passed), mypy src/ strict (clean), ruff check (clean).

[jkbennitt]

Self-review pass — found 6 issues, all fixed in 26d6bcf:

Bug

• Duplicate _wait_for_save_written in scripts/create_scenario_saves.py — second definition was shadowing the first. Deleted the redundant copy and kept the one using module constants (SAVE_WAIT_TIMEOUT, SAVE_MIN_SIZE_MB).

Minor

• _h_repair_rect / _h_destroy_rect now coerce coords to int via int(params.get(...)), matching the pattern in _h_move. LLM output can surface as strings.
• _h_equip now checks thing_id is None instead of falsy — thing_id=0 would have been silently skipped.
• get_resources_stored logs a one-line warning when items are missing def_name. Previously they were silently dropped, which would mask upstream RIMAPI schema drift. Added logger = logging.getLogger(__name__) to client.py.

Nits

• spawn_pawn docstring now calls out the nickname → nick_name / bio_age → biological_age / chrono_age → chronological_age wire-name remappings so callers don't need to read the source.
• _default_rimworld_save_dir() now chains path segments via / instead of embedding / inside strings.

Verification: ruff clean on the three changed files, mypy clean on client.py and action_executor.py, 79 unit tests pass (test_rimapi_client.py + test_action_executor.py).