Status

Checked in the Live Game

Reader

• Gold, population, popularity, 20 stored goods and own-troop totals by type match the game's panels.
• Date, housing, idle peasants, rations, popularity factors, selected-building workers, placement mode and camera position match the game.
• Storage, on Oasis by the Sea on 2026-10-09: each stockpile pile's good, amount and capacity and each granary's food and capacity, read from the engine's building records, followed purchases and sales at the market. Buying stopped when a good's piles were full and no pile was empty (wood at 140 with stone in the fourth pile; iron refused outright), with no text message from the game; an emptied pile took the next good. The model's storage summary showed the stockpile full at that point.
• Popularity readings against the game's panels on 2026-10-10 (a scratch reader build, on the reference games' end saves and agent saves): the keep panel's tax figure (tax_amount) at every level for 202 people, 0.2 gold a person per step from Low (121) to Crueler than cruel (404) and the bribes' −121, −161 and −202; the rations factor at all five levels (−8 to +8), with the food factor equal to the rations and food-variety factors added up; worker efficiency (140% with the fear factor at −4, 110% at −1). The game's religion line was 3 points above the reader's religion factor in both reference games, which have a church and a cathedral: the bonus the reader reads as church_bonus. Read from the committed reader on 2026-10-10, it is 75 on both end saves (the factors' 25 per point) and 0 on a save with only a chapel; the factors in the model's stats then add up to the game's coming change on all three.
• The player's buildings by type (buildings), from the committed reader on 2026-10-10: on both reference games' end saves they match the count made earlier with a test build (57 bakeries, 24 hovels and 54 bad things on Oasis by the Sea; 95 bakeries, 11 inns and 21 bad things on Large Barren Desert), a save with only a chapel lists it, and each building placed in the tool checks below appeared under its type at once. The keep panel's tax figure read 150 gold a month at Low taxes for 250 people.
• Placement and resource warning messages are captured and reach the model with their age.
• Sampling every 100 ms has no measurable cost to the game.
• The collector watchdog against a real collector stall, in a 25-game-minute GLM 5.3 Flash episode on 2026-10-10 (benchmark 1.1.4): the collection count stopped at 334, and after 20 seconds the reader forced a collection (25 ms). The count then moved on by itself every few seconds, and the game's memory stayed at 3.1 GiB. The collector's variables confirmed the cause: the thread-stack total had wrapped around (2^64 minus about 2.9 GB, so a thread's stack pointer read about 2.9 GB above the base of its stack), the threshold for the next collection was 3.1 × 10^18 bytes, and collection was neither disabled nor in manual mode.
• The same watchdog against a simulated stall, on a paused Oasis by the Sea on 2026-10-10. A test build switched off automatic collection (Unity's manual mode), so the collection count stopped and the heap's used bytes grew 7.5 MiB a second, as in a real stall. After 20 seconds the reader forced a collection, which took 23 ms and brought used bytes back from 499 to 364 MiB, and it did so again every 20 seconds until automatic collection was switched back on; the heap grew from 516 to 548 MiB in all. The collector variables it recorded were sensible: a 23 MiB threshold (the game allocates about 24 MiB between collections), 162 KiB of thread stacks, and manual mode on.

Control

• Clicks, right clicks, drags (selecting units), mouse-wheel list scrolling, keys, P (for the host), camera zoom with Z and X, and game speed with + and -.
• The game-window capture excludes windows covering the game.
• Stop from the dashboard and from the control monitor, and the focus check that blocks input while another window is active.
• The run overlay, shown with a test status: on the screen above the game, absent from the game-window capture the agent sees, never active, with an empty input region, and with keys from the host still reaching the game. Not yet watched through a model's run.
• The runner's overlay feed for a round on an earlier version, on 2026-10-13 through Muse Wonder's run on Oasis by the Sea at 1.5.1 (checked over SSH, not on the screen): this checkout's overlay.py copied to the game machine before the run, one feed and one overlay process running, and the feed's clock, tool calls, cost and figures under a second old. A feed left by a start that failed earlier the same day was still running with its runner gone; feeds now exit after 60 seconds without a line.

Tools

• market_trade in the game's own steps (1.5.0), on 2026-10-11. Clicking the marketplace through the host on a saved game: a purchase is 5 units a click for every good tried (wood, apples), and a sale is 20 units while 100 or more are in stock, 10 from 50, else 5 (wood, stone and apples; 103 sold 20 and 91 sold 10, 53 sold 10 and 46 sold 5); purchases land about 0.45 s after their clicks at any pace and at the turn of the month. Then four 3-game-minute GLM 5.3 Flash checks from the 1.4.0 episode's end save, each told to make the same trades; the last, on the final code, matched the game in every one: 20 lots of 102 stone sold exactly 100 in 13 clicks (gold +700); 5 lots of 120 apples sold 20 and stopped, saying the next click would sell 10 of the 5 left; 20 lots of wood bought 100 in 20 clicks with the month's taxes landing among them (gold −363, that is −400 and +37); 10 lots of bread bought 50. The earlier checks had each missed one purchase that landed in the very reader sample of the taxes and a delivery. Before the fix, two runs that day sold 105 wood for 10 lots (494a5f78) and all 225 for 20 (cdecaf85). The simulated market in economy.test.ts covers the same cases and a purchase the game refuses while planks are being unloaded.
• build_structure placement, its verification by cost and structure count, the camera guard, and retries at nearby spots.
• center_on, place_near, expand_storage, flat_view and saved views.
• expand_storage anchored on the stockpile's piles (2026-10-09): stockpiles placed on the up-right, down-left and down-right sides at the first try (up-left, the keep's side, fell through to the next side), and two granaries beside the granary. Before, it anchored on the keep beside the stockpile, and in every recorded run only spots on the stockpile's down-right side were placed. market_trade names gold or storage room when a purchase stops.
• map_overview, go_to_tile, set_tax and market_trade have been used in live runs; marketplace prices on screen match the price table.
• Run against the live game through the harness's own tool code, without a model, on 2026-10-10 (benchmark 1.2.0). set_tax on the Large Barren Desert reference game (250 people): Low to Average returned 200 gold a month, to the generous bribe −250, back to Low 150, with the popularity effects −4, +7 and −2, and the keep panel closed each time. set_rations from none: half, extra, double, full and none returned the ration factors −4, +4, +8, 0 and −8 as the reader showed them; asking for the current rations clicked nothing; the granary panel closed each time. build_structure on "Oasis by the Sea benchmark start" placed a dancing bear, gallows, a hovel straight after the gallows (so the tray left the fear-factor page), a town garden, a dungeon, a shrine and a woodcutter, each confirmed by its cost or the structure count; the town garden's first two spots, offered by find_sites, were silently refused and it went 3 tiles over. With three good things and two bad ones in a settlement of 10 people, the fear factor read +1 and worker efficiency 90%: good and bad things offset each other. inspect_building opened a panel only with the game running, as it is during play.
• Run against the live game through the harness's own tool code, without a model, on "Oasis by the Sea benchmark start" on 2026-10-12 (the changes of 1.6.0):
  • build_structure placed an Armory (from the Castle page), a Fletcher's, a Poleturner's and a Blacksmith's Workshop and a Marketplace, each confirmed by one more of its type in the reader's buildings. A tile read showed each footprint on exactly the tile asked.
  • New workshops make bows, spears and swords. Each panel's left button sets bows, spears or swords and its right one crossbows, pikes or maces. The building record holds the product in progress at +0x1bc and the next at +0x1be, as goods numbers 17 to 22, which the reader now reports as selected_workshop. set_workshop set a Fletcher to crossbows from a corner tile of its footprint and read it back; inspect_building later showed it making crossbows. Asked for swords there, it answered wrong_workshop; given the Armory, not_a_workshop. In the first of four calls in a row no panel opened, so the click is now tried twice.
  • A 10-game-minute GLM 5.3 Flash trial on 1.6.0 (unversioned, run d41d24dc) played valid with no tool errors. It placed an iron mine that find_sites had found out of view (moved_to_tile), an Armory, and a Fletcher's Workshop that was moved one tile off it (touches_storage). It made 2 bows and a sword. Its set_workshop on a Fletcher opened no panel in two clicks; on its end save the same call worked at every try. A panel shows in the reader about 0.6 s after the click, so each click now waits 4 s and a reader without readings is reported as such.
  • Armories, on "The Sinkhole start" on 2026-10-12: one holds 50 weapons and armour of any mix, counted singly. Buying a lot of each of the eight types stored 40; 20 spears and 5 of six other types filled it at 50 (the record's amount and capacity at +0xb2 and +0xb6 read 50 and 50); then the market refused another lot of spears with gold to pay. GLM 5.3 Flash's armory held 11 bows, 25 spears, 11 swords, 2 leather and 1 metal armour at 50 and 50 too, and a person's test with 43 swords and 5 leather armour got "No space in the armory." for 5 more. An empty armory reads 0 and 0. A second armory away from the first was refused: "Needs to be placed adjacent to the armory." No game message reached GLM while its 14 workshops stood at a full armory: the game says "The armory is full, my lord" aloud (a person's observation), and spoken lines without text are not read, so the stats' full note stands in for it.
  • The 1.6.2 reader on a person's end save with 45 leather armour and 2 swords: the stats showed the armory at 47 of 50; with the game running, the blacksmiths' swords took it to 50, where the full note appeared, and market_trade then bought no swords and named expand_storage. In that crowded town the only two spots touching the armory also touched other buildings; expand_storage (kind armory) placed a second armory on one of them, and the stats read two armories with room 49.
  • A click on a building's middle can open a neighbour's panel: a Hovel in front of a Blacksmith is drawn over the Blacksmith's middle. set_workshop now checks the opened building against the tile map and then clicks the back corner of the footprint, which opened the Blacksmith.
  • Choosing a Stockpile, Granary or Armory in the menu centres the view on the first one of its kind. In placement mode the minimap does nothing, while going to a game bookmark works and keeps the building selected. A Stockpile given by a tile 40 tiles from the first came back to its tile through bookmark 9 (moved_back_to_tile), and the game then refused it for adjacency, as it should.
  • Tile targets off screen: the camera moved there with one minimap click and the building was placed (moved_to_tile).
  • Terrain height: a Hovel given by a plateau tile at height 130 was refused at its flat position (which showed the cliff below it) and four retries. The sixth click, 15 rows higher, placed it on the very tile asked, while the tool reported another. With the height-aware projection, a second Hovel on the plateau went onto its tile at the first click, and one on low ground (height 8) did too.
  • market_trade bought 20 wood and reported goldChange −80 at the market's price.
  • find_sites with tile_x/tile_y listed spots around tiles out of view, and build_structure placed there.
  • The 1.6.2 round on The Sinkhole, seven models, 2026-10-12/13: all seven played the full 45 game minutes and every reported placement matched the game's final count. expand_storage (kind armory) placed second armories in three runs and gave the no-room remedy in one; set_workshop set fletchers, poleturners and blacksmiths by tile and by pixel, though a building in front (three times the Armory, once a Hovel) still opened instead four times. Seen live and fixed in 1.6.3 (below): place_near could not place a first Armory (three calls threw "Game image expired" for Claude Haiku 5.5, one for Kimi K3, and GPT-6.1 Sol got 20 probes and "No room around your armory"); apple purchases stopped three times as not shown in the reader, and in the one checked (Claude Opus 5.5) the reader had shown it, with apples eaten and part of the taxes in the same sample; GLM 5.3 Flash's end-of-run save typed only a space as its name. The other six saved normally.

Runs

• The prompt cache key for Meta's models, on 2026-10-13: Muse Wonder's run on Oasis by the Sea at 1.5.1, its host given the key by round-preload.ts, read the cache on all 106 requests, 96.9% of its input tokens. Its run the same morning without the key missed on 3 of its first 11 requests (67% of input tokens cached).

• The Large Barren Desert scenario's map facts, on 2026-10-11, from "Large Barren Desert benchmark start" loaded without an agent: 5,000 gold, 192 wood (the stockpile's four piles of 48 full) and 55 meat about 27 game seconds after load, with the rest of the package listed at the top left as still to come; the granary and stockpile already standing; and from the whole-map summary no oasis grass, stone about 100 tiles west of the keep and iron ore about 60 east. The summary's 154 "tree" tiles were cacti: two woodcutters of Claude Haiku 5.5 placed among them delivered no wood in 30 game minutes. In a GLM 5.3 Flash run the rest (50 stone, 15 iron, 20 pitch, 20 ale) arrived within 90 game seconds, once building had used some wood; the total wood (about 245) is read from the screen's list, not counted.

• Placing next to the desert stockpile, on 2026-10-12, by replaying Kimi K3's clicks on "Large Barren Desert benchmark start" through the host without an agent: with the camera on the stockpile (tile 506,373), the first flush Marketplace click built it on the stockpile's up-right side, and the reader's own-building counts showed it (type 26) at the next check. The tile map then had it and offered expand_storage two flush spots (tiles 513,375 and 508,380). Kimi's run had clicked pixels that match those spots seen through a camera one tile away, and the game refused both without a message. Clicked again here, the first of Kimi's pixels placed a pile and the second was refused, so a one-tile error can go either way; the map's second spot then placed a pile (its first was covered by the new pile by then). Kimi's recorded frames show that choosing the Marketplace again with one standing centres the view on it, and that later Marketplace clicks placed nothing.

• Ox tethers away from the quarries, on 2026-10-12 (Kimi K3, benchmark 1.5.5, run 8ef9720b, stopped at 9 game minutes): five tethers beside the stockpile, about 100 tiles from three staffed and working quarries, delivered no stone in over 7 game minutes; the oxen stood at their tethers. In every other run the tethers stood 4 to 12 tiles from a quarry and stone arrived. Seen in that run, not tested apart; from 1.5.6 the prompt says to build tethers beside the quarries.

• New peasants stop arriving while 24 stand idle: in the 1.5 runs on both maps the population stopped with 24 idle, and grew again only when jobs took some of them (Claude Sonnet 5.5's four iron mines brought 8 people on Large Barren Desert, GPT-6 Luna's quarries 5 to 11 each time). Seen in runs, not read from the game; the prompt says it from 1.5.6.

• Timed runs with game-time budgets, pausing during thinking, compaction and the final pause, with several models including Kimi K3 and GLM 5.3 Flash.

• Unattended episodes end to end: launch, load by name, run, scorecard, close.

• The 21 runs in the evaluation (seven models on each of Oasis by the Sea, Large Barren Desert and The Sinkhole, benchmarks 1.5.0 to 1.6.2, 2026-10-11 to 2026-10-13), from their event logs:

  • All played their whole budget (33⅓ game minutes, 45 on The Sinkhole) except GLM 5.3 Flash's on Oasis by the Sea, which four replies without tool calls ended about 5 game seconds early (from 1.6.0 the rest of the budget passes instead). None was cut short by the game.
  • The host's game-speed step ran before every budget and measured 39.4 to 40.4 ticks a second. The saves already run at 40, so it pressed no key.
  • observe waited out the rest of the minimum turn 31 times in 14 runs and said so in its result. Every reply's last tool result carried the "Game time left" line.
  • The collector watchdog broke 16 stalls in 12 runs. Each showed the wrapped thread-stack total of the Wine bug (see Reader), and each forced collection set the collection count moving again.

• The results and the interview, on 2026-10-10: three dry runs of "Oasis by the Sea benchmark start" wrote their scorecards and results in folders of their own, and a 3-game-minute GLM 5.3 Flash episode (an unversioned trial) wrote its results, was saved and closed, and was then interviewed. The model answered all four questions in 1 min 40 s without calling a tool; 96% of the interview's 152,164 tokens were read from the provider's cache, and it cost $0.006, kept in interview.json. The run's own cost in run.json stayed the sum of its gameplay and preparation requests, and events.jsonl has no interview events. The first question said the budgets were not like for like (3 minutes beside the person's 33).

• Run recording and video rendering, with scripted inputs and in a 2-game-minute GLM 5.3 Flash episode on 2026-10-08.

• On a paused game: minimap clicks and Z move and zoom the camera, and the keep hotkey does nothing. The final overview's camera sequence (minimap to the keep's tile, then Z until the view stops widening) was run by script on 2026-10-08, and at the end of real runs since: five GLM 5.3 Flash runs on Oasis by the Sea (2026-10-10 and 2026-10-11) each recorded the paused minimap move to the keep as arrived.

• Acting tools' own steps taking less game time (see the 1.3.0 changelog), in a 15-game-minute GLM 5.3 Flash run on Oasis by the Sea on 2026-10-11 (run 0b08f7e0, unversioned because of uncommitted changes). All five go_to_tile calls came first among their reply's acting tools; in each, the minimap clicks came before the unpause (recording/inputs.jsonl), none of the five screenshots showed the "Game Paused" overlay, four arrived and one stopped with the tile in view, and no pause toggle was retried. Their time to the first input fell from 1.28 s to 0.40 s on average. The one expand_storage call placed both stockpiles. Its tile read overlapped the building selection but took about 4 s, against about 3 s in earlier runs, so the first terrain click still came 2.8 s after the selection (3.15 s before); one call cannot tell whether the overlap slows the read. Reusing the latest capture applied in 4 of 15 calls, saving about 0.25 s each: in the rest it was over 20 s old, because text-only observations take no screenshot.

• Benchmark 1.4.0 beside the published 1.1.0 run, on 2026-10-11: one 25-game-minute GLM 5.3 Flash episode (series 20261011T114226-a46cbd06, run cdecaf85) with the settings of episode 1 of the published series 20261009T122236-3aae6fc4 (run 171a1a71): the same save, objective text, game speed 40, budgets and model profile, but no playbook (the published series had one, still empty in its episode 1). Both were valid and used the whole budget. Against the published episode: cost $0.129 against $0.207, with 109k tokens of uncached input against 714k (97% of the input read from the cache); the agent's 25 game minutes took 34 min 53 s of real time against 41 min 06 s, and the agent started 41 s after the series did against 2 min 59 s; the median wait for a reply's first token was 8.6 s against 10.7 s, and inference 871 s against 1,126 s over 80 turns against 84; acting tools took 204 s against 428 s for about the same number of clicks (about 230), the game machine's tile reads 9 s against 70 s; one compaction against two; the host sent 6 screenshots and 47 text observations against 57 screenshots. The first gameplay request read 19,712 tokens from the cache and 1,732 uncached, after a preparation request of 19,718 that called no tool. Net worth was 6,231 against 4,748 (popularity 79 against 64); one episode each, and single GLM episodes vary widely, so the score difference is not established. The stockpile ended with 3 of 12 piles empty. With 5 game minutes left the agent saw wood room 111 (two empty piles counted in it) and sold wood anyway to keep piles free for iron and stone; it took all 225 (see Known issues).

• The current harness and prompt, in two 5-game-minute GLM 5.3 Flash episodes on 2026-10-08 that completed: the compact observation summary, reading tools at the start of a reply running while the game stays paused, the dashboard fetching a screenshot only when the run has a new one, and the model settings and harness version recorded in run.json. The second ran with a 60,000-token budget and compacted twice; each handoff request repeated the gameplay request and was served about three-quarters from the provider's cache.

• The minimum turn length (8 game seconds), in a 2-game-minute GLM 5.3 Flash episode on 2026-10-08: every turn that ran the game took at least 8 game seconds, short turns were topped up with the model told why, and reading-only turns stayed free.

• A learning series of three 3-game-minute GLM 5.3 Flash episodes on 2026-10-08: each episode started from the playbook the one before left, the agent changed it during play, the request after each episode rewrote it, and the series records and report table were written. What OpenRouter billed was recorded for every request, about $0.04 per episode.

• The fixed game speed, set before a 25-game-minute idle baseline on 2026-10-08: the save started at 19.9 ticks a second, and four presses of + (5 ticks a second each) brought it to 39.9. Doing nothing then scored 1,304 net worth; the game placed the granary by itself.

• Full-length runs: a learning series of three 25-game-minute GLM 5.3 Flash episodes on 2026-10-10 at benchmark 1.1.4 (run unversioned, before the change was committed). All three ran their whole budget and were scored valid (net worth 5,032, 3,021 and 5,038), with the game's memory at 3.1–3.3 GiB throughout. The third had the collector stall described under Reader, which the watchdog broke.

• Stopping and resuming a series, in a learning series of two 1-game-minute GLM 5.3 Flash episodes on 2026-10-08 at benchmark 1.0.0: Ctrl-C during episode 1 stopped the dashboard run, closed the game and paused the series within 2 seconds; --resume ran episode 1 again as attempt 2 and then episode 2 from its playbook, and the report counted episode 1 by its second attempt. Every run recorded its version, attempt and a clean commit, and the runner's check of the game machine's files caught an out-of-date one before the run.

• The automatic re-run after an infrastructure failure, in two series on 2026-10-09 (20261009T122236-3aae6fc4 and 20261009T164620-3dd2f4e8): one series' episode 3 failed on attempt 1 and was valid on attempt 2; the other's episode 2 failed on attempts 1 to 4 and was valid on attempt 5. Each time the runner started the next attempt from the same playbook on its own, without pausing the series.

• Claude through Anthropic's own API, in a learning series of two 3-game-minute Claude Haiku 5.5 episodes at medium effort on 2026-10-09 with a 75,000-token context budget: both episodes ran their full budget; all 27 buildings it tried to place were placed (3 at a nearby spot after a retry), as expected since Anthropic's documentation says these models see a 1920 × 1080 screenshot unscaled; compaction and the playbook request worked; and episode 2 started from the playbook. The cache marks paid off: each request read the conversation the previous one wrote (70–72% of all input from the cache), and the handoff and playbook requests read it too. Anthropic reports no cost; the costs of that day's ten Claude runs were worked out afterwards with npm run cost from their recorded tokens at Anthropic's list prices (about $0.60 in all; $0.31 for the three published episodes of series 20261009T100849-21b4ea9d, $0.38 with the attempt that was run again), not checked against Anthropic's billing. A 60,000-token budget is too small for Claude: its fresh conversation after a compaction (about 39,300 tokens) did not fit under 65% of it, which ended that run with an error.

• The game on the test laptop's NVIDIA GPU, on 2026-10-13 with Ubuntu kernel 7.0.0-38 and NVIDIA driver 610.57.04, the harness disconnected except where named: it reached the main menu with DXVK (three launches on Proton Experimental 11, one on Proton 9.0-4) and with WineD3D/OpenGL, and nvidia-smi listed it as a graphics client with about 3 GB. A dry run loaded "The Sinkhole start" 52 seconds after launch, and a 45-game-minute do-nothing run on it scored 2,195 at 39.9 ticks a second, as on the Intel GPU the day before. In September, with kernel 7.0.0-31, the game stalled at the end of loading on the NVIDIA GPU with both APIs, on Proton 9.0-4 and Proton Experimental alike, with or without the reader. The cause was not found. Proton is ruled out (the same Proton 9.0-4 build stalled then and loads now), and so are the game, the NVIDIA driver, Mesa, Mutter, Xwayland and the Steam client, which did not change in between. The kernel update is the likeliest fix; the only other change near the display path was FreeRDP, which GNOME's remote desktop uses.

Tested in Code Only

• The shims that bring two harness changes to rounds played on earlier versions (cache-key.test.ts, overlay-feed.test.ts, test_overlay.py; both also used in a run, above): the key added to the harness's own requests as Pi sends them and by the preload through an earlier host's global fetch, with none for other models; the overlay's figures read from an earlier host's state, its clock run on from the game tick between that host's updates, tool calls counted from the run's events, and the feed's figures used only while fresh.

• The 1.6.3 fixes (model-observation.test.ts, remedies.test.ts, economy.test.ts, game-save.test.ts), not yet run live:

  • place_near placing the first Armory (or first of any storage kind) against its anchor from the tile map, and the no-room remedy kept for a storage's own anchor.
  • place_near and expand_storage sending clicks through a frame captured afresh once the one they started with is 20 s old.
  • market_trade counting a purchase when the stock rose and the gold fell in one sample.
  • The end-of-run save trying once more when no save under the new name appears.

• The 1.6.0 changes not yet run live (model-observation.test.ts, tile-map.test.ts, economy.test.ts, controller.test.ts):

  • find_sites searching around the nearest ground of a building's own kind on the map when the view has none.
  • The one tile of open ground kept around granaries and armories (access: touches_storage).
  • demolish_building moving the camera to a tile target off screen.
  • The right-click that leaves placement mode before a later placement's camera move.
  • market_trade waiting 5 seconds for a trade to show.
  • Four replies without a tool call in the last game minute letting the rest of the budget pass.
  • None has been used in an agent's run yet.

• The run overlay on the game machine (test_overlay.py, test_control_status.py): what it shows from the bridge's status and how it draws it, and the host's status carrying the session's game clock, tool calls, net worth, popularity and cost.

• The placement fixes of 1.5.5 (construction-ui.test.ts, model-observation.test.ts, building-types.test.ts): one more of the player's own buildings of the placed type confirms a placement over a refusal message; a second Marketplace is refused before any input; place_near aims its tile-map spots through the live camera and probes the view when every spot it gave was refused without a message; build_structure's closing right-click stays on screen when selecting moved the view. Not yet run live in an agent's run; that the own-building count changes within the 120 ms settle sample, as the structure count does, is assumed from both coming from the same reader sample.

• Trees told from cacti by the ring of 0x2000 tiles a tree has around it (tile-map.test.ts for find_sites' tree counts), from tile data read live on 2026-10-11: on Oasis by the Sea every trunk tile on a palm had a ring tile within 3 tiles and every one on a cactus had none, and Large Barren Desert has no ring tiles at all. The reader's map summary draws trunks without a ring as cacti (c) instead of trees.

• Clearing the Save dialog's whole name field when the click lands inside a long loaded name (game-save.test.ts), the cause of the failed saves on Large Barren Desert, where the game saved "glm 5-3 flash 1-5-3 ea2895rt". Since the fix, every desert run of 2026-10-11 and 2026-10-12 saved under its own name.

• A Moonshot profile's reasoning level sent as Kimi K3's reasoning_effort (low, high, max), with default still sending none and other levels refused (model-config.test.ts, store.test.ts). Moonshot's guide lists those three values; not yet checked live.

• The 1.4.0 request savings (see the changelog), with a fake provider: the preparation request carries the gameplay tools and the first gameplay request repeats them (controller.test.ts, model-config.test.ts), a preparation reply that calls a tool is asked again, and the observation stats leave out goods with none stored (model-observation.test.ts). The enums' validation was checked against pi-ai's validator by script. The cache read it brings was seen live (see Runs above); the retry of a preparation reply that calls a tool has not been needed in a run.

• A good's room in storage counting the empty piles, and an empty granary adding 250 to the granary room (anchors.test.ts), from pile and granary readings shaped like those seen live on 2026-10-09 and 2026-10-11. One live run has played with it (see Runs above); an empty granary's 250 has not come up live.

• Acting tools' own steps taking less game time (see the 1.3.0 changelog), with fake game devices (controller.test.ts, anchors.test.ts, model-observation.test.ts): go_to_tile and navigate_minimap click the minimap paused only when nothing in the reply has unpaused the game yet, and add no pause toggle when it is already running. navigate_minimap has not been called in a live run since.

• Compacting a run's raw recording frames into recording/frames.mp4 (tools/video/compact-recording.py) and rendering from it instead of the JPEGs: with synthetic recordings (tools/video/test_compact_recording.py), encoding, ffprobe verification, a verification failure keeping the JPEGs, and a render from the compacted video matching the JPEG render (mean pixel difference under 6 of 255 on sampled frames, the extra JPEG → H.264 round trip's own loss). An interrupted render (SIGTERM or KeyboardInterrupt) leaving no partial file is tested the same way, and confirmed once with a real SIGTERM to a running render-video.py process. The 10–25× size and SSIM ≥ 0.987 in Run videos are from earlier offline encoding experiments, not reproduced by this test suite. Neither the compaction nor the decode-from-video path has run against a real recorded run or a live episode.

• The fixed game speed's key presses in an agent run, and its failures when the keys do nothing or step past 40. Live, the idle baseline pressed the keys (above), and every agent run so far started at 40 and pressed none.

• wait_and_observe waiting out the minimum turn length when asked to wait less than the turn still owes. No live run has asked for that yet; observe's wait has run live (above).

• Context compaction across many turns, stop and deadline handling, and failure paths, with fake models and game devices (npm test). This includes the retries added before the first 25-game-minute run: failed handoff and playbook requests, cut-off replies with tool calls, and unconfirmed pause toggles. The headroom check's switch to the same anchored estimate as the 65% trigger, against a fake model that reports a small usage figure over a long conversation, is tested the same way; it has not yet been seen to change when a real run compacts.

• Saving the finished game (see Running episodes), live on Oasis by the Sea (2026-10-09) through the same routine the runner calls: Game Options → Save, the name typed over the loaded save's, the file copied off the game machine (3.5 MB) and the benchmark save's hash unchanged; the game stayed paused. The name field takes 32 characters (a longer name was cut), and saves made this way load from the in-game Load dialog. Real episodes have since saved themselves at their end (GPT-6 Luna on 1.1.3, GLM 5.3 Flash on 1.1.4). The overwrite guard (copy, hash check, restore) is tested with a simulated save folder.

• Reference games (--human, see Scoring), live on Oasis by the Sea (2026-10-10) with a 30-game-second budget and the unpause sent as a player would: the runner paused the game after 31 game seconds, wrote the scorecard and a 1920×1080 screenshot, saved the game and closed it; the benchmark save's hash was unchanged. Not yet used for a full game.

• Open sides for buildings with workers (see Tools), live on Oasis by the Sea (2026-10-09): a woodcutter aimed at a 3×3 hole boxed in by four hovels came back access: "no_open_side" and was built 4 tiles away without the hole being clicked, and find_sites did not offer the hole. Placed there with exact, it warned, showed the game's no-entry sign, read "no access to keep" and got no worker while 19 peasants stood idle; the moved one was staffed and working. Hovels are engine type 1. Spot choice, the check before clicking and the background tile read are also tested with tile fixtures and the simulated game (npm test).

• Tool changes made before the first 25-game-minute run, not yet used live: woodcutter spots ranked by nearby trees, placement checks waiting for a valid reader sample, and cost in placement results.

• build_structure's per-placement checks, with fake game devices (npm test), not yet used live: a target outside the screenshot or over the bottom HUD reports invalid_target on its own (nothing sent, the rest of the batch still runs) instead of stopping the whole call; affordability is checked from the reader before each placement's first click, re-read so an earlier placement in the same batch is accounted for; and a quarry, iron mine or pitch rig refused for its ground is retried nearby on matching ground the same way a farm is retried on oasis.

• set_tax's popularity factor read live from the reader (popularity_factors.tax, the same points conversion as set_rations) instead of the static table, with a fake keep panel (economy.test.ts). The live run recorded above (benchmark 1.2.0) predates this change and shows the static table's values.

• The per-tool-call frame cache for input-only clicks and hotkeys (anchors.ts's frame(), reused for up to FRAME_REUSE_MS after a capture) and build_structure remembering the construction tray's open category across calls (falling back to the full category click when the reader does not confirm it), with fake game devices (anchors.test.ts, device-lifecycle.test.ts, model-observation.test.ts). Neither the 20-second reuse window nor the category memory's savings have been measured live.

• Targeting build_structure placements by map tile (tile_x/tile_y) instead of a screenshot pixel, and find_sites running without an observation at all (both read the live camera, waiting out an unavailable sample): a tile placement on screen, one off screen reported not_on_screen without sending input, one projected again when the view moves during selection, a batch mixing pixel and tile targets (only the pixel target's clicks keep the screenshot age limit), a blocked tile target retried at a nearby tile the map shows free (also after the camera moved since the last screenshot, and stopped when the live camera moves) or by fixed offsets reporting the tile it went to, find_sites giving pixels and using near_x/near_y only while a screenshot shows the live view, its map reused by the next placement's retry, and a shape error (neither or both coordinate pairs given) reported invalid_target; all with fake game devices (model-observation.test.ts). Not yet used live.

• wait_and_observe's screenshot: false (waits the same, no capture, the old screenshot still targets clicks) and the fallbackObservation setting (text after a turn with only text-only actions under when-needed; a screenshot after a game_action click or key, after a reply with no tool calls, or under always; the reading-only top-up; the briefing for each value; the default recorded in run.json), with fake game devices (model-observation.test.ts, controller.test.ts, protocol.test.ts). Not yet run live; how models use the text-only observation is unmeasured.

• Screenshots kept in the conversation until compaction (a batch prune only past 16), so every request repeats the previous one unchanged: checked across a fake run whose every turn brings a screenshot (controller.test.ts) and for the cap (context.test.ts); Claude's cache marks on where the previous request ended and the newest message (model-config.test.ts). The cost effect is not yet measured live; the 1.3.0 partial run of 2026-10-11 (only the two newest screenshots kept) showed the cache misses this removes.

• A remedy on every failed or partial tool result and fix-it wording in thrown errors (remedies.test.ts, economy.test.ts, model-observation.test.ts). Whether models follow them is not yet seen live.

• Recognising the main menu and the Load Game dialog from pixel colours during save loading: the classifier against colours measured on captures of 2026-09-28 (menu-screen.test.ts) and the sampler with synthetic images (tools/runner). Not yet used in a live launch; the seconds it saves, and whether the dialog is ready to type into a second after it is recognised, need one.

• Screenshots scaled down by the game-machine bridge with clicks scaled back up: live on 2026-10-11 through the host's own GameDevice, a 1280 × 720 capture of a 1920 × 1080 window, and a click on the main menu's Load Game at its 1280-wide coordinates opening the Load Game dialog. The window-coordinate scaling and the run applying the setting are tested in code (protocol.test.ts, controller.test.ts). No agent run has used the smaller screenshots yet.

• Clicking through the startup screens: live on 2026-10-11, clicking the screen centre once a second skipped the Firefly and Stronghold intro videos and the promo splash and brought the main menu 25 s after launch instead of 49 s, with nothing else opened; built into the runner after that and not yet used by npm run episodes.

• demolish_building: the demolish button (the X on the red strip beside the construction tray at 1537, 1016 in the 1920 × 1080 layout) putting the game in demolish mode (reader placement action 6), a click on a woodcutter's hut removing it at once with no confirmation (the map's structure count 68 to 67 within 1.5 s) and a right-click leaving the mode, checked live on 2026-10-11 through the host's input API. The tool's own statuses and remedies are tested with fake game devices (model-observation.test.ts); no agent has used it yet.

• find_sites grouping its spots by ground, and harness-chosen spots keeping off other buildings' ground (oasis grass, stone, iron ore, oil): tile-map.test.ts, model-observation.test.ts. Not yet used live.

• The construction guide sent as construction-trays.jpg when installed, else the PNG: file choice tested with temporary files (visual-atlas.test.ts). The JPEG at quality 90 was checked by eye on 2026-10-10 (labels readable), and runs on 1.3.0 used it. The 80% guide of 1.4.0 (1530 × 2270) was checked by eye on 2026-10-11: titles, button numbers and the 13-point notes are readable; run cdecaf85 used it.

• Good and bad things' footprints are from static analysis of the game files, the same method as the other buildings'; the town garden's refused spots (above) may mean its footprint is larger than recorded.

• Reading-only replies costing game time from the 12th in a row instead of ending the run, and the warning after each reply without a tool call. Not yet run live.

• Results written for a reference game or a do-nothing baseline by the runner: only dry runs and an agent episode have written them live. The two reference games' results were written afterwards from their recorded readings.

• The interview's refusal of tool calls, its retries and the 30-minute stop, with fake models (npm test): the live trial's model called no tool. The interview files in published bundles are tested in code only. The four questions asked as one combined request under titled headings, and the reply split back into four answers by them (or kept as one answer when the headings do not line up), is tested the same way; the live trial predates this and asked one question per request.

• Benchmark versions: the fingerprint, benchmark-versions.json and the test that enforces it. Resumable series: the episode outcomes, which attempts count, where a series resumes and the settings a resume must match (the automatic re-run after an infrastructure failure is confirmed live; see Runs above).

• A learning series starting the next episode's game (launch, connect, load, pause) while the previous episode's interview is still going, instead of waiting for it first (episodes.ts's prepareEpisode/abandonPrepared). episodes.ts drives a live dashboard and game machine through fetch and ssh and has no unit tests, so this is checked by reading the code, not run live or under npm test.

• Dropping a tool call's image from the two extra copies events.jsonl would otherwise keep (omitDuplicateToolImages, context.test.ts) and skipping image data in key redaction (store.test.ts, mirroring tools/dataset/scrub.py's own rule): the wiring that applies the first to a live run's event stream (index.ts) has no test of its own.

• Pricing requests on endpoints other than OpenRouter with the profile's prices (see Context): against fake Anthropic and OpenAI-format streams, including Claude Haiku 5.5's higher prices for prompts over 100,000 tokens, and the refusal to start a run without prices. Not yet used in a live run.

• Removing duplicate messages that appear in several channels at once.

• Prompt caching marks for Claude through OpenRouter: checked against the request body the model library builds, not yet against a live provider. The session-affinity header sent with every OpenRouter request, and the upstream provider name (upstream, e.g. "DeepInfra") read from a cost chunk when OpenRouter reports one, are checked against fake responses only (model-config.test.ts).

• Host shutdown ending a run with the final pause, and the host staying ready when a run cannot be set up.

• Publishing runs (Publishing): packaging, scrubbing, the tables and the upload checks are tested in code. The first series (GLM 5.3 Flash, benchmark 1.0.0, 20261008T191852-c1d986e9) was uploaded to the Hugging Face dataset on 2026-10-09, and its three tables load from the Hub with load_dataset. Making bundles smaller (images as WebP files, token deltas dropped, 720p video) is tested in code. Both series then in the dataset (that one, and Claude Haiku 5.5 on 1.1.0, 20261009T100849-21b4ea9d) were packaged again with it and replaced with --replace on 2026-10-09: their tables came out the same, they went from 546 to 227 MB and from 770 to 249 MB, the files on the Hub match their manifests, and the tables load from the Hub. The replaced versions were then deleted from the dataset's storage with permanently_delete_lfs_files (history rewritten): the 12 old events.jsonl files and 1080p videos, 1,302.5 MB. The dataset now stores 376 files, 468.8 MB, all used by main; both series still match their manifests, and the tables and a video download from the Hub. The three series in the dataset then (those two and GLM 5.3 Flash on 1.1.0, 20261009T122236-3aae6fc4) were packaged again with cost_source and replaced with --replace on 2026-10-09. Compared value by value with the bundles before: the GLM series changed only by cost_source (billed), their costs unchanged; the Haiku series gained its costs from list prices ($0.31, prices) and its run files the cost events. The files on the Hub match the manifests, and the tables load from the Hub.

Not Verified or Not Built

• Messages: low-popularity warnings and enemy messages are not verified; the skirmish chat area is not read; spoken lines without text cannot be captured.
• Mouse-wheel camera zoom does not work through the bridge (use Z and X).
• Other setups: other Linux distributions and desktops, native Wayland, Windows or macOS game machines, window sizes other than 1920 × 1080, other game builds.
• Agent runs on the NVIDIA GPU: none yet. Only a 45-game-minute do-nothing run has played there (above).
• The text-only menu guide that a fresh checkout uses: no model has played with it yet.
• Benchmark: there is no military scoring. Early single-model pilots used older harness versions and are not comparable.

Known Issues

• Garbage-collector stalls. Under Proton, the game's garbage collector can stop starting collections, paused or running, with or without the reader: about once per 1,300–1,500 collections, or 0.75 times an hour of play, in recorded runs up to 2026-10-09. The managed heap then grows 7.5 MiB a second, and the game's memory 300–470 MiB a minute, until the memory guard closes the game and ends the run (the scorecard then uses the last valid reading). In all 13 recorded stalls the collection count stopped right after an ordinary collection. The cause is Wine bug 59333 (also Proton #9430): found by reading the collector's source and disassembling the game's Mono runtime, and confirmed by the collector's state in a stall on 2026-10-10 (above). The reader's collector watchdog forces a collection after 20 seconds of stall, which ended that stall at no cost to the run. The memory guard stays as the backstop.
• Map-wide structure count. The reader's structure count includes other players, so it cannot attribute a change to the agent on maps with opponents.
• Reads are not atomic. Each sample is read twice and checked, but it is not a guaranteed consistent snapshot.
• Guide images are private. A fresh checkout runs with a text-only menu guide. All recorded runs used the image guide, which you can make from your own screenshots (see Setup).