The route

Prerequisite: Tests, and How we know it's right for what the playthrough is for.

This page is the route itself: what the twenty-seven segments cross, how much of it to run while you work, what driving each of the game's minigames costs, and the conventions a new segment will trip on. The suites that run it are the test reference.

The game is won. The route plays from the cold boot to the credits in one continuous session and the closing narration comes out mission = "good" — all four artifacts the ending is scored on in Frank's hands, onehappens, twohappens and revhappens all false. Both hosts agree: npm run test:playthrough 30/30 headless, and the browser gate 91 beats over 27 segments with no divergence at all — every one of them carried, zero checkpoint loads, 23.7 min, ending on credits.mov.

The twenty-seven segments

Each row is one test in playthrough.ts, one function in segments.ts, and one .ti in out/checkpoints/ named for the state it starts from.

segment	from	crosses	ends
1	the boot	mission 0 → mission 1 phase 0	the London flat, the bomb, the crossing
2	m1p0	m1.0 → m1.1	boarding, the gym, Penny's briefing
3	m1p1	m1.1 → m1.2	the wireless room, the trunk, the Enigma
4	m1p2	m1.2 → m1.3	the turbine plant, and the engineer's thanks
5	m1p3	m1.3 → m1.4	the Rubaiyat hidden in a coal bunker, then taken back
6	m1p4	m1.4 → m2.0	the necklace sub-plot, then Penny's debrief; mission 1 is complete
7	m2p0	m2.0	the Purser's errand, and into the wireless room
8	m2gram	m2.0	the set switched to transmit, Thayer's telegram sent
9	m2sent	m2.0	the report, the passenger manifest read, purs at "none2"
10	m2man	m2.0	the cufflink errand, and Straus's cufflink out of a reception chair
11	m2link	m2.0	the cufflink handed over, the car keys off the office wall
12	m2keys	m2.0 → m2.1	the hold, the car's headlight, the crate, the painting
13	m2p1	m2.1	Max, Smethells, the squash court, and the first fencing bout won 5-0
14	m2fence	m2.1 → m2.2	Smethells hands over Willy's ring for one win
15	m2p2	m2.2 → m2.3	Clariss Limehouse identifies the ring, and keeps it
16	m2p3	m2.3 → m3.0	Penny's debrief; mission 2 is complete
17	m3p0	m3.0	Willy's body in the Turkish bath, and the Rubaiyat clue
18	m3clue	m3.0	the rope on Scotland Road, and the Hacker's phrase for the clue
19	m3phrase	m3.0	the Old Reds off Willy and Zeitel's table in the cafe
20	m3cigs	m3.0 → m3.1	the cigarettes to Max in the smoking room
21	m3p1	m3.1	the Chief Engineer's second favour: the turbine plant again, and the engine-room door
22	m3thanks	m3.1 → m3.2	the fight with Vlad, won 512-0
23	m3p2	m3.2	nine floors of the false smokestack, solved, to the notebook's platform
24	m3top	m3.2 → m3.3 → mission 4	the notebook taken, and given to Zeitel; mission 3 is complete
25	m4p0	m4.0	Penny recaps the case — then Zeitel's cabin, the bomb, and the notebook back
26	m4penny	m4.0	Clariss's shawl; Zeitel's deal for the painting refused
27	m4anti	m4.0 → the end	the boat deck, a place in a lifeboat, and the closing narration

Some of those checkpoint names are historical and no longer describe what happens at them — segment 27 still resumes from m4anti although the antidote errand it was named for is retired. The name is the file on disk; the row says what the segment does.

Two segments are missing from that history and both are worth knowing about, because each was deleted rather than fixed:

• There used to be a twenty-eighth that went six decks down to trade Clariss's shawl for the real necklace in the turbine room. It could only ever load — the trip crosses the boat deck and answers away the Gorse-Joneses' one-shot lifeboat offer, leaving the ending segment with nobody at the rail. The necklace sub-plot gets that necklace in mission 1 instead, so the trade had nothing left to buy.
• Segment 26 used to hand the painting to Zeitel for Lady Georgia's antidote and buy it back with a boat pass won at blackjack. savegeorgia()'s 102 — "No. You're bluffing." is the one answer that never enters givepaint(), so the painting is simply never let go of. That retired two segments' worth of work and with them the only part of the run whose outcome was a property of the RNG stream rather than of the play. Lady Georgia dies of the poison for it; the closing narration does not score her, and segments.ts says so where it makes the choice.

How much to run

The headless suite is the commit gate and is cheap enough to run every time. The browser suite only ever says something new about the segment just written — every divergence found so far came from the new segment, never from an old one — and a full run spends most of its time replaying segments untouched for days.

while	run	cost
writing a segment	npm test && npm run test:playthrough	~2 min
that segment's cross-host check	SEGMENTS=13 npm run test:browser:seg	20 s – 5 min
a mission is finished	npm run test:browser:m0 / :m1 / :m2	1 – 8 min
before a long break, or after touching nav/aim/drivers	npm run test:browser	~24 min

The last row is the one that matters: a change to the shared navigation, aiming or driver code is exactly what CAN break an old segment, and that is when the whole thing earns its twenty-four minutes. It was 36.5, and what the three parts of that difference were is TODO §4a — the accost sweep's dud clicks, one wait in the ending, and the deck plan the map lands you on; none of them the game's own pace.

npm test                        # the gate
npm run test:playthrough        # 27 segments + 3 property tests — writes out/checkpoints/*.ti
TAOOT_RECORD=1 npx vitest run --config vitest.playthrough.config.ts   # re-record goldens
TAOOT_RECHECKPOINT=1 …          # …and rebuild the .ti checkpoints (after a save change)
npm run watch:m2p0              # watch segment 7 in a real window
SEGMENTS=9 npx tsx tests/browser/playthrough.ts    # one segment alone, ~35 s

The minigames, and what driving one costs

Six of the game's mechanics are on the route. Each needed something the one before it did not, and the shapes recur.

The turbine plant (segments 4 and 21) is a simulation with a fixed point, and the route waits for it rather than for a number of ticks: PLANT_STEADY holds all four flows stationary, because iterateone computes the flows from the levels before moving them and whether one more iteration fits before a beat depends on frames drawn. It has to be re-dialled in segment 21, and that is the save's fault, not the plant's — the twelve plant globals have no record in the shipped template and only three free node slots exist to make records in, so a checkpoint cannot carry them. The dial loop is idempotent, so a carried session sets the same numbers and waits on a fixed point it is already at.

Fencing (segment 13). One click does two jobs: FENCE.STG mousedown reads the ATTACK quadrant out of the click point, while playeridle re-reads mouse() every tick and sets the BLOCK from the X alone — so the route clicks in the column that guards the side Haderlitz has chosen, and leaves the cursor there. willieidle fills its four defence slots from where the cursor is, weakest at the cursor's own quadrant, so attacking where you hover is attacking what he guards least. The route wins 5-0, and that is not showing off: pointgoesto moves fencelevel by four per point, so a point conceded would leave the two hosts running different difficulties and diverging on every random draw after it.

The fight with Vlad (segment 22) is arithmetic once the combo is right — right cross, left cross, uppercut, repeat. vladdamage reads the combo rather than random(), so it is 121 blows and vladpower = -52 in both hosts, with playerpower still its opening 512. Only Vlad's blows roll dice, and cycling three different blows on alternating sides trips none of the four repetitions that would hand him a turn. The trap is the ending: the end-of-fight test lives in Vlad's own idle handler, his idle only runs when the loop vladdamage arms comes round, and every click cancels that loop so a blow can interrupt him. A driver clicking as fast as it can pump starves its own win condition — 400 blows, vladpower −520, the flat still open. The route stops at vladpower < -50 and waits, which is what a player does.

The smokestack maze (segment 23) is smstack2, one floor of an eight-scene ring replayed eleven times. Climbing is four views and is never blocked; walking round the ring is sixteen views the maze shuts by setting blocks to a comma-list of closed gaps per (maze, level). mazenumber = random(4) is drawn live at the door, so the two hosts need not be dealt the same maze and the golden's value is masked. That costs nothing, because nav/smokestack.tssolves whichever maze it draws — a breadth-first over (level, position) against the eight blocks strings — and asserts every move against that solution, which is a stricter test than a fixed value. It has to solve: one of the sixteen (maze, entry) pairs is a dead end, maze 4 into scene39, both gaps closed. Maze 4 is the hard one and it climbs in 18 moves.

The necklace sub-plot (segment 6) is what clears onehappens, and it is the longest chain on the route: neckphase 0 → 1 → 2 → 5 → 6 → 7 → 8 → 9, through Georgia on three decks, Charles in the smoking room, the A-deck fusebox twice (cut the power to draw Sasha out of his cabin, restore it so the bag can be seen into), and patty.stg — a four-dial combination lock whose combination is written down in the game's own debug cheat (solvedoll(): 6, 0, 0, 0). Getting it wrong is fatal. a14.Set c54 checks that the fake necklace is Vlad's when Sasha comes back; taking the real one without leaving the fake in its place is playerdeath = "by sasha". So the swap has to be an EXCHANGE, and inven1.stg doneck() enforces it from the other side — the bag can never hold both.

The bomb (in segment 25) is three numbers, and bomb.stg's own debug cheat solvebomb() names them: unibomdoor = 0, propdeg("key") = 5, unibompower = -1. The last one comes from one place — hammershake(), which only runs with unibomflag = 0, which is only true while switch1 = 0, which needs the key unturned. So the order is forced: switch1 down, power on, solenoid up, wait 58 iterations (~42 s) for the timer to run out harmlessly, switch1 back up, then the key. changedone() runs after every control change and five of its nine branches are boomer(). It was defused first time and costs three minutes of the endgame clock.

Blackjack is measured but off the route: the deck is 52 draws off a stream every earlier gesture has been moving, so winning is a property of the world rather than of the play, and the route stopped needing the boat pass. The same hand has been measured winning and busting with nothing about the play changed, so a future route must not assume the first one wins.

Conventions a new route will trip on

Learned the hard way, all verified:

• In mission 4, a conversation costs two minutes. gang.cst prepuppet() does min = min + 2 at mission = 4, and the sinking's phase is a clock (the endgame is a countdown). "Ask everyone everything" is not free the way it is in missions 1–3.
• Walking IS pressing up, and some standpoints are wired to that press. The planner's walk within a room is a series of pressUps, so a route crossing the 2nd-class staircase changes deck by accident (nav/stair2c.ts). faceStandpoint avoids pressing up at any standpoint the ship graph knows is an exit — which is why a transition the graph does NOT know about is worse than a missing edge.
• A parked movie is a question. nav.rush refuses one deliberately. Use clickMovie(r => …) and name the region by what it does (type === 2 pages, target === "openit" opens the Purser's window).
• A poll loop must not be awaited. INVEN1.STG dobook() parks in while not button() waiting for the player's next press — startClick fires the gesture that starts it, and the press that follows is holdUntil, because viewer.click moves the cursor without pressing.
• A minigame can end on a timer your own clicks postpone. Look for this wherever the thing that ENDS a minigame is a makeloop rather than a script the gesture itself runs; the fight is the worked example above.
• Some buttons need pressing until they take. A flat refuses input while a script is still running, and clickThing answers "it landed on the button" either way. closeBunker in segment 5 presses up to four times: headless the loop it waits on is already over, in a browser it is not.
• A trackbut must be clicked inside the button's REGION, not where the dark prop's pixels are. The shipped trackbut (BOOTFILE 0002) answers pointinbutton(currentflat(), target, mouse()) — the flat's named click region, where target is the region the click arrived on — and the bevel it is NAMED with is only the highlight it shows while you hold. In practice the two coincide, which is why the aim points here did not move when the engine stopped transcribing this helper: measured, cuff.stg and wireless.stg's ok regions are 429..485, 338..362, and fuse.stg's fuseoklit prop is 428,338 57×25. What still fails is aiming at the DARK prop: fuseokdark is 419,330 76×42 and overhangs on every side, so a click on its edge samples outside the region and trackbut answers 0, silently, with the flat left open. cuffok is the same geometry.
• Props can be behind other props. The Rubaiyat and the boiler switch have no clickable footprint until the bunker door is opened: boildoor's frame is 82×120 with all 9840 pixels opaque, at propdist -3, in front of both.
• An actor outranks a hotspot in the hit order, and dismissing one takes frames, not an answer. The seaman in front of the A-deck fusebox ends his own lines with putdownactor(); until that lands the second click cannot even be aimed. Waiting ~40 ticks after his last line is the whole fix.
• An instrument's state can live in propvalue, not propdeg. The wireless needle's frequency is its value (14..200, tuned() wants 34..40 to transmit); its deg is only which of ten pictures the small view draws. NavDriver.propValue exists for that, and both drivers mirror it.
• Set a puzzle up BEFORE opening the flat that uses it. WIRELESS.STG's openflat() calls setuptx() only if the sender, the breaker and the tuner are already right; switch them afterwards and the morse key silently does nothing. The same shape will turn up wherever a flat's openflat reads prop owners.
• A ratchet dial's direction is arithmetic, not a guess. fixdeg256 buckets the cursor's bearing and limiter takes the sign of the bucket change, so a rising atan2 bearing raises the value; the swing per move has to clear a bucket (256/6 ≈ 43 for the tuner, hence ARC = 48) without tripping the dial's own seam test. Tuning the wireless from 200 to the transmit band is ~80 swings and 12 s in a browser — there is no acceleration in adjustneedle.
• Owner is not presence. actorOwner("ga") reads "none" while Georgia is standing right there. The runtime record (visible, worldX/Y/Z, starName) is what to read, and reading the owner that way once cost a session an hour.
• Name the standpoint in a big room; do not sweep. nav.accost caps at MAX_GESTURES_PER_ROOM = 60 and the boat deck has 27 scenes, so a blind sweep never reaches the far end of it.
• A click that lands is not a click that worked. inven.shp's stdmouse gates every object lying in a room on realdist(what) < hotdist(), so a click from across the room reaches the thing and is then thrown away. nav.hunt judges a standpoint on what the click moved — the screen-level four, handitem, the thing's own owner/view/visibility/deg/value, the globals minus the self-moving ones — and otherwise keeps sweeping, ending with clicked X from N standpoint(s) and nothing moved instead of a false ok that surfaces three assertions later. Segment 23's notebook is the case with numbers on it: View53 is inside a 4100 hotdist by 58 units and loses anyway. (accost is the same shape for people, and stays its own thing — there the effect worth waiting for is somebody answering, not the first thing that moved.)
• A click nobody answers is worth 1.5 s, not 8. The sweep still clicks from every standpoint — that is how a walking character gets time to arrive — but it stops waiting once the person it clicked has held still for ACCOST_WATCH_MS. Measured: every dud wait in the route (Zeitel 15×, Georgia 4×, Charles 3×) is somebody whose position never changes by a unit, and the wait was 8.1 s of browser time each. Movement, never distance: every answered click came from beyond hotdist() as the route computes it (nav/reach.ts).
• Find a hotspot's script by container id. ObjectEntry.locationScript in src/df/set.ts maps a view's object to the container that handles it — that is how Sasha's door turned out to be the "door" of Scene51/View57 and not the "knock" beside it.
• Bevel ids, never positions. PENNY1.PUP's zeitelgram() calls puppetscramble(), which shuffles its plaques.
• out/scripts/ is the first place to look for any of this (npx tsx tools/dumpscripts.ts gamefiles out/) — but it is not complete for BOOTFILE: five codes are dumped where boot1 and boot2 carry 78 between them, so progress looks undefined when it is not. session.bootScripts[n].script.codes is the honest list. npx tsx tools/disasmcmd.mts <name|0xVA> disassembles TI.EXE when the scripts are not enough.

Back to the reference index.