fix(pursuit): a stealth leg is walked crouched, a death on a pursued step is a measurement, and a build is staged only with every story world passed by stellarfeline · Pull Request #1346 · stellarfeline/delvewright

stellarfeline · 2026-11-11T10:54:46Z

The problem

A delve's critical-path bot walked every declared stealth leg standing up.
configureLeg turned the crouch on once, when the leg started, and
mineflayer-pathfinder clears every control state on setMovements, on every
setGoal and on arrival. The leg calls the first two on every hop, so the
crouch was gone before the first step: on the real pathfinder over the real
client physics, the old code walked a two-hop sneak leg standing on 83 of 83
ticks.

Four more defects were found on the same path:

A trigger step could never be walked crouched. The compiler always gave it
sneak: false, and the harness refused the key on a trigger row.
Moves a crouched body cannot make had no single owner. Neither the bot nor
the sculk model had a shared answer for a climb down, a step down or a jump,
and the sculk model treated a whole sneak leg as silent.
On a pursued step, the bot made those moves blind.
climbHop's descent could end standing at the top of the ladder, on the
edge of the floor beside the column. The pathfinder's next hop finished the
descent only while that hop was walked standing; with the crouch held, the
walk got stuck there.

Separately, the staging gate read only the critical-path run report and the
written-world record, so a build whose branch runs were red was still
admitted.

The fix

The crouch is held on every tick (movement.ts configureLeg). It
yields only for the leg's exported stand_ups.
A trigger performed in front of an objective is walked as that objective
declares (plan::path). It takes the sneak of the first step after it
that is not a trigger.
One answer to "which moves are made standing".
compiler::waypoints::stand_ups returns the climb downs, the steps to a
lower route cell outside any climb, and the declared jumps. It is exported
on a sneak leg as stand_ups and read by the bot. The sculk model hears
exactly the cells those moves stand in (standing_cells →
leg_vibrations).
A death on a pursued step is a measurement (spec-0132 §5.2 and
criterion 10, amended, and declared as a loosening of what died
asserted).
- Each attempt's record states the attacker, and the run line counts
  deaths: …; closest approach d; x death(s).
- The bot respawns as a player does, through the sequencer's one recovery,
  respawnInPlace, which retryOnDeath now shares. It then walks the step
  again from the respawn: the leg cursor rewinds, and the leg resumes from
  the proven waypoint nearest the bot.
- The death-loop bound guards termination, not survival: a step may take
  at most as many deaths as the proven legs it walks have waypoints, and
  the death past that fails the step by name. Whether a life got further
  than the last is never asked.
- What stays red: pursuer_lost, unread, staged/restored, and a
  death loop.
A downward climb takes hold before it slides (climbHop). The old
descent could end standing at the top, on the edge of the floor beside the
column.
A run that walked its branch and exited non-zero is never a passed row
in branch-runs.json. A run-level check such as the sculk ear is not a
report stage, and the gate reads the row.
Each darkness the bot receives is printed, with its position, the step,
the deaths so far and the nearest warden the client sees.
A warden's darkness is judged on the server's positions. The sculk
ear's excuse is read off the pursuit census (SculkEar.onWardenReading):
the reading before and the one after the darkness, within 20 blocks, with
no tolerance constant. Where the client and the census disagree, the
disagreement is printed by name.
The ladder scripts that build the delve image (bot-run.sh,
world-save.sh, reset-when-empty-run.sh) run compose from the engine
root. buildx bake refused the engine's Dockerfile.delve from the
content tree without an fs.read entitlement.
The staging gate requires the branch runs. A build whose
validation/branch-plan.json names a branch owes --branch-runs, and
every planned world must have passed on this build: each row carries the
manifest_sha256 its bot run stamped. playtest-server.sh up takes and
passes the flag.
Gallery.
- obj/board-the-ferry and obj/cross-the-strait declare stealth, and
  trigger/ferry-tiller in front of the crossing inherits it.
- Leg 21 exports a step-down and a jump in stand_ups.
- Each declaration, removed, takes its sneak lines and stand-ups out of
  the emitted path.

Evidence

harness:
- tsc --noEmit is clean; the whole suite passes, 811 of 811. The docs-job checks pass, except check-demo-queue, which needs the content checkout.
- Red checks on the old code:
  - the one-time crouch: the sneak-leg test fails, standing on 83 of 83
    ticks;
  - the old descent: climb-down.test.ts fails, letting go at
    [5.5, 87.0, 4.8] instead of [5.5, 64.0, 4.8];
  - the old died failure: the death-as-measurement test fails;
  - the sequencer's pursued-death branch removed: both sequencer cases fail;
  - the bound removed: the death-loop case fails;
  - the exit mark removed from branch-runs.sh: the new merge case fails;
  - judging the darkness on the client flag again: both new sculk cases
    fail (client 20.5 / server 19.5 passes; server beyond 20 fails);
  - the cd removed from bot-run.sh: the compose-root test fails.
Rust:
- delvewright-dsl lib: 228 passed.
- delvec lib, waypoints + sculk: 25 passed.
- tests/it, native_entrance + teleport_link: 45 passed.
- The trigger-sneak test fails when built without the derivation.
- Clippy over all targets with -D warnings: clean.
tools/tests: the staging-gate and branch-runs tests pass (157). With
the gate's call removed, 8 of the 10 new tests fail.
check-gallery-coverage.py exits 0.

The warden maze on this branch (unreleased build, no override)

Engine edf9246aa, build manifest 613d5d93…. The whole ladder ran from the
content tree, and it is green. Deaths are the measurement:

Rung	Result
PackTest	61 of 61 passed
Critical path	passed, 0 deaths
Hatch stand	passed
`world/0`	passed, 3 deaths (one each on steps 6, 8, 9)
`world/1`	passed, 0 deaths
`world/2`	passed, 0 deaths
Written world	passed
Staging gate (`--branch-runs`)	admitted

On every darkness, the client and the census agreed about the warden in
reach.

A sneak leg sprints for the proven sprint jump it makes standing: the crouch leg's sprint lock is lifted per hop for a stand-up hop of a jump declared sprint (the gallery's ferry-deck plank).

🤖 Generated with Claude Code

https://claude.ai/code/session_019qiQFhKtTpTbYnn9NZBbSd

configureLeg set the crouch once, at the leg's start. mineflayer-pathfinder clears every control state on setMovements, on every setGoal and on arrival (resetPath / fullStop -> bot.clearControlStates), and walkLeg calls the first two on every hop, so the crouch was gone before the body took a step: a declared stealth leg was walked upright. The crouch is now held on each physicsTick for the whole leg, after the library's steering and before the tick's position packet. A crouched body cannot do three moves a player stands up for: a downward climb (it holds on the ladder, and will not walk off the brink), a step down onto a lower floor, and a declared jump (vanilla's edge back-off, which the client physics reproduces). The hold yields for exactly those hops (climbHop for a down climb, hopStandsUp for the rest). Measured on the real pathfinder over the real client physics (test/sneak-leg.test.ts): with the set-once crouch the body was upright on 83 of 83 ticks of a two-hop sneak leg; with the hold, 0 of the leg's ticks. Held through a one-block step down, the body stops at the brink; with the leg's rule it arrives, crouched on every other tick. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019qiQFhKtTpTbYnn9NZBbSd

…ked sneaking A trigger step's sneak was always false, and the harness refused a `sneak` key on a trigger row and walked every trigger upright. A trigger the path performs in front of an objective is walked as part of the walk to it, so it now takes the `sneak` of the first step after it that is not a trigger (a spliced link press included); critical-path.json carries `sneak: true` on such a trigger row, the harness accepts it there and every walk of fireTrigger takes it. The warden maze's world/0 walks its 53-block approach to the bridge as a trigger step in front of the walk to Wren; it was walked upright whatever the objective after it declared. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019qiQFhKtTpTbYnn9NZBbSd

…ith every world passed The gate read the critical-path run report and the written-world record and nothing else, so a build whose branch runs were red was admitted: the warden maze's world/0 died on its step 9 and the gate minted a token for that build. A build whose validation/branch-plan.json names a branch now owes --branch-runs, validation/branch-runs.sh's merged branch-runs.json: every planned world must have a row that ran and passed on THIS build, and a declared branch no world realizes is refused. A build with no plan, or a plan naming no branch, owes none. The record binds to the build: branch-runs.sh stamps each driven row with the manifest_sha256 its bot run stamped, and the gate compares it with the tree's. playtest-server.sh up takes --branch-runs and hands it to the gate; the skill's hand-over and ladder pages name it. tools/tests: 157 passed. With the gate's call to the branch-run check removed, 8 of the 10 new branch tests red. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019qiQFhKtTpTbYnn9NZBbSd

…igger steps, and how the bot holds it Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019qiQFhKtTpTbYnn9NZBbSd

…el hears exactly those moves A sneak leg's crouch now holds, and the warden maze still lost its bot: the one loud move left is the tower ladder's descent, which no body makes crouched, and the bot made it blind. An rcon observer on world/2 read the bot absent from the warden's anger list for about four minutes of crouched walking; its first entry (34) came on the standing slide down the ladder, 14.2 blocks from the warden in three axes, 121 at 5.6, and the bot died at the foot. spec-0132 §5.2, amended: on a pursued step, before each move a crouched body cannot make, the bot waits crouched until every live pursuer whose kind hears stands beyond its reach of the move (least three-axis distance to the move's segment), on the step's own readings; the wait is bounded by the largest forget_ticks (the most anger the kind holds over its decay: the warden's 150 at 1 a second, 3000 ticks), and a spent bound fails the step by name. The route is otherwise unchanged. Criterion 17 states what it asserts. One answer to "which moves are made standing": compiler::waypoints::stand_ups (a climb down, a step to a lower route cell outside any climb, a declared jump), exported on a sneak leg as `stand_ups` and read by the bot (the harness's own y-rule is removed); the sculk model hears exactly the cells those moves stand in (standing_cells -> leg_vibrations). The hearing is delvewright_dsl::senses, one record, carried on the pursuer's row as `hearing`; warden-behaviour.md §8 records the reach and the bound. Tests: dsl senses 1 passed; delvec lib waypoints+sculk 25 passed; it native_entrance+teleport_link 45 passed; harness tsc clean, waypoints, critical-path, movement, sneak-leg and the four wait tests pass. With quietBeforeStanding returning at once, 2 of the 4 wait tests red. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019qiQFhKtTpTbYnn9NZBbSd

… press in front of it with it obj/cross-the-strait declares `stealth`, so the exported path carries `sneak` on it and on trigger/ferry-tiller performed in front of it: the gallery's element for the sneak marker on an objective step and on a trigger step (spec-0039). The executor's crouchYielded doc names the exported stand_ups. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019qiQFhKtTpTbYnn9NZBbSd

climbHop's descent aligned the body within 0.25 of the column's centre and then slid "while not on ground"; a body entered from above whose 0.6-wide hitbox still overlapped the floor beside the column stood on that edge, so the loop ended at once and the climb "let go" at the top. The pathfinder's next hop used to finish the descent; with a sneak leg's crouch now held on that hop, a crouched body holds on the ladder and the walk stuck there for the hop's whole timeout (the warden maze's world/2, step 7: "let go at [47.5, 87.0, 48.5]", then stuck at y 86.8). The descent now steps to the column's centre until the body drops into it, and fails by name if it never does. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019qiQFhKtTpTbYnn9NZBbSd

…'s stand-ups are emitted obj/board-the-ferry declares `stealth`; its leg steps down a block and takes the declared jump, so the waypoints export both hops in `stand_ups` (step-down [18,69..70,20], jump [16,69,20]->[12,69,20]) and the path carries `sneak` on it. Measured on the gallery primary (release build of this branch, prefabs by tools/ci/gallery-prefabs.py): with both declarations, `sneak` on steps 27, 28 (trigger/ferry-tiller) and 29, and leg 21's two stand-ups; with board-the-ferry's `stealth` removed, no stand-ups and `sneak` on 28 and 29; with cross-the-strait's removed (previous commit), the two `sneak` lines on 28 and 29 go. The build's vibration lines are identical either way. check-gallery-coverage.py exits 0. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019qiQFhKtTpTbYnn9NZBbSd

The wait (spec-0132 §5.2's amendment, criterion 17, quietBeforeStanding and its forget_ticks bound, and the delvewright_dsl::senses registry and the pursuer row's `hearing` that only it read) is removed. Measured on the warden maze, it cannot work: a warden comes to a body that stands still — three of three waits at the top of the tower spent the bound with the warden 2.5 to 3.3 blocks from the ladder's foot, and two of eight waits at the dais ended with the bot killed while waiting (warden-behaviour.md §8 keeps the record). The shared stand_ups answer, the sculk model's agreement with it, the held crouch, the trigger sneak and the climb's take-hold fix stay. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019qiQFhKtTpTbYnn9NZBbSd

…e pursuer row's `hearing`, its parse, and the record The previous commit removed only crates/dsl/src/senses.rs (a pathspec error dropped the rest of the staging); this one carries the remainder, so the tree builds again here. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019qiQFhKtTpTbYnn9NZBbSd

…pawns and walks the step again, and only a death loop is red Owner ruling, a loosening of what `died` asserted (spec-0132 §5.2 and criterion 10, amended in their own words): whether a body survives a vanilla-AI pursuer is the designer's call, and a solo player passes by dying and retrying. judgePursuedStep no longer fails on `died`; each attempt's record states the attacker and the run's line counts deaths (`...; closest approach d; x death(s)`, `deaths` in the report). The sequencer's one recovery after a death (respawnInPlace: wait out the respawn, re-arm where the bot stands — what retryOnDeath already did, now factored and shared) is taken on every death on a pursued step; the executor then rewinds the leg cursor to where the step began and resumes its leg from the proven waypoint nearest the respawn (resumePursuedStep). A life must get further along the step's proven route than every earlier life on it, measured as the furthest proven waypoint reached, leg by leg; one that does not is a death loop and fails the step by name. So a step is walked at most once more than it has proven waypoints: the bound is the route's own length. The climb's take-hold fix gains its test (test/climb-down.test.ts: the real client physics, the tower in miniature; the old descent lets go at [5.5, 87.0, 4.8], the fixed one at [5.5, 64.0, 4.8]). Red->green: the old `died` failure put back reds the death-as-measurement case; the sequencer's pursued-death branch removed reds both sequencer cases; the progress comparison removed reds the death-loop case. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019qiQFhKtTpTbYnn9NZBbSd

…oute has waypoints; progress is never asked The bound guards termination, not survival. Asking each life to get further than the last re-imported the judgement of the pursuer's vanilla will that the owner ruled out. A pursued step now takes at most as many deaths as the proven legs it walks have waypoints, whatever each life reached, and the death past that fails the step by name (death, bound, legs). The executor notes the legs a step's walks match (noteStepLeg) in place of a per-hop progress value. spec-0132 §5.2 and criterion 10, playtest-methodology.md and the skill's ladder page say so. Red->green: with the bound removed, the death-loop case reds (the fifth death on a four-waypoint leg is walked again). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019qiQFhKtTpTbYnn9NZBbSd

… never a passed row On the warden maze's world/0 the bot walked its branch to the end and its report recorded every stage passed, but the run exited non-zero because the sculk ear failed it (`darkness 1` that no warden in reach explained). The merge copied the report's `passed: true`, so the staging gate, which reads that row, admitted the build. The exit code is now the verdict: a branch whose run wrote a report and still exited non-zero is merged as not passed, with the exit code in its reason. The harness also prints each darkness it receives, with where the bot stood, the step, the deaths so far and the nearest warden the client saw, so a darkness the ear counts can be read against the deaths and respawns around it. Red->green: tools/tests/test_branch_runs_keeps_bot_report.py's new case (a stub run exits 3 after a passing report) fails without the exit mark, and passes with it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019qiQFhKtTpTbYnn9NZBbSd

…mpose from the engine root Compose builds through buildx bake, and bake refuses to read a file outside the working directory unless it is granted an `fs.read` entitlement. The skill runs the ladder from the content tree, where the engine's validation/Dockerfile.delve lies outside the working directory, so the image build was refused before the hatch stand ran: additional privileges requested: pass "--allow=fs.read=../wt-maze-death/validation/Dockerfile.delve" The same `bot-run.sh --only-hatch-stand` on the same build passed when run from the engine root. bot-run.sh, world-save.sh and reset-when-empty-run.sh now cd into the engine root before compose (branch-runs.sh already did). Every path they hand compose is absolute or resolved against validation/ first; world-save's --seed is made absolute before the cd. tools/tests/test_ladder_compose_runs_from_engine_root.py: every script that exports DELVE_DOCKERFILE changes into the engine root before its COMPOSE line. It found reset-when-empty-run.sh as a third instance, and it fails with bot-run.sh's cd removed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019qiQFhKtTpTbYnn9NZBbSd

…the pursuit census, not the client's entity table The sculk ear excused a darkness when the client saw a warden within 20 blocks. The pulse is applied on the server's positions, and the client's entity table lags them at the edge of the reach. On the warden maze's world/0, one darkness of 34 read past 20 on the client and failed the run. A diagnostic run printed every darkness: the nearest warden the client saw ran up to 19.0 blocks. The excuse is now read off the pursuit census, one source: each reading of a declared warden's actor feeds the ear its server distance to the bot (SculkEar.onWardenReading, from readPursuers). A darkness is a declared warden's when the reading before it or the one after has the warden within 20 blocks and no shriek was heard within a second. No tolerance constant. Where the client and the census disagree about a warden in reach, the run prints the disagreement by name (`[sculk] darkness at <t>: the client …, the census …`); the client never decides. spec-0132 §5.7, compiler.md and playtest-methodology.md say so. Red->green (test/sculk.test.ts): a darkness the client places at 20.5 and the census at 19.5 passes; one the census places at 20.5 and 21 fails, whatever the client saw. Judging on the client flag again reds both. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019qiQFhKtTpTbYnn9NZBbSd

… standing A walk declared stealth that crosses a declared sprint jump could not be walked: `configureLeg` turns `allowSprinting` off for the whole sneak leg, and mineflayer-pathfinder 2.4.5's parkour move lands at most two cells forward with sprinting off (`getMoveParkourForward`, `maxD`), so the jump's landing was no neighbour of its take-off cell and the hop answered `No path to the goal!`. The compiler already names the jump among the leg's `stand_ups` and the crouch hold stands the body up for it; the sprint lock was never lifted. The gallery's ferry walk became stealth on this branch and crosses the broken plank (`jump/the-broken-plank`, `sprint: true`, a three-wide gap): the gallery bot stopped at the plank's take-off on `anchor/ferry-deck` waypoint 7/11. `hopMovements` (harness/src/movement.ts) is the one per-hop setter the leg calls: the parkour move as before, and sprint on for a plain leg, or for a sneak leg's hop that is a stand-up of a jump declared `sprint`. Direction: the defect can only turn a proof red (a stealth leg over a sprint jump halts); it cannot let anything ship. Test: test/sneak-leg.test.ts "a sneak leg makes its proven sprint jump standing, at a sprint, and lands" walks the real pathfinder over the client physics across a three-wide gap. With the sprint held off on the sneak leg it reds ("stopped at [9.50, 64.00, 2.50]", the take-off cell); with the fix, 4 of 4 tests in the file pass. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019qiQFhKtTpTbYnn9NZBbSd

stellarfeline and others added 16 commits November 10, 2026 23:25

docs(compiler): the critical path's sneak marker, on objective and tr…

5015d62

…igger steps, and how the bot holds it Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019qiQFhKtTpTbYnn9NZBbSd

stellarfeline merged commit 85fc614 into main Nov 11, 2026
39 checks passed

stellarfeline deleted the fix/maze-pursued-leg branch November 11, 2026 18:43

Navigation Menu

Repository navigation

fix(pursuit): a stealth leg is walked crouched, a death on a pursued step is a measurement, and a build is staged only with every story world passed - #1346

fix(pursuit): a stealth leg is walked crouched, a death on a pursued step is a measurement, and a build is staged only with every story world passed#1346
stellarfeline merged 16 commits into
mainfrom
fix/maze-pursued-leg

stellarfeline commented Nov 11, 2026 •

edited

Loading

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

stellarfeline commented Nov 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

The problem

The fix

Evidence

The warden maze on this branch (unreleased build, no override)

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

stellarfeline commented Nov 11, 2026 •

edited

Loading