dayz-mcp-verify
Auto-test a DayZ mod in-game through dayz-mcp without keyboard or OCR: spawn a classname, orbit/capture, raycast collision, inspect placement/attachments/telemetry and emit PNG+JSON evidence. Compose with dayz-test-ingame; dayz_test_run/dayz_test_stop own the managed lifecycle, t
Install
npx skills add https://github.com/willy92wins/DayZ-Modding-Knowledge-Pack/tree/main/skills/dayz-mcp-verify
claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install willy92wins-dayz-modding-knowledge-pack@llmmart
git clone https://github.com/willy92wins/DayZ-Modding-Knowledge-Pack.git
The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole willy92wins/dayz-modding-knowledge-pack collection as a plugin from our marketplace. Git is the plain clone.
Skill manifest
DayZ MCP verify — auto-test in-game vía tools MCP
GATE 0 — ask WHO is driving, before touching anything (added 2026-08-07)
(until 1.29: managed dayz_test_run / dayz_test_stop and Mode=all capture still hold.) (since 1.30: Diag Script Console tabs register via ScriptConsoleTabRegistry; IsHeadless() names roboclients; SP PerformActionStart no-ops while pending. Details: dayz-1-30-mcp-verify.md.)
Before the first dayz-mcp tool in a verification run, ask the user whether
they are driving or the MCP is. Neither is assumed. A single question per run, not per
capture.
Why it is a gate and not a preference: with the user present, their manual cycle is faster
and yields better results — spawns with VPP in seconds and evaluates the whole scene at a glance,
while MCP loop needs to position camera, capture, and chain tools. MCP wins when
the user is NOT there: early morning, unattended run, or while attending another front. Furthermore,
changing mode is not free midway (clean manual test requires relaunching without
@DayZ_MCP), so the decision is made at the start of the cycle or paid twice.
| Situation | Who drives |
|---|---|
| User present and available | Them (VPP + direct judgment) — MCP only if they ask |
| User absent / unattended run | MCP, and report is left for them with evidence |
| Doubt | Ask. Never assume |
What does NOT change: MCP remains mandatory for what a human cannot measure by eye (telemetry, numerical raycast, exact placement, repeatable series). There you do not ask, you use it.
Measurement that originated gate (2026-08-07, corpus of 276 sessions): 63% of wall-clock time in a full cycle falls on user side, and 71% of their actual verifications are automatable with today's tools. That it is automatable does not imply it should be automated: with them present, their cycle rules.
WHAT THIS DOES
Drives an already launched DayZDiag client, via dayz-mcp MCP server tools, to
verify a mod without human intervention: spawns object, views it from multiple angles
(camera + window-grab), checks collision via raycast, reads placement/telemetry, and
produces a report with pass/fail criteria and evidence. It is the automated observation
loop bridging the gap between "PBO compiles/deploys" and "looks and behaves properly
in-game", for the class of properties that DO NOT require keyboard input.
Server-authoritative + passive pixel capture. Control and data are engine-native
(spawn, raycast, telemetry, camera); sole non-native piece is visual capture
(window-grab of rendered client — MakeScreenshot is broken in diag, T165276).
COMPOSITION — public lifecycle + MCP verification
Build/deploy/launch belongs to dayz-test-ingame (its environment preflight runs on its own:
P:\ mounted, junction P:\Mods, AddonBuilder, allowFilePatching). This skill invokes it
with MCP mod added and then drives. Bridge verbs do not start or terminate
processes; public test tools do orchestrate managed lifecycle.
Before operational flow, read canonical protocol:
<runbooks>\dayz-mcp-agent-session-protocol.md.
[EXACT][CLAIM-R21-MCP-ORCHESTRATED-TEST] Execute lifecycle with
dayz_test_run and stop it with dayz_test_stop on exact run_id. Both
tools possess FIFO queue, lease, heartbeat, and release; do not wrap them in
a second session_acquire. Reserve session_* primitives for
low-level mutations not already encapsulated. Every lifecycle is
identified by run_id: sharing mod does not grant ownership. Under retail
quarantine only reads are allowed; if whoever opened retail cannot close it via
UI, declare manual_cleanup_required.
Bridge startup sequence (rev. 2026-09-06): dayz_test_run → session_acquire_wait →
verbs. Never probe before adopting. "Do not wrap launch in another lease" does not mean run
remains yours: upon completion, dayz_test_run RELEASES its owner lease (daemon audit:
session_release_finished reason=owner_release right after launch) and run transitions to RUNNING_IDLE
without owner (release_owner, process_lifecycle.py:958-976). From that instant the idle run fence
rejects ALL /enqueue on that run, reads included: _enqueue_run_rejection discards flag
mutation (loopback.py:1293-1296) and for RUNNING_IDLE returns run_not_owned (:1317-1318) with
hint "This run has no owner (RUNNING_IDLE). Adopt the existing run before dispatching." (:99-102,
:1323-1329). The counter-intuitive part: bridge_status counter is named
fence.mutation_rejects_by_code (:2821-2822) and counts reads as well. Every verb crossing
bridge falls there (query_*, telemetry, raycast, captures, world_spawn…) and also
wait_for(players_at_least|players_at_most), which polls with query_all_players (server.py:2645-2650).
wait_for(log_matches) does NOT fall: reads log files without touching bridge (server.py:2603-2637,
:2686-2697).
dayz_test_run(...)→ retainrun_id.- Optional without adopting:
wait_for(log_matches, "…")to wait for startup via RPT. session_acquire_wait(purpose=…): grant adopts the ONLY ownerlessRUNNING_IDLErun and declares it inadopted_run(_adopt_on_grant,loopback.py:3119-3120,:3193-3230). Checkadopted_run.okbefore first verb. With multiple idle runs it answersadopted_run.error = "multiple_idle_runs"and adopts none (:3202-3208):dayz_test_stopsurplus run and repeat.wait_for(players_at_least, 1)and remaining verbs.- [EXACT] The adopted lease expires by itself (added 2026-09-27, LFPowerGrid, three runs): it lasts 120 s and is NOT renewed internally, fresh MCP client or not. Send
session_heartbeatat least every ~90 s during the whole run, including while a human plays. Past the deadline the run goes ownerless again (grace.remaining_scounts down) and the daemon stops it (lifecycle_stop_outcome: stopped), even with the user in front of the game. (measured in game, DayZ 1.30.164014 Exp)
- [EXACT] The adopted lease expires by itself (added 2026-09-27, LFPowerGrid, three runs): it lasts 120 s and is NOT renewed internally, fresh MCP client or not. Send
session_release→dayz_test_stop(run_id).
With an MCP client prior to 8f5727f of DayZ_MCP (2026-09-06) rejection arrived as a
BARE remote_error, without code or hint (ledger entry fb-20260906-193626-f45d, first attributed to a
client without polling); with current client it arrives as run_not_owned: This run has no owner (RUNNING_IDLE)…. In both cases remedy is adopting, neither retrying nor relaunching. This rule corrects
step 1 of recipe SP-292 and clarifies first bullet of SP-152, below.
Two lease failures around long waits (measured in game sessions, 2026-09-28): (1) wait_for over the log does NOT renew the lease — a wait longer than ~2 min expired with lease_expired; interleave session_heartbeat between waits. (2) When a user message interrupts an in-flight call (a long wait_for), the next session_heartbeat answers lease_invalid: token was never valid for this client; if the run is still alive, session_acquire_wait readopts it. Also: dayz_test_close shuts the run down in order, with a termination line on both sides. [EXACT]
Companion externo: dayz-labs
[EXACT][CLAIM-R21-MCP-COMPANION-AUTHORITY]
external_companion_no_lifecycle_authority. Release dayz-labs v0.1.35,
reviewed at pinned commit
dbd6ad3e54e30c81a9aeb88fcb9f60f007804c2a, can serve as reference or
optional companion. Its start/stop/restart verbs are excluded while
DayZ_MCP holds a run: only dayz_test_run / dayz_test_stop on exact
run_id govern that lifecycle. Do not install or update companion
as part of a gate; always document the examined pinned_version.
Its interface is WPF and captures its own application:
wpf_not_layout_evidence. A correct capture from dayz-labs does not validate the
parser, proportions, clipping, or semantics of a DayZ .layout; UI
viewer needs its own render and comparison against in-game evidence.
The dayz_test_run request must select Mode=all, build when PBO
changed, and @DayZ_MCP as additional dependency:
-Mode all(server + client) is MANDATORY: visual capture reads from rendered client. A headless-Mode serverhas no window to grab.@DayZ_MCPis loaded alongside the mod under test; both peers (server + client bridge) poll the MCP server loopback.
PREREQUISITES (gate before driving)
dayz-mcpregistered and connectable. PRIMARY gate: are toolsmcp__dayz-mcp__*available in session? (ToolSearch finds them → gate PASS).claude mcp listis SECONDARY and NOT reliable: gives false× Failed to connectwith server operational — do not block on it if ToolSearch sees tools. Tools are only available in a session started WITH server already registered (broker registration--client, see TROUBLESHOOTING).Bridge config seeded. In-game bridge reads
dayz_mcp.json(url + key + pollHz) from its profiles; MCP server uses<DayZ_MCP_dev>\tools\.dayz_mcp.key. Seed the SAME key in launch profiles withinstall-mcp.ps1pointing to directories generated bydayz-test.ps1:<DayZ_MCP_dev>\tools\install-mcp.ps1 ` -ServerProfiles "<TargetMod>_dev\_server\profiles" ` -ClientProfiles "<TargetMod>_dev\_client\profiles" ` -MissionPath "<...>\mpmissions\dayzOffline.chernarusplus"[verify on first run]— generic seeding (arbitrary mod +@DayZ_MCPviadayz-test.ps1) was until now only exercised throughrun-fase3.ps1. The FIRST run of this skill validates that bridge finds config in these paths; ifbridge_statusreportslast_poll_age_s = null, config is not where bridge looks for it — adjust path and declare it.bridge_statusgreen. ALWAYS first call:bridge_status.server_peermust have freshlast_poll_age_s(not null). For capture,client_peeras well.never polled→ bridge did not start (config/key/startup) → abort and diagnose, do not keep driving blindly.Mods passed via
extra_mods/base_modsmust be REAL DIRECTORY underP:\Mods(→!Workshop); Steam junctions ARE NOT valid. Two layers, confusing them costs a session:- Public boundary
_valid_public_mod(DayZ_MCP_dev\tools\dayz_mcp\dayz_test_tool.py:75-85) rejects:,\and/across all projects → absolute workshop path never passes through tool. That is by design, NOT a gap in your project's policy. (Updated 2026-10-03: the boundary now also accepts an absolute path inside the selected project'smod_roots;_valid_public_modcallsdayz_test_request._valid_mod_entry(dayz_test_tool.py:269-272,dayz_test_request.py:222-233, DayZ_MCP_dev506de5d). A workshop path lies outside them and still fails, withbad_mod.) - Subscribed mods exposed by Steam in
!Workshop(@CF,@Dabs Framework,@VPPAdminTools) are Junctions tosteamapps\workshop\content\221100\<id>, and path identity guard rejects them →dayz_test_failed, which is the generic catch-all ofserver.py:1101and swallows the real cause.
Operational consequence: CF/Dabs/VPP are only loaded by declaring them with absolute path in project
default_base_modsinside launcherrequest-policy.json— which is how SUB_BRZ and LFHeli declare them. Mods that are real directories (@DayZ_MCP,@LFHeliCore, mod under test) do go viaextra_modswith relative name.Diagnosis in one command:
Get-Item '<!Workshop>\<@Mod>' -Force | Select-Object LinkType→Junction= relative does not work. Measured 2026-08-02 withpreflight=true(pure dry-check:dayz_test_worker.py:533-534returns before build and launching anything): without mods OK ·@DayZ_MCPOK ·@A6_SR2MOK ·@CFFAIL ·base_mods=["@CF"]FAIL.[EXACT][CLAIM-MCPV-JUNCTION-EXTRA-MODS-130] (Corrected 2026-10-26, stable box, DayZDiag 1.30.164200.) A Steam junction did pass in
extra_mods:@SurvivorAnims, a mount-point junction in!Workshop(Workshop item 2918418331), passedpreflight=truewith projectDayZ_MCPand then loaded in two runs, server and client (-mod=P:\Mods\@DayZ_MCP;P:\Mods\@SurvivorAnims;...in the RPT header,Worldmodule at 2349 files with its scripts). The rejection above was measured on 2026-08-02 with@CF. Run the preflight on the junction you need before going throughrequest-policy.json.When the mod under test does not compile, test a probe copy of it (added 2026-10-03, measured 2026-10-02, dayz-mcp run 606a5dbb) [EXACT][CLAIM-MCPV-PROBE-COPY], for example when its build packed another session's untracked test scripts:
- copy the deployed PBO entry for entry without those entries (every other entry byte-identical,
the same header properties, the SHA-1 trailer recomputed) into a probe folder of its own under
P:\Mods; - host it in another approved project whose
default_base_modscarry what the mod needs (CF and Dabs there): an absolute workshop path inbase_modsfails withbad_mod, since each entry must be a folder name or an absolute path inside the project'smod_roots(dayz_test_tool.py:399,dayz_test_request.py:222-233); - writing under
P:\Modschanges what every later run can load: ask the owner in chat first (in that session the agent harness's auto-mode permission check blocked the write until the owner agreed), and move the probe out ofP:\Modswhen done.
In that run the copy (641 of the 645 entries) compiled the mission module, and its models were the deployed ones byte for byte.
⚠ Before blaming the bridge for
version_blocked/last_poll_age_s=null: check thatkeyofdayz_mcp.jsonin project profiles is the SAME asDayZ_MCP_dev\tools\.dayz_mcp.key. A stale key gives 401 and bridge never polls, with symptom identical to "mod is not loaded". Key rotations leave behind projects not exercised since then (2026-08-02: MERCEDES_AMGLF and LFPowerGrid with old key, SUB_BRZ/LFHeli/DayZ_MCP with current key).⚠⚠ The symptom is NOT always
version_blocked/last_poll_age_s=null— and believing so costs the session just the same (measured 2026-08-17 in LFPowerGrid, with this warning already written above and overlooked). With stale key the peer can readversion_state: okwithlast_poll_age_sof 1.485 s: it is data from LAST good poll, which may belong to another already dead run. A dead bridge reads like healthy this way. The telltale signature is in game script log:[MCP-POC] poll error=5(server) /[MCP-CLIENT] client poll error=5(client). That 5 isEREST_ERROR_CLIENTERROR(scripts\3_game\http\restapi.c:16-17), meaning daemon returned 401 — 401 DOES NOT appear anywhere on game side. Two-command gate before blaming anything:- Compare key of each
dayz_mcp.jsonthe run will use against keyfile. These are two paths and$profile:wins over$mission:(MCPBridge.c:145-149,MCPClientBridge.c:234-237); if profiles key is missing, bridge falls back to mission key, which is usually the stale one. Same length (43) is NOT proof: compare the bytes. Invoke-WebRequest "http://127.0.0.1:8765/poll?key=<keyfile content>"→ 200{"commands":[]}proves daemon and live key are fine, isolating failure to run JSON. And fixing it DOES NOT require server restart:ReloadKeyAfterFailure(MCPBridge.c:373-398) rereads key when backoff hits ceiling. Fix JSON, wait ~20 s,[MCP-POC] poll key reloadedappears and peer returns tolast_poll_age_s0.2. That saves the boot.
⚠ Player spawn needle cannot presuppose gender. The example
Create entity type 'SurvivorM_—suggested bywait_fordescription itself— fails in silence with a female character: 300 s timeout with player already in world andCreate entity type 'SurvivorF_Helga'written to RPT. UseCreate entity type 'Survivor.- Public boundary
BRIDGE GOTCHAS THAT COST ONE RUN EACH (added 2026-08-18)
Verified in-game on 2026-08-18 during ATM cell test battery. All six cost at least one run each, and five of six produce a verdict accusing the mod without the mod having any issue (6 runs, 0 actual mod failures). Check them BEFORE writing first probe.
dayz_test_runDOES NOT load@DayZ_MCPby default, and returnssucceededanyway. Both processes launch alive, the game works, and the bridge stays silent. The symptom you receive (server_poll_stale/client_not_polling/ready.reason=no_run) points to bridge or key, not to mod set, so time is lost in the wrong place. Gate: passextra_mods=["@DayZ_MCP"]and confirm in server log that defineDayZ_MCPappears in all four modules, or look for[MCP-POC]lines.logs_sincedoes not accept its ownmarker. It RETURNS it as dict {path: [offsets]} and its parameter requires string: Pydantic ValidationError. Incremental draining is impossible. Workaround: read entire log at end of sequence and assign responses in order, verifying alignment with some field from event itself (a type, an id). If two probes share that field, markambiguous_alignmentinstead of guessing.wait_for(log_matches)does not work to wait for YOUR response. With defaultlookback_lines=200, a line from previous probe satisfies pattern instantly. And withlookback_lines=0it has been seen timing out with line already present in client log, returning a line from another flow inobserved. Use it as opportunistic wait; verdict comes from pollinglogs_since.Player inventory is NOT reset between runs. Second run finds what first left behind, nothing fits anymore and dotation fails silently (
create_failed, or endow of 0 units). To chain:dayz_test_stop, delete<mission>\storage_1\players.dbwith backup, and relaunch (~3 min). WARNING:clean=trueDOES NOT do this — forcesBuild=trueand passes-clearto AddonBuilder, meaning rebuilds the PBO and breaks binary identity, which is precisely what an A/B test cannot afford. Balance/state that mod saves in server profile SURVIVES deletion.inventory_givereturnscreate_failedwhether classname does not exist or does not fit. They are indistinguishable. Before blaming inventory, verify classname exists: grep in missiontypes.xmland test with a control vanilla item (Appleworks). And do not presuppose mod config classnames exist: they can point to a mod not loaded, in which case exact classname count gives 0 forever and entire feature looks broken without being broken.Do not believe a verdict without looking at client log. Many mod events are written by CLIENT, not server. In test battery originating this section, automatic verdict said FAIL or INCONCLUSIVE six times in a row while client log showed operations executing properly. Dangerous failure mode of a cell is not "does not measure": it is "measures poorly and accuses the mod", and that propagates to project documents. If a verdict says FAIL, open client log before recording it.
world_spawnwithok=1DOES NOT prove object stays where you asked (added 2026-08-21, DayZ-MCP council).IsSpawnReady(MCPBridge.c:2729-2753) takes ACTUAL position of object itself, searches objects at that position, and considers itself ready if it finds itself. An object is always where it is: gate is tautological and satisfied for any live object, even if physics is already ejecting it. Does not compare against REQUESTED position, does not measure drift, and does not require inter-tick stability. Real case: spawn at urban point withok=1and vehicle ending up at Y = -40 km. Distinct from inverted coords failure above, which times out: this gives false PASS. Rule: after anyworld_spawnthat matters, confirm withobject_inspectorentities_querythat distance to requested position is as expected, and repeat reading a few seconds later to rule out drift. A playbook verdict looking only atokis invalid.
THE LOOP
- Launch via
dayz-test.ps1 … -ExtraMods "@DayZ_MCP"(wait until client and server are in; BUG-009: client autoconnect is flaky — higher-ServerWait/ retry). - Gate
bridge_status(PREREQUISITES.3). - Scene conditions (comparable captures):
world_time_setat noon (all 5 argsyear/month/day/hour=12/minute=0are mandatory) andworld_weather_set(overcast=1.0). Overcast sky = diffuse light: eliminates specular glint from direct sun that burns weapons/materials to white and hides detail (A6_SR2M finding 2026-06-17). DO NOT usetime_multiplier=0before pending animations (freezes sim). - Playbook depending on mod type (below).
- Report with evidence.
PLAYBOOKS (pass/fail criteria by type)
All start from spawn. world_spawn(type=<classname>, pos=[x,y,z]) → PASS if ok and without
unknown_type/spawn_failed; save actual pos for subsequent steps.
Static object / container / building
Load: spawn ok (above). FAIL → classname does not resolve (mod not mounted: absolute
!Workshoppaths, see dayz-test-ingame; orCfgPatchesdoes not register).Visible + textured + winding: orbit camera (≥4 poses — front/side/back/overhead) with
camera_set+capture_screenshot; inspect each PNG. PASS = object visible (not invisible), without missing textures (no solid magenta/white/black), plausible proportions, without inverted faces or holes (winding). Hole/missing-face from one angle and solid from opposite = inverted winding. Before failing a model on textures, look at the vanilla terrain and sky in the same frame: if they are flat too, the client did not load its textures and the frame says nothing about the model (dayz-test-ingamereferences/session-findings.md, "H2R import (2026-10)").Collision:
scene_raycast(from_pos, to)(defaultmethod="rvproxy": LOD intersection in theintersectmode) aimed at the object in each ofgeom,viewandfire: the battery below for a small object or an item, multi-point for a building (walls, corners, floor). PASS = rays that should hit yieldhit=truewithobject_type/object_classof object. No hit where it should hit = ViewGeo/FireGeo missing or improperly resolved (LODs). (Changed 2026-10-03: this bullet asked for rays "from ≥2 angles", which do not tell the causes below apart.)Ray battery and its controls (added 2026-10-03; the battery of
dayz-p3d-audit"Absolute winding check" rule 6, fired on DayZDiag 1.29.163709 on 2026-10-02, dayz-mcp run 606a5dbb) [EXACT][CLAIM-MCPV-COLLISION-BATTERY]: nine rays per mode.- One vertical ray down the object's axis. On a solid box whose collision spans its own axis,
like the measured kit box, a LOD recentred by a missing
autocenter=0(dayz-p3d-auditkiller #3) still sits on that ray, so a miss there is not the recentring. On an object with gaps or several separate parts the shifted collision can leave the axis: check that the ray crosses the collider both where it is modelled and where the recentring would put it. - Four horizontal rays through the middle, one from each side, and two more heights on one side.
- Two rays from inside the object outward: a sound object answers
entry 0, exit 1at distance 0.
Fire the same battery, in the same run, at two controls: an object from the same PBO and load path, which separates the model from the build and the loading, and a vanilla object of the same physics layer (
physLayer;WoodenCratesets none and takesitem_smallfromInventory_Base,DZ/data/config.cpp:3152; for a static, theLand_Container_1Aohcontrol under WHAT IT DOES NOT COVER). In that run the LFPowerGrid kit box took 0 of 27, while the same PBO's logic-gate kit andWoodenCratetook 21 of 21 each (the battery without the two extra heights). The battery locates a miss; it does not name the cause. When the suspect is the winding, only a pair that changes the winding alone decides: the kit's paired run in rule 6 did; the Rule 12 pair, which also turned the Visual LOD and the normals, supports the cause without deciding it.- One vertical ray down the object's axis. On a solid box whose collision spans its own axis,
like the measured kit box, a LOD recentred by a missing
Picking an item up proves nothing about its collision (added 2026-10-03; vanilla 1.29 and bridge source read, the pickup seen in the same run) [EXACT][CLAIM-MCPV-PICKUP-NO-RAY]. Vanilla targeting casts
RaycastRVProxyfrom the camera, 5 m, with the defaultObjIntersectView(4_world/classes/useractionscomponent/actiontargets.c:241-255,3_game/global/dayzphysics.c:88), and at a camera pitch of −45° or lower it also takes the objects of a 30°, 3 m cone and scores them by their distance to that ray, with no hit on them needed (actiontargets.c:324-325,:531-532,:843-848). The take action's target condition then checks only the distance to the object's position:ActionTakeItemToHandsusesCCTObject(4_world/classes/useractionscomponent/actions/interact/actiontakeitemtohands.c:13,4_world/classes/useractionscomponent/targetconditionscomponents/cctobject.c:10-24), and its other conditions (takeable, not being placed or deleted, attachment state, room in the hands, and since 1.30 a code lock the player cannot detach;actiontakeitemtohands.c:31-46) do not look at collision either.action_usebuilds its world target from the nearest object of the class, withcomponentIndex -1and the object's position as the cursor hit, and casts no ray (MCPClientBridge.c:3769,:3783,:3861). In that runaction_use(ActionTakeItemToHands)put the kit box in the player's hands while all 27 rays of its battery missed it. Read collision fromscene_raycastagainst a same-layer control.intersect="view"is the cursor's mode, but the bridge casts a 0.05 m sphere when the radius is 0 (MCPBridge.c:2496,:2528-2531), where the vanilla ray has none. The cone path is read in the source, not measured: no MCP verb sets the camera pitch.Physics and player collision (added 2026-10-02, measured on DayZDiag 1.29.163709) [EXACT][CLAIM-MCPV-COLLISION-PROBES]:
scene_raycast(method="bullet", radius=0)castsDayZPhysics.RayCastBulletin the server's physics world, restricted to the layers BUILDING, DOOR, VEHICLE, ROADWAY, TERRAIN, ITEM_SMALL, ITEM_LARGE and FENCE (MCPBridge.c:2983-3004); the defaultrvproxyintersects the LODs instead. Count a hit only whenobject_typeis the target and the position is where its face should be: a vertical ray through a box without collision returnshit=1on the terrain. A miss alone does not prove the object has no physics shape (its layer may be outside the mask), and a hit does not mean the player collides (theWoodenCratebelow). For player collision the run teleported the player 4 m from the object's centre and calledplayer_move(to=<4 m past the centre, on the far side>, speed="walk", hold_s=10), then read the end position withquery_player_state(server). Blocked meansarrived:falseAND the player stopped just before the face, after the full hold (about 0.36 m before it on that run's 2 m boxes);released_by:"hold"only says the hold expired, andarrived:falseanywhere else is INCONCLUSIVE. Free meansarrived:true(7.5 m in about 5 s); walk the same distance and bearing on open ground as the free reference. Every walk of that run went north nearly straight ahead (applied_angle_deg0, and 0.04 on one walk), so it did not exercise a steering angle or its sign. A walk-through alone does not prove missing collision geometry: a vanillaWoodenCrate(item_small, low) took every ray and did not stop the player, while aHescoBox(item_large) stopped it; that run did not separate the crate's layer from its height.Axis-aligned test fixtures: try
rotation=64, then check the pose (added 2026-10-02, same run):rotation=0leaves the bridge'sRF_DEFAULT(512, the config's placement), which yawed anInventory_Baseprobe about 10 degrees (a side-face reading moved 0.22 m across 1.2 m of the face);rotation=64spawned the same box with its faces on the world axes (equal readings at both offsets), on flat concrete.scripts/3_game/ce/centraleconomy.cgives 64 two names,RF_IGNORE("object will spawn as model was created", :57) andRF_RANDOMROT(:63); the result agrees with the first comment and does not settle how the engine reads the value. Before trusting face coordinates, check the pose: two parallel rays at one height on a vertical face catch a yaw (unequal readings) but not a tilt, so also readscene_raycast(method="bullet")normals on two faces that are not parallel, such as a side and the top (a tilt about one face's normal leaves that normal unchanged), and check that each lies along its axis, allowing for the ground: the statics below read up to 0.31 degrees off the axis, the terrain about as much, the items on it. Read the fixture again a few seconds later. It may change an item's configured resting side; slopes and settling were not measured. Thenormalof a defaultrvproxyreply is not a face normal (added 2026-10-02, two runs) [EXACT][CLAIM-MCPV-RVPROXY-NORMAL]: the bridge copies the engine'sRaycastRVResult.dir(MCPBridge.c:2958), for a ray the "direction and size of the intersection" (scripts/3_game/global/dayzphysics.c:104), asenforce-script-reference("Surface normals") found on house walls. In this run and in run cf2d6bb3 it lay along the ray in all 190rvproxyhits on an object, the box yawed about 10 degrees included ((-1.963, 0, 0) and (-1.966, 0, 0)), and its length is the ray's path through the object: (-2.1, 0, 0) across a 2 m box is its width plus the 0.05 m sphere at each side.bulletreturns the hit face's normal (MCPBridge.c:2998): a unit vector in all 40 of its hits, up to 0.31 degrees off the axis on theHouseNoDestructstatics and on the axis on the items. (Corrected 2026-10-02: this entry read the pose from the returnednormallying along the axis, citing (-2.1, 0, 0); with the default method it always does.) An item can come out with its sides swapped, and none of these checks shows it (measured 2026-10-02 on DayZDiag 1.29.163709, dayz-mcp run cf2d6bb3) [EXACT][CLAIM-MCPV-ITEM-YAW]: one MLOD held two 2 m boxes, A (x -3..-1) inComponent01and B (x 1..3) inComponent02, and a twin had B in no component; each model was spawned, binarized and as MLOD, withflags=8389668androtation=64asHouseNoDestructand asInventory_Base(item_large). On the four statics the west box was A, as modelled: it answered as component 0 (Component01;Component02reads 1), and on the twins it was the only box that took rays. On the four items the boxes had swapped sides, as a turn of 180 degrees about Y leaves them: the west box answered as component 1, and on the twins every hit was component 0 on the east box while the west box took no ray. The boxes are symmetric in z, so these rays cannot tell that half-turn from a mirror in x. Parallel rays read equal on the items too, and the normals pointed the same way as on the statics, since the swap keeps every face on the axes; a fixture symmetric about its origin, like the single box above, cannot show it, andobject_inspectreturns memory points in model space (GetMemoryPointPos,MCPBridge.c:2240), not the pose. So give the fixture sides a ray can tell apart and check which answers where: one component per side, read from thecomponentofrvproxyhits (geom,vieworfire; the bridge leaves it 0 onbullethits,MCPBridge.c:2983-3004), or a collider on one side only. Put the parts off-centre in z as well to tell a half-turn from a mirror, and probe along z too: a quarter-turn moves parts that lie along x onto the z axis.telemetry_read(mode="object_at")returns the object'sGetOrientation()(MCPBridge.c:3137); that run did not read it. Not measured: whether the swap is fixed or random (all four items swapped; 64 is alsoRF_RANDOMROT), otherrotationandflagsvalues, base classes and layers, and the cause.Standing on top: move the player before reading its height (added 2026-10-02, measured on DayZDiag 1.29.163709 in a later run, on 2 m boxes) [EXACT][CLAIM-MCPV-STAND-PROBE]: a second reading of the object's physics body next to the walk-into probe above, on its top instead of a side face. An idle player is not a probe. After
player_teleport(pos=[x, <top + 0.07>, z], skip_clearance_check=true)the player kept the teleport height (y 340.6000) and did not fall, on a box top and above open ground alike: the server read it twice over 3-4 s (query_player_state), and above open ground the client's physics position (thestart_posthatplayer_movereports) still read it 15 s after the teleport (the worn-items section below saw the same after a teleport to y=160). Everyquery_player_stateof these probes read godmode on, the DayZ_MCP default; the run did not separate the bridge, godmode and the engine as the cause. A short move applies the fall: the run teleported the player above the centre of each top, calledplayer_move(angle_deg=0, speed="walk", hold_s=0.5), thenquery_player_state. Above open ground the player dropped to the ground (y 340.6000 to 338.459); on the boxes, threeInventory_Baseitems at least 4 m from their neighbours (two outward-wound, MLOD and binarized, and one inward-wound), the server's y matched the top to within 0.0001 m (spawn y + 2.000 m, wherescene_raycast(method="bullet")hit it) and its x, z lay 0.70-0.75 m from the centre, inside the 2 m top. Read the server's position after the move: with x, z still inside the top's footprint, y at the top's height there means the player stands on it and y at the ground means it fell through; an endpoint off the footprint, or any other height, is INCONCLUSIVE. Start above the centre of an isolated top (no other support at that height within the walk's reach, or the player can end on it), and run the same teleport and move above open ground as the falling reference. Keep the move short: 0.5 s took the player 0.70-0.75 m by the server's position read after the move, with the edge 1 m away; a top whose nearest edge is closer than that can be walked off in 0.5 s, and the endpoint check then reads INCONCLUSIVE.Placement:
telemetry_read(mode="object_at", type=<classname>, pos=<spawn_pos>, radius=2)→found=true,pos~ spawn, reasonableorientation. PASS = not buried or floating (cross-referencepos.ywith visuals).
Item / arma
- Load + visible/textured/winding as above, with emphasis on proportions vs real-world reference and on post-import geometry (orientation, winding) — where generated/imported meshes tend to fail.
telemetry_read object_at→attachment_count,health01. Telemetry PASS = found + healthy stats.- Collision: as for a static object, the ray battery with its two controls. Picking the item up, by
hand or with
action_use, does not test it (static-object playbook, "Picking an item up proves nothing about its collision").
Vehicle (placement/structure only)
- Load + visible + collision + telemetry.
vehicle_enter(pos)→seated=trueconfirms seat.telemetry_read→engine_on_server,wheel_count,fuel_fraction. vehicle_enterSEATS player (placement/seat). To DRIVE: drivability ladder (§DRIVABILITY +references/acceptance-ladder.md) covers autonomous driving via owner-side verbs (vehicle_get_in_client/engine_set/vehicle_control/vehicle_telemetry).
CAPTURAR ARMA ALZADA / ADS (raise client-side) [VERIFIED-SR2M gate iter36→37]
To validate RAISED/aimed weapon pose (e.g. support hand grip) there is no player input
tool. Weapon raise is forced IN THE MOD, client-side — captured character is client's LOCAL
player, so server init.c override DOES NOT alter the pose client renders
(exact symptom: server log states raised=1 but capture shows weapon LOWERED).
Drop-in: a modded class MissionGameplay in a gate mod (declares a missionScriptModule in its
class defs, files[]={"Mod/Scripts/5_Mission"}; build -PackOnly so .c survives packing):
modded class MissionGameplay
{
override void OnUpdate(float t)
{
super.OnUpdate(t);
PlayerBase p = PlayerBase.Cast(GetGame().GetPlayer()); if (!p) return;
HumanInputController h = p.GetInputController(); if (!h) return;
h.OverrideRaise(HumanInputControllerOverrideType.ENABLED, true); // ENABLED persiste; ONE_FRAME parpadea -> idle
}
};
Key takeaways:
- 3rd-person aim pose depends SOLELY on
IsRaised()(dayzplayerimplement.c:1742, AimingModel) → sustained raise is enough. DO NOT callSetIronsights(): forces ironsight camera which fights MCP free-cam, and server-side left weapon untextured. WeaponADS()(human.c:86) is INPUT flag without script override (human.c:253-274) → always 0 even if ADS works. Success signal =IsRaised()/ client log, NOTWeaponADS().- Keep
OverrideRaise(ENABLED)server-side (init.c) as well so server agrees. - Judge grip by hand at NATIVE resolution (crop orbit frame), never rescaled
contact-sheet. Grip mechanism + geometric parity: skill
dayz-animation-pipeline(references/weapon-in-hands.md).
TOOL → WHAT IT VERIFIES
| Tool | Verifies | Failure signal |
|---|---|---|
world_spawn |
classname loads | unknown_type / spawn_failed → mod not mounted |
camera_set + capture_screenshot |
render: visible, textures, winding, proportions | invisible / magenta / holes |
scene_raycast (default rvproxy) |
LOD collision (Geometry/ViewGeo/FireGeo, by intersect) |
no hit where it should hit |
scene_raycast(method="bullet") |
a physics shape on the masked layers (static-object playbook) | no hit on the target where it should hit; a miss alone is not proof |
player_move walk + query_player_state |
the player stopped by the object (static-object playbook) | arrived:true through it; a walk-through alone is not proof |
telemetry_read (object_at) |
placement, orientation, attachments, health | found=false / buried pos |
bridge_status |
peer liveness (gate) | last_poll_age_s=null → bridge down |
world_time_set / world_weather_set |
reproducible scene for captures | — |
WHAT IT DOES NOT COVER (ALWAYS declare in report)
Player and UI actions: open/close doors, interactive inventory, shooting, reloading, menus. There is no generic player input tool. (Driving IS covered — ladder §DRIVABILITY /
references/acceptance-ladder.mdwith owner-side verbs.) For a building, geometry/collision YES; doors NO → manual test.exec_enforcedoes not execute on the headless diag server (GATE4B-LIM, engine limitation like MakeScreenshot) — do not rely on it to "execute arbitrary verification logic".telemetry_readis exposed as-is (BUG-010/011/012, hardening pending): does not certify large JSONL fixtures or extreme ranges.CONTINUOUS actions (with progress bar):
action_useSTARTS them but does NOT COMPLETE them (measured 2026-09-07, LFPowerGrid, ledger entryfb-20260907-184749-3fc1). Returnsok=1, started=1and the server effect never arrives:OnFinishProgressServerdoes not fire and the target remains in the world (entities_queryat 0.003 m after two attempts with 60 s and 45 s waits).key_pressis not the way out: it documents itself "not OS input, key-up, hold", and a continuous action needs SUSTAINED input. This immediately kills dismantling, deploying/deploy, and crafting, which is precisely how a mod moves money and persistent objects.- What CAN be affirmed, and it is not little: the CONDITION is evaluated for real, so
started=1means "valid candidate" and serves as a measure that the mod offers the action on that target in that state. Always verify it with a negative control that must come out red (an excluded target returnscondition_failed), orstarted=1proves nothing. - Corollary that bites separately: since deploying from a kit is a continuous action, there is no
way via MCP to create a truly persistent object —
world_spawnuses default flags and does not survive restart. Add thatdayz_test_stopdoes not shut down gracefully (the next startup prints... was not closed. Always shut down the server gracefully) and no test ofOnStoreSave/OnStoreLoadbetween restarts is conclusive via MCP today. If the result comes out "did not persist", that is the harness, NOT the mod: do not report it as a bug.
- What CAN be affirmed, and it is not little: the CONDITION is evaluated for real, so
class Doors(Building) doors do NOT open by MCP withoutdoor_index(measured 2026-09-07, LFSecure I-0/I-4, PBO L4):action_use(ActionOpenDoors, classname=<door>)returnscondition_failedeven with the player 1.4 m away, because the bridge builds theActionTargetwithcomponentIndex=-1andBuilding.GetDoorIndex(-1)(native,3_game/entities/building.c:28, "index of the door based on the view geometry component index") resolves no door.object_anim(source=<door source>, phase=1.0)answersphase=1, but the engine's door controller returns it to its "wanted" phase (re-readphase=0) andIsDoorOpen(index)stays false, so every action that needs an open door also returnscondition_failed. Withoutdoor_index, opening/closing, sound and sync were a MANUAL test (the user's F) and MCP could only verify "starts closed" (object_animreadsphase=0); see the update below.Update (measured 2026-10-01, DayZDiag 1.29.163709, test building whose door component is a button on a moving piece) [EXACT]:
action_usewithActionOpenDoors/ActionCloseDoorsanddoor_indexDOES open and close the door even there.object_anim(SetAnimationPhaseNow) on the door's source still cannot hold a static pose: the door settles atajarwith phase 0.07 instead of the requested phase — doors cannot be posed statically. For the verdict, useobject_doorsand rays.setup_failedis a FALSE NEGATIVE for local instantaneous actions (IsLocal() && IsInstant():ActionTogglePlaceObject,ActionDropItemSimple; measured 2026-09-07): the bridge checksGetRunningAction()==nullright afterPerformActionStart, and an instantaneous action has already finished. The hologram appeared and the object fell to the ground in both cases. Rule: withsetup_failedon an instantaneous action, verify the EFFECT (capture,entities_queryof dropped object) before marking the action as failed.scene_raycast:entry=0is TERRAIN andnormalis NOT a unit normal (measured 2026-09-07 withintersect=fireandgeom): a vertical ray against the ground returnshit=1,object_type="",surface_type=cp_grass|cp_concrete2,entry=0,exit=0; against an object it returnsentry=1. Thenormalfield isRaycastRVResult.diras-is, which in line-object collision is "direction AND SIZE of the intersection" (3_game/global/dayzphysics.c:104): measured magnitudes 0.16-0.42. Its direction is the ray's own ("Axis-aligned test fixtures" above), so it says nothing about the face: never read it as a normal nor use it as a threshold (dot >= 0.9is never reached); a face normal comes frommethod="bullet". (Corrected 2026-10-02: this line said to use it only as direction (sign/axis).) A mod that filters!hit.entrydiscards natural ground (hologram that never snaps to ground: LFSecure I-1). (Qualified 2026-10-03: a ray that starts inside an object also readsentry=0: the inside rays of run 606a5dbb readentry 0, exit 1at distance 0 on the controls. Tell terrain by its emptyobject_type, not byentry=0alone;entry=1holds for hits from outside.)Collision probe with vanilla CONTROL before blaming the mesh (measured 2026-09-07, LFSecure L3 -> L4): if
scene_raycastdoes not hit a mod static, repeat the SAME ray (view, fire, and geom; from outside AND from inside) against a vanilla static spawned next to it (world_spawn(type="Land_Container_1Aoh"): flat walls,House,IsBuilding()true, also useful as a wall for holograms). Control HIT + mod MISS in all modes and from both sides = reversed collision convex components (inverted winding in Geometry/View/Fire: ODOL preserves them, engine does not see them); control HIT + mod HIT only from inside = reversed faces only in visuals. With the source winding restored (L4) the same ray gave the front face at +0.079 m and a wall of 0.42 m. The correction rule lives indayz-p3d-audit(0% agreement = discrepancy, not direction). (Qualified 2026-10-03: the two readings above name the likely cause, not a measured one; only a pair that changes the winding alone decides it, as "Ray battery and its controls" in the static-object playbook says.)
REPORTING
A report with: criteria table (criterion · PASS/FAIL/INFO · evidence), PNGs and telemetry/raycast JSONs as evidence, global verdict, and a "to manual test" section with what is not covered (doors, shooting, etc.). Context cost: each capture weighs ~25k tokens (~240-320 px) — batch orbit-pose captures and do NOT re-read a PNG except to verify something concrete (images inflate context quickly). Hard budget ~25k tokens/image: do not request higher resolution.
TROUBLESHOOTING
| Symptom | Cause | Fix |
|---|---|---|
claude mcp list → × Failed to connect |
claude mcp list is NOT reliable: gives false Failed with the server operational. E4 contention on port 8765 (ExclusiveThreadingHTTPServer lock, fail-closed) was the PRE-broker failure mode |
verify FIRST with ToolSearch (mcp__dayz-mcp__* tools available? → all OK, ignore Failed). Broker registration --client (since 2026-06-24) overcomes E4 contention: N sessions share the server without fighting for 8765, and orphan-guard releases orphans alone. If tools are truly missing: register in broker mode (--client) and open NEW session (tools load at startup) |
| tools do not appear in session | server was registered AFTER opening session | open a new session (MCPs load at startup) |
bridge_status.server_peer.last_poll_age_s = null |
server bridge does not poll | check dayz_mcp.json in server_profiles + key; confirm @DayZ_MCP mounted (absolute paths !Workshop) |
client_peer … null (server ok) |
client does not connect or client_profiles\dayz_mcp.json missing |
BUG-009 (flaky autoconnection): larger -ServerWait / retry; seed client config |
version_state = legacy_blocked |
--require-version ON against a bridge not sending ver= |
deploy 4B PBO (sends ver=4~…), or register server without --require-version for that run |
| byte-identical captures between poses | grab caught a stale desktop frame (not render) | [EXACT] PrintWindow(hwnd, hdc, PW_RENDERFULLCONTENT=2) on the client's main window DOES capture the D3D game area — full scene in >20 captures even with another window covering it and without taking focus, as long as the Windows session is unlocked and the display is on (measured in game, DayZDiag 1.29, 2026-09-28). A black game area means the display is off or the session is locked, not that the route cannot read D3D. CopyFromScreen only works with the window in front, and a covered DayZDiag also drops to ~20 FPS; validate on a crop of the game area, not whole-frame variance (corrects the LL-247 entry of 2026-08-12) |
| tool timeout and then "zombie" commands on reconnect | BUG-024: timeout leaves command queued; bridge executes it on return | after a timeout, reconcile with bridge_status before continuing |
REFERENCES
dayz-test-ingame— build/deploy/launch (this skill composes it with-ExtraMods "@DayZ_MCP").DayZ_MCP_dev\tools\README-mcp.md— startup order + bridge troubleshooting.DayZ_MCP_dev\HANDOFF.md(LIVE-STATE) — server invariants, 11 tools, GATE4B-LIM, backlog._shared\dayz-conventions.md— L2 (LODs, ViewGeo/FireGeo, DayZ response format).- Precedent of camera+orbit capture: A6's
gate-mcp.ps1(window-grab by orbit, dayz-test-ingame finding 2026-06-17) — alternative batch sweep to agent driving tools.
VEHICLE: repeatable smoke recipe (added 2026-06-24)
Visual smoke of a CarScript vehicle driven by MCP, verified end-to-end on MercedesAMGLF 2026-06-24. Repeatable without re-deriving:
Mission = stock
dayzOffline.chernarusplus(from DayZServer install), NOT a Phase 0 mount-probevoid main(): mount-probe spawns no player → client does not render the world and capture comes out black. Stock spawns player (CreateCharacter→CreatePlayer) → free-cam with rendered world. Pass it with-Mission "<...>\dayzOffline.chernarusplus".Bridge seed in profiles generated by
dayz-test.ps1:dayz_mcp.json{"url":"http://127.0.0.1:8765/","key":"<key>","pollHz":5}(ASCII) in<Mod>_dev\_server\profiles\and..\_client\profiles\. The bridge reads it from$profile:(serverMCPBridge.c:125, clientMCPClientBridge.c:211); key =DayZ_MCP_dev\tools\.dayz_mcp.key. Seed BEFORE launch (the bridge reads config in its init, only once). Do not useinstall-mcp.ps1 -Register(re-registers broker mode--client).Launch:
dayz-test.ps1 -Mod <Mod> -Mode all -Build -PackOnly -ExtraMods "@DayZ_MCP" -Mission "<...>\dayzOffline.chernarusplus" -ServerWait 240.-PackOnlymandatory in mods with.c(binarize drops them → NO_IGNITER). Call the.ps1via absolute path (P:\ is subst).Readiness: stock runs CE (
InitOffline) ~1-3 min; during CEbridge_statuscomes out STALE (last_poll_age_sgrows even thoughversion_state=ok) — do NOT spawn there. Wait forconnected to serverin server'sscript*.logand forbridge_statusto return fresh (<1 s). Poll log host-direct with PowerShell, NOT bash Monitor (inherits bindfs cache, LL-142).Smoke (grouped, R5):
world_time_set(year,month,day,hour=12,minute=0)(all 5 mandatory) +world_weather_set overcast=1.0→world_spawn <Class> pos≈[player+~10]→ wait 30 s (survives without native crash = LL-099 ruled out) →camera_set(cam_mode "lookat",cam_pos+look_at) +capture_screenshot+scene_raycast+telemetry_read object_at. Healthy telemetry =found=1,health01=1,velocity 0.Flaky capture: grab sometimes catches loading screen or client's "Continue" menu overlay instead of render. If it happens: wait for settle (~30-40 s in background — foreground sleep is blocked in this environment) and recapture; use several angles. Top-down that clearly reveals proxy (mis)alignment:
cam_pos=[carX+3, 16, carZ+1]look_at=[carX, ground+0.1, carZ]. ~25k tokens/image — batch captures, do not re-read a PNG except to verify something concrete. A frame flat everywhere, vanilla terrain and sky included, is not a flaky grab of the model: withhold its texture verdict, check that the client window has focus and capture again after settling, since textured runs have opened with 4-11 s of flat frames; if the scene stays flat, its visual evidence isSETUP_FAILand the run is relaunched (dayz-test-ingamereferences/session-findings.md, "H2R import (2026-10)").User's eye > capture when render is ambiguous: in MercedesAMGLF grab was flaky and user reading was reliable diagnosis of winding and proxy alignment. If there is a human in the loop, cross-check with them. (Case s3 2026-06-24: I was about to sign PASS on faint renders; user's eye caught proxy misalignment I could not resolve → correct FAIL.)
SHARED SESSION. Port 2302 and Steam client remain single resources, but exclusion is coordinated with the runbook's FIFO lease, not attributing processes nor evicting other sessions.
dayz_test_runacquires and maintains exclusion; retainrun_idand terminate that same run withdayz_test_stop. NEVER kill a process to unlock the box: if state does not reconcile, retain the process and declare degraded shutdown."Wheels = sheet/flat disk" is NOT a proxy bug. A
world_spawnof a CarScript without attachments leaveswheel_count=0/attachment_count=0→ only hub/disk is visible, without tires. Expected in visual smoke; real wheels require attachments (physics phase), out of scope for smoke.legacy_blocked("poll did not include ver=") is usually INCOMPLETE INIT, not version mismatch (added 2026-06-28). Refines the homonymous row in TROUBLESHOOTING: with 4B PBO deployed, right after launching serverbridge_statuscan come outlegacy_blocked/last_poll_age_s=nullbecause the bridge has not yet completed the handshake; transitions tookwithver=4~…when server mission LOADS completely (~1-2 min). Do NOT redeploy nor re-register for that — wait and re-check. It is only a real mismatch if it stayslegacy_blockedwith mission already loaded (client connected, world rendered). Origin: SUB_BRZ Phase 5 2026-06-28.Historical pattern superseded 2026-07-15.
cmd start/.batavoided a background job losing its child, but created a process outside the registered lifecycle. For any current smoke, use exclusivelydayz_test_run, retain itsrun_idand terminate only that run withdayz_test_stop. No unmanaged fallback exists to bypass the lifecycle guard.
DRIVABILITY + ACCEPTANCE LADDER (summary — detail in reference)
Phase 5 of DayZ-MCP project added and gated in-game owner-side verbs that DO drive the
car (client takes ownership and handles throttle/steer): vehicle_get_in_client(pos) (seats +
ownership), engine_set("start"/"stop"), vehicle_control(throttle, steer, brake, handbrake, hold_ttl_s) (SUSTAINED control, fail-closed), vehicle_telemetry(), vehicle_release(), and
query_get_in_condition(pos, component) (server peer, diagnoses which of the 7 gates of
ActionGetInTransport blocks). On top of them runs the acceptance ladder rip→drivable:
ordered rungs R1 spawns → R2 render → R3 get-in available → R4 seated → R5 drives → R6
wheel direction, each reading in-game ground-truth and mapping its failure to a known fix in the
SUB_BRZ taxonomy (dayz-vehicles/references/). Reference orchestrator references/drive_ladder.py
(drives R1→R6, outputs verdict.json; does NOT apply fixes nor rebuild) + fixtures tests/test_drive_ladder.py.
Full detail (verbs, verified owner-side mechanism, preconditions, spawn placement,
R2.5 restore-gameplay, anti-false-green guardrails, ladder table, and failure→fix mapping) →
references/acceptance-ladder.md.
(added 2026-07-14) No restore-gameplay tool: for user MANUAL test, relaunch without @DayZ_MCP
camera_set ALWAYS suppresses player control (SuppressGameplay() -> PlayerControlDisable, see
§R2.5) and there is NO MCP tool that fires RestoreGameplay() — only internal gate drive_probe_client
does it, and it is not exposed as a verb/tool. Practical consequence: after an MCP smoke that used camera_set
(free-cam or static-cam lookat), the USER cannot take player control to test the car by hand
— their input remains suppressed and there is no verb to revert it on the fly.
Rule: when workflow passes from MCP smoke (automatic) to user manual test (driving, judging
feel/aesthetics), relaunch the game WITHOUT @DayZ_MCP (or at least with a client that never received
camera_set in that session). A clean client has normal control. If the client was already neutered:
reconnecting (Esc -> Disconnect -> Reconnect) recreates the player with clean camera/control; there is no tool
shortcut. Origin: SUB_BRZ s35 — time was lost looking for a non-existent restore tool and user had
to close the game and ask to relaunch normally.
(added 2026-07-14) Autonomous smoke: world_spawn takes engine [x, y_up, z], and capture with sleeping display = black frame (SP-060)
Dos caveats de smoke MCP autonomo verificados in-vivo (SUB_BRZ s32):
world_spawntakes vector in ENGINE order[x, y_up, z_north](MCPBridge.c:1638Vector(x,y,z)), but bridge connect-log prints player position as<x, z_north, y_up>. Passing the log triplet VERBATIM spawns the object at ~6 km altitude ->IsSpawnReady(2.0 m radius) is never satisfied -> job timeout +found=0, and engine auto-deletes the orphan (NETWORK (E): Will delete object ... outside world coords). Cost 3 timeouts in a row. Rule: convert<x,z,y>->[x,y,z]before everyworld_spawn/camera_set; after a spawn timeout, grep RPT foroutside world coordsBEFORE retrying (distinguishes bad-coords from slow-spawn). Verifiable withscene_raycastto terrain (gives real y_up).- Capturing with the display asleep = BLACK frame even though the client runs (on 2026-07-11 a locked session with LockApp also gave black frames; since 2026-09-14 the capture backend fails outright on a locked session instead, see point 4). Fix: bring the client window to the foreground + input wiggle (mouse / F15) before EVERY
capture_screenshot. Misleading symptom: it looks like "the mod's render is broken" and it is the compositor. - (minor) Spawn adjacent to player can fall INSIDE a building -> probe 3-4
scene_raycastto clear terrain before choosing pos. - [EXACT] A locked Windows session is a distinct capture failure (added 2026-09-14, SUB_BRZ s93): with
LogonUIalive,capture_screenshotdoes not return a black frame — it fails outright withcapture_backend_failedon every attempt while the game client stays alive and probing. Before budgeting captures in an unattended run, runGet-Process LogonUIfirst: with a locked session there is no host capture at all, for the MCP capture and for any window-grab script. (measured in game, DayZ 1.30.164014 Exp)
Origin: SUB_BRZ s32 MCP smoke (2026-07-11): 3 world_spawn timeouts with raw log pos + 2 black captures with LockApp; both resolved with the above.
(added 2026-07-14) Telemetry numeric gates: calibrate against MEASURED PHYSICAL FLOOR, not ideal fixtures (SP-061)
Invariant for any acceptance-ladder / numeric gate on in-game telemetry (car drive_ladder, spikes, future harnesses):
- A threshold calibrated with ideal mathematical fixtures is a gate that no real physics passes. The engine has a native noise floor (body dither between SetVelocity and reading; client network interpolator). Before setting a threshold: MEASURE the floor in a real run (metric p95 on real data), threshold = floor x margin (>=2x), and document the measured numbers next to the constant. Two
Files (dayz-modding-knowledge-pack)
-
references
-
acceptance-ladder.md 21.2 KB
Loading… -
dayz-1-30-mcp-verify.md 15.2 KB
Loading… -
drive_ladder.py 21.4 KB
Loading…
-
-
tests
-
test_drive_ladder.py 9.4 KB
Loading…
-
-
SKILL.md 109.4 KB
--- name: dayz-mcp-verify description: > Auto-test a DayZ mod in-game through dayz-mcp without keyboard or OCR: spawn a classname, orbit/capture, raycast collision, inspect placement/attachments/telemetry and emit PNG+JSON evidence. Compose with dayz-test-ingame; dayz_test_run/dayz_test_stop own the managed lifecycle, then this skill drives the bridge. Use for "auto-probar el mod", "verificar in-game con MCP", visual smoke/re-test, "comprobar que el .p3d carga/se ve/colisiona", automated spawn+capture+raycast, or the vehicle acceptance ladder (spawn→render→get-in→drive→wheel direction). Covers static objects, items/weapons, buildings without doors and vehicle placement/drivability. Player UI, door interaction, inventory use and firing remain manual. Also: ScriptConsoleTabRegistry, IsHeadless roboclient, UA_AM_INIT action inject. --- # DayZ MCP verify — auto-test in-game vía tools MCP ## GATE 0 — ask WHO is driving, before touching anything (added 2026-08-07) (until 1.29: managed `dayz_test_run` / `dayz_test_stop` and Mode=all capture still hold.) (since 1.30: Diag Script Console tabs register via `ScriptConsoleTabRegistry`; `IsHeadless()` names roboclients; SP `PerformActionStart` no-ops while pending. Details: [dayz-1-30-mcp-verify.md](references/dayz-1-30-mcp-verify.md).) **Before the first `dayz-mcp` tool in a verification run, ask the user whether they are driving or the MCP is.** Neither is assumed. A single question per run, not per capture. Why it is a gate and not a preference: **with the user present, their manual cycle is faster and yields better results** — spawns with VPP in seconds and evaluates the whole scene at a glance, while MCP loop needs to position camera, capture, and chain tools. MCP wins when the user is NOT there: early morning, unattended run, or while attending another front. Furthermore, changing mode is not free midway (clean manual test requires relaunching without `@DayZ_MCP`), so the decision is made at the start of the cycle or paid twice. | Situation | Who drives | |---|---| | User present and available | **Them** (VPP + direct judgment) — MCP only if they ask | | User absent / unattended run | **MCP**, and report is left for them with evidence | | Doubt | **Ask.** Never assume | What does NOT change: MCP remains mandatory for what a human cannot measure by eye (telemetry, numerical raycast, exact placement, repeatable series). There you do not ask, you use it. Measurement that originated gate (2026-08-07, corpus of 276 sessions): 63% of wall-clock time in a full cycle falls on user side, and 71% of their actual verifications are automatable with today's tools. That it is automatable does not imply it should be automated: with them present, their cycle rules. ## WHAT THIS DOES Drives an already launched DayZDiag client, via `dayz-mcp` MCP server tools, to verify a mod without human intervention: spawns object, views it from multiple angles (camera + window-grab), checks collision via raycast, reads placement/telemetry, and produces a report with pass/fail criteria and evidence. It is the automated **observation** loop bridging the gap between "PBO compiles/deploys" and "looks and behaves properly in-game", for the class of properties that DO NOT require keyboard input. Server-authoritative + passive pixel capture. Control and data are engine-native (spawn, raycast, telemetry, camera); sole non-native piece is visual capture (window-grab of rendered client — `MakeScreenshot` is broken in diag, T165276). ## COMPOSITION — public lifecycle + MCP verification Build/deploy/launch belongs to **`dayz-test-ingame`** (its environment preflight runs on its own: `P:\` mounted, junction `P:\Mods`, AddonBuilder, `allowFilePatching`). This skill invokes it with MCP mod added and then drives. Bridge verbs do not start or terminate processes; public test tools do orchestrate managed lifecycle. Before operational flow, read canonical protocol: `<runbooks>\dayz-mcp-agent-session-protocol.md`. [EXACT][CLAIM-R21-MCP-ORCHESTRATED-TEST] Execute lifecycle with `dayz_test_run` and stop it with `dayz_test_stop` on exact `run_id`. Both tools possess FIFO queue, lease, heartbeat, and release; do not wrap them in a second `session_acquire`. Reserve `session_*` primitives for low-level mutations not already encapsulated. Every lifecycle is identified by `run_id`: sharing mod does not grant ownership. Under retail quarantine only reads are allowed; if whoever opened retail cannot close it via UI, declare `manual_cleanup_required`. **Bridge startup sequence (rev. 2026-09-06): `dayz_test_run` → `session_acquire_wait` → verbs. Never probe before adopting.** "Do not wrap launch in another lease" does not mean run remains yours: upon completion, `dayz_test_run` RELEASES its owner lease (daemon audit: `session_release_finished reason=owner_release` right after launch) and run transitions to `RUNNING_IDLE` without owner (`release_owner`, `process_lifecycle.py:958-976`). From that instant the idle run fence rejects ALL `/enqueue` on that run, **reads included**: `_enqueue_run_rejection` discards flag `mutation` (`loopback.py:1293-1296`) and for `RUNNING_IDLE` returns `run_not_owned` (`:1317-1318`) with hint "This run has no owner (RUNNING_IDLE). Adopt the existing run before dispatching." (`:99-102`, `:1323-1329`). The counter-intuitive part: `bridge_status` counter is named `fence.mutation_rejects_by_code` (`:2821-2822`) and counts reads as well. Every verb crossing bridge falls there (`query_*`, telemetry, raycast, captures, `world_spawn`…) and also `wait_for(players_at_least|players_at_most)`, which polls with `query_all_players` (`server.py:2645-2650`). `wait_for(log_matches)` does NOT fall: reads log files without touching bridge (`server.py:2603-2637`, `:2686-2697`). 1. `dayz_test_run(...)` → retain `run_id`. 2. Optional without adopting: `wait_for(log_matches, "…")` to wait for startup via RPT. 3. `session_acquire_wait(purpose=…)`: grant adopts the ONLY ownerless `RUNNING_IDLE` run and declares it in `adopted_run` (`_adopt_on_grant`, `loopback.py:3119-3120`, `:3193-3230`). Check `adopted_run.ok` before first verb. With multiple idle runs it answers `adopted_run.error = "multiple_idle_runs"` and adopts none (`:3202-3208`): `dayz_test_stop` surplus run and repeat. 4. `wait_for(players_at_least, 1)` and remaining verbs. - [EXACT] The adopted lease expires by itself (added 2026-09-27, LFPowerGrid, three runs): it lasts 120 s and is NOT renewed internally, fresh MCP client or not. Send `session_heartbeat` at least every ~90 s during the whole run, including while a human plays. Past the deadline the run goes ownerless again (`grace.remaining_s` counts down) and the daemon stops it (`lifecycle_stop_outcome: stopped`), even with the user in front of the game. (measured in game, DayZ 1.30.164014 Exp) 5. `session_release` → `dayz_test_stop(run_id)`. With an MCP client prior to `8f5727f` of DayZ_MCP (2026-09-06) rejection arrived as a BARE `remote_error`, without code or hint (ledger entry `fb-20260906-193626-f45d`, first attributed to a client without polling); with current client it arrives as `run_not_owned: This run has no owner (RUNNING_IDLE)…`. In both cases remedy is adopting, neither retrying nor relaunching. This rule corrects step 1 of recipe SP-292 and clarifies first bullet of SP-152, below. Two lease failures around long waits (measured in game sessions, 2026-09-28): (1) `wait_for` over the log does NOT renew the lease — a wait longer than ~2 min expired with `lease_expired`; interleave `session_heartbeat` between waits. (2) When a user message interrupts an in-flight call (a long `wait_for`), the next `session_heartbeat` answers `lease_invalid: token was never valid for this client`; if the run is still alive, `session_acquire_wait` readopts it. Also: `dayz_test_close` shuts the run down in order, with a termination line on both sides. [EXACT] ### Companion externo: dayz-labs [EXACT][CLAIM-R21-MCP-COMPANION-AUTHORITY] `external_companion_no_lifecycle_authority`. Release dayz-labs v0.1.35, reviewed at pinned commit `dbd6ad3e54e30c81a9aeb88fcb9f60f007804c2a`, can serve as reference or optional companion. Its `start/stop/restart` verbs are excluded while DayZ_MCP holds a run: only `dayz_test_run` / `dayz_test_stop` on exact `run_id` govern that lifecycle. Do not install or update companion as part of a gate; always document the examined `pinned_version`. Its interface is WPF and captures its own application: `wpf_not_layout_evidence`. A correct capture from dayz-labs does not validate the parser, proportions, clipping, or semantics of a DayZ `.layout`; UI viewer needs its own render and comparison against in-game evidence. The `dayz_test_run` request must select `Mode=all`, build when PBO changed, and `@DayZ_MCP` as additional dependency: - `-Mode all` (server + client) is MANDATORY: visual capture reads from **rendered client**. A headless `-Mode server` has no window to grab. - `@DayZ_MCP` is loaded alongside the mod under test; both peers (server + client bridge) poll the MCP server loopback. ## PREREQUISITES (gate before driving) 1. **`dayz-mcp` registered and connectable.** PRIMARY gate: are tools `mcp__dayz-mcp__*` available in session? (ToolSearch finds them → gate PASS). `claude mcp list` is SECONDARY and NOT reliable: gives false `× Failed to connect` with server operational — do not block on it if ToolSearch sees tools. Tools are only available in a session started WITH server already registered (broker registration `--client`, see TROUBLESHOOTING). 2. **Bridge config seeded.** In-game bridge reads `dayz_mcp.json` (url + key + pollHz) from its profiles; MCP server uses `<DayZ_MCP_dev>\tools\.dayz_mcp.key`. Seed the SAME key in launch profiles with `install-mcp.ps1` pointing to directories generated by `dayz-test.ps1`: ```powershell <DayZ_MCP_dev>\tools\install-mcp.ps1 ` -ServerProfiles "<TargetMod>_dev\_server\profiles" ` -ClientProfiles "<TargetMod>_dev\_client\profiles" ` -MissionPath "<...>\mpmissions\dayzOffline.chernarusplus" ``` `[verify on first run]` — generic seeding (arbitrary mod + `@DayZ_MCP` via `dayz-test.ps1`) was until now only exercised through `run-fase3.ps1`. The FIRST run of this skill validates that bridge finds config in these paths; if `bridge_status` reports `last_poll_age_s = null`, config is not where bridge looks for it — adjust path and declare it. 3. **`bridge_status` green.** ALWAYS first call: `bridge_status`. `server_peer` must have fresh `last_poll_age_s` (not null). For capture, `client_peer` as well. `never polled` → bridge did not start (config/key/startup) → abort and diagnose, do not keep driving blindly. 4. **Mods passed via `extra_mods`/`base_mods` must be REAL DIRECTORY under `P:\Mods` (→ `!Workshop`); Steam junctions ARE NOT valid.** Two layers, confusing them costs a session: - Public boundary `_valid_public_mod` (`DayZ_MCP_dev\tools\dayz_mcp\dayz_test_tool.py:75-85`) rejects `:`, `\` and `/` across **all** projects → absolute workshop path never passes through tool. That is by design, NOT a gap in your project's policy. *(Updated 2026-10-03: the boundary now also accepts an absolute path inside the selected project's `mod_roots`; `_valid_public_mod` calls `dayz_test_request._valid_mod_entry` (`dayz_test_tool.py:269-272`, `dayz_test_request.py:222-233`, DayZ_MCP_dev `506de5d`). A workshop path lies outside them and still fails, with `bad_mod`.)* - Subscribed mods exposed by Steam in `!Workshop` (`@CF`, `@Dabs Framework`, `@VPPAdminTools`) are **Junctions** to `steamapps\workshop\content\221100\<id>`, and path identity guard rejects them → `dayz_test_failed`, which is the generic catch-all of `server.py:1101` and **swallows the real cause**. **Operational consequence**: CF/Dabs/VPP are only loaded by declaring them with **absolute path** in project `default_base_mods` inside launcher `request-policy.json` — which is how SUB_BRZ and LFHeli declare them. Mods that are real directories (`@DayZ_MCP`, `@LFHeliCore`, mod under test) do go via `extra_mods` with relative name. Diagnosis in one command: `Get-Item '<!Workshop>\<@Mod>' -Force | Select-Object LinkType` → `Junction` = relative does not work. Measured 2026-08-02 with `preflight=true` (pure dry-check: `dayz_test_worker.py:533-534` returns before build and launching anything): without mods OK · `@DayZ_MCP` OK · `@A6_SR2M` OK · `@CF` FAIL · `base_mods=["@CF"]` FAIL. [EXACT][CLAIM-MCPV-JUNCTION-EXTRA-MODS-130] *(Corrected 2026-10-26, stable box, DayZDiag 1.30.164200.)* A Steam junction did pass in `extra_mods`: `@SurvivorAnims`, a mount-point junction in `!Workshop` (Workshop item 2918418331), passed `preflight=true` with project `DayZ_MCP` and then loaded in two runs, server and client (`-mod=P:\Mods\@DayZ_MCP;P:\Mods\@SurvivorAnims;...` in the RPT header, `World` module at 2349 files with its scripts). The rejection above was measured on 2026-08-02 with `@CF`. Run the preflight on the junction you need before going through `request-policy.json`. **When the mod under test does not compile, test a probe copy of it** (added 2026-10-03, measured 2026-10-02, dayz-mcp run 606a5dbb) [EXACT][CLAIM-MCPV-PROBE-COPY], for example when its build packed another session's untracked test scripts: - copy the deployed PBO entry for entry without those entries (every other entry byte-identical, the same header properties, the SHA-1 trailer recomputed) into a probe folder of its own under `P:\Mods`; - host it in another approved project whose `default_base_mods` carry what the mod needs (CF and Dabs there): an absolute workshop path in `base_mods` fails with `bad_mod`, since each entry must be a folder name or an absolute path inside the project's `mod_roots` (`dayz_test_tool.py:399`, `dayz_test_request.py:222-233`); - writing under `P:\Mods` changes what every later run can load: ask the owner in chat first (in that session the agent harness's auto-mode permission check blocked the write until the owner agreed), and move the probe out of `P:\Mods` when done. In that run the copy (641 of the 645 entries) compiled the mission module, and its models were the deployed ones byte for byte. ⚠ **Before blaming the bridge for `version_blocked`/`last_poll_age_s=null`**: check that `key` of `dayz_mcp.json` in project profiles is the SAME as `DayZ_MCP_dev\tools\.dayz_mcp.key`. A stale key gives 401 and bridge never polls, with symptom identical to "mod is not loaded". Key rotations leave behind projects not exercised since then (2026-08-02: MERCEDES_AMGLF and LFPowerGrid with old key, SUB_BRZ/LFHeli/DayZ_MCP with current key). ⚠⚠ **The symptom is NOT always `version_blocked` / `last_poll_age_s=null`** — and believing so costs the session just the same (measured 2026-08-17 in LFPowerGrid, with this warning already written above and overlooked). With stale key the peer can read **`version_state: ok` with `last_poll_age_s` of 1.485 s**: it is data from LAST good poll, which may belong to another already dead run. A dead bridge reads like healthy this way. The telltale signature is in game script log: `[MCP-POC] poll error=5` (server) / `[MCP-CLIENT] client poll error=5` (client). That **5 is `EREST_ERROR_CLIENTERROR`** (`scripts\3_game\http\restapi.c:16-17`), meaning daemon returned 401 — 401 DOES NOT appear anywhere on game side. Two-command gate before blaming anything: - Compare key of **each** `dayz_mcp.json` the run will use against keyfile. These are two paths and `$profile:` wins over `$mission:` (`MCPBridge.c:145-149`, `MCPClientBridge.c:234-237`); if profiles key is missing, bridge falls back to mission key, which is usually the stale one. Same length (43) is NOT proof: compare the bytes. - `Invoke-WebRequest "http://127.0.0.1:8765/poll?key=<keyfile content>"` → 200 `{"commands":[]}` proves daemon and live key are fine, isolating failure to run JSON. **And fixing it DOES NOT require server restart**: `ReloadKeyAfterFailure` (`MCPBridge.c:373-398`) rereads key when backoff hits ceiling. Fix JSON, wait ~20 s, `[MCP-POC] poll key reloaded` appears and peer returns to `last_poll_age_s` 0.2. That saves the boot. ⚠ **Player spawn needle cannot presuppose gender.** The example `Create entity type 'SurvivorM_` —suggested by `wait_for` description itself— **fails in silence** with a female character: 300 s timeout with player already in world and `Create entity type 'SurvivorF_Helga'` written to RPT. Use `Create entity type 'Survivor`. ## BRIDGE GOTCHAS THAT COST ONE RUN EACH (added 2026-08-18) Verified in-game on 2026-08-18 during ATM cell test battery. All six cost at least one run each, and **five of six produce a verdict accusing the mod without the mod having any issue** (6 runs, 0 actual mod failures). Check them BEFORE writing first probe. 1. **`dayz_test_run` DOES NOT load `@DayZ_MCP` by default, and returns `succeeded` anyway.** Both processes launch alive, the game works, and the bridge stays silent. The symptom you receive (`server_poll_stale` / `client_not_polling` / `ready.reason=no_run`) points to bridge or key, not to mod set, so time is lost in the wrong place. **Gate**: pass `extra_mods=["@DayZ_MCP"]` and confirm in server log that define `DayZ_MCP` appears in all four modules, or look for `[MCP-POC]` lines. 2. **`logs_since` does not accept its own `marker`.** It RETURNS it as dict {path: [offsets]} and its parameter requires string: Pydantic ValidationError. Incremental draining is impossible. **Workaround**: read entire log at end of sequence and assign responses in order, verifying alignment with some field from event itself (a type, an id). If two probes share that field, mark `ambiguous_alignment` instead of guessing. 3. **`wait_for(log_matches)` does not work to wait for YOUR response.** With default `lookback_lines=200`, a line from previous probe satisfies pattern instantly. And with `lookback_lines=0` it has been seen timing out with line already present in client log, returning a line from another flow in `observed`. Use it as opportunistic wait; verdict comes from polling `logs_since`. 4. **Player inventory is NOT reset between runs.** Second run finds what first left behind, nothing fits anymore and dotation fails silently (`create_failed`, or endow of 0 units). To chain: `dayz_test_stop`, delete `<mission>\storage_1\players.db` with backup, and relaunch (~3 min). **WARNING**: `clean=true` DOES NOT do this — forces `Build=true` and passes `-clear` to AddonBuilder, meaning **rebuilds the PBO** and breaks binary identity, which is precisely what an A/B test cannot afford. Balance/state that mod saves in server profile SURVIVES deletion. 5. **`inventory_give` returns `create_failed` whether classname does not exist or does not fit.** They are indistinguishable. Before blaming inventory, verify classname exists: grep in mission `types.xml` and test with a control vanilla item (`Apple` works). And do not presuppose mod config classnames exist: they can point to a mod not loaded, in which case exact classname count gives 0 forever and entire feature looks broken without being broken. 6. **Do not believe a verdict without looking at client log.** Many mod events are written by CLIENT, not server. In test battery originating this section, automatic verdict said FAIL or INCONCLUSIVE six times in a row while client log showed operations executing properly. Dangerous failure mode of a cell is not "does not measure": it is **"measures poorly and accuses the mod"**, and that propagates to project documents. If a verdict says FAIL, open client log before recording it. - **`world_spawn` with `ok=1` DOES NOT prove object stays where you asked** (added 2026-08-21, DayZ-MCP council). `IsSpawnReady` (`MCPBridge.c:2729-2753`) takes ACTUAL position of object itself, searches objects at that position, and considers itself ready if it finds itself. An object is always where it is: gate is tautological and satisfied for any live object, even if physics is already ejecting it. Does not compare against REQUESTED position, does not measure drift, and does not require inter-tick stability. Real case: spawn at urban point with `ok=1` and vehicle ending up at **Y = -40 km**. Distinct from inverted coords failure above, which times out: this gives false PASS. **Rule: after any `world_spawn` that matters, confirm with `object_inspect` or `entities_query` that distance to requested position is as expected, and repeat reading a few seconds later to rule out drift.** A playbook verdict looking only at `ok` is invalid. ## THE LOOP 1. **Launch** via `dayz-test.ps1 … -ExtraMods "@DayZ_MCP"` (wait until client and server are in; BUG-009: client autoconnect is flaky — higher `-ServerWait` / retry). 2. **Gate** `bridge_status` (PREREQUISITES.3). 3. **Scene conditions** (comparable captures): `world_time_set` at noon (all 5 args `year/month/day/hour=12/minute=0` are mandatory) and `world_weather_set(overcast=1.0)`. Overcast sky = diffuse light: eliminates specular glint from direct sun that burns weapons/materials to white and hides detail (A6_SR2M finding 2026-06-17). DO NOT use `time_multiplier=0` before pending animations (freezes sim). 4. **Playbook** depending on mod type (below). 5. **Report** with evidence. ## PLAYBOOKS (pass/fail criteria by type) All start from spawn. `world_spawn(type=<classname>, pos=[x,y,z])` → PASS if `ok` and without `unknown_type`/`spawn_failed`; save actual `pos` for subsequent steps. ### Static object / container / building - **Load**: spawn ok (above). FAIL → classname does not resolve (mod not mounted: absolute `!Workshop` paths, see dayz-test-ingame; or `CfgPatches` does not register). - **Visible + textured + winding**: orbit camera (≥4 poses — front/side/back/overhead) with `camera_set` + `capture_screenshot`; inspect each PNG. PASS = object visible (not invisible), without missing textures (no solid magenta/white/black), plausible proportions, without inverted faces or holes (winding). Hole/missing-face from one angle and solid from opposite = inverted winding. Before failing a model on textures, look at the vanilla terrain and sky in the same frame: if they are flat too, the client did not load its textures and the frame says nothing about the model (`dayz-test-ingame` `references/session-findings.md`, "H2R import (2026-10)"). - **Collision**: `scene_raycast(from_pos, to)` (default `method="rvproxy"`: LOD intersection in the `intersect` mode) aimed at the object in each of `geom`, `view` and `fire`: the battery below for a small object or an item, multi-point for a building (walls, corners, floor). PASS = rays that should hit yield `hit=true` with `object_type`/`object_class` of object. No hit where it should hit = ViewGeo/FireGeo missing or improperly resolved (LODs). *(Changed 2026-10-03: this bullet asked for rays "from ≥2 angles", which do not tell the causes below apart.)* - **Ray battery and its controls** (added 2026-10-03; the battery of `dayz-p3d-audit` "Absolute winding check" rule 6, fired on DayZDiag 1.29.163709 on 2026-10-02, dayz-mcp run 606a5dbb) [EXACT][CLAIM-MCPV-COLLISION-BATTERY]: nine rays per mode. - One vertical ray down the object's axis. On a solid box whose collision spans its own axis, like the measured kit box, a LOD recentred by a missing `autocenter=0` (`dayz-p3d-audit` killer #3) still sits on that ray, so a miss there is not the recentring. On an object with gaps or several separate parts the shifted collision can leave the axis: check that the ray crosses the collider both where it is modelled and where the recentring would put it. - Four horizontal rays through the middle, one from each side, and two more heights on one side. - Two rays from inside the object outward: a sound object answers `entry 0, exit 1` at distance 0. Fire the same battery, in the same run, at two controls: an object from the same PBO and load path, which separates the model from the build and the loading, and a vanilla object of the same physics layer (`physLayer`; `WoodenCrate` sets none and takes `item_small` from `Inventory_Base`, `DZ/data/config.cpp:3152`; for a static, the `Land_Container_1Aoh` control under WHAT IT DOES NOT COVER). In that run the LFPowerGrid kit box took 0 of 27, while the same PBO's logic-gate kit and `WoodenCrate` took 21 of 21 each (the battery without the two extra heights). The battery locates a miss; it does not name the cause. When the suspect is the winding, only a pair that changes the winding alone decides: the kit's paired run in rule 6 did; the Rule 12 pair, which also turned the Visual LOD and the normals, supports the cause without deciding it. - **Picking an item up proves nothing about its collision** (added 2026-10-03; vanilla 1.29 and bridge source read, the pickup seen in the same run) [EXACT][CLAIM-MCPV-PICKUP-NO-RAY]. Vanilla targeting casts `RaycastRVProxy` from the camera, 5 m, with the default `ObjIntersectView` (`4_world/classes/useractionscomponent/actiontargets.c:241-255`, `3_game/global/dayzphysics.c:88`), and at a camera pitch of −45° or lower it also takes the objects of a 30°, 3 m cone and scores them by their distance to that ray, with no hit on them needed (`actiontargets.c:324-325`, `:531-532`, `:843-848`). The take action's target condition then checks only the distance to the object's position: `ActionTakeItemToHands` uses `CCTObject` (`4_world/classes/useractionscomponent/actions/interact/actiontakeitemtohands.c:13`, `4_world/classes/useractionscomponent/targetconditionscomponents/cctobject.c:10-24`), and its other conditions (takeable, not being placed or deleted, attachment state, room in the hands, and since 1.30 a code lock the player cannot detach; `actiontakeitemtohands.c:31-46`) do not look at collision either. `action_use` builds its world target from the nearest object of the class, with `componentIndex -1` and the object's position as the cursor hit, and casts no ray (`MCPClientBridge.c:3769`, `:3783`, `:3861`). In that run `action_use(ActionTakeItemToHands)` put the kit box in the player's hands while all 27 rays of its battery missed it. Read collision from `scene_raycast` against a same-layer control. `intersect="view"` is the cursor's mode, but the bridge casts a 0.05 m sphere when the radius is 0 (`MCPBridge.c:2496`, `:2528-2531`), where the vanilla ray has none. The cone path is read in the source, not measured: no MCP verb sets the camera pitch. - **Physics and player collision** (added 2026-10-02, measured on DayZDiag 1.29.163709) [EXACT][CLAIM-MCPV-COLLISION-PROBES]: `scene_raycast(method="bullet", radius=0)` casts `DayZPhysics.RayCastBullet` in the server's physics world, restricted to the layers BUILDING, DOOR, VEHICLE, ROADWAY, TERRAIN, ITEM_SMALL, ITEM_LARGE and FENCE (`MCPBridge.c:2983-3004`); the default `rvproxy` intersects the LODs instead. Count a hit only when `object_type` is the target and the position is where its face should be: a vertical ray through a box without collision returns `hit=1` on the terrain. A miss alone does not prove the object has no physics shape (its layer may be outside the mask), and a hit does not mean the player collides (the `WoodenCrate` below). For player collision the run teleported the player 4 m from the object's centre and called `player_move(to=<4 m past the centre, on the far side>, speed="walk", hold_s=10)`, then read the end position with `query_player_state` (server). Blocked means `arrived:false` AND the player stopped just before the face, after the full hold (about 0.36 m before it on that run's 2 m boxes); `released_by:"hold"` only says the hold expired, and `arrived:false` anywhere else is INCONCLUSIVE. Free means `arrived:true` (7.5 m in about 5 s); walk the same distance and bearing on open ground as the free reference. Every walk of that run went north nearly straight ahead (`applied_angle_deg` 0, and 0.04 on one walk), so it did not exercise a steering angle or its sign. A walk-through alone does not prove missing collision geometry: a vanilla `WoodenCrate` (`item_small`, low) took every ray and did not stop the player, while a `HescoBox` (`item_large`) stopped it; that run did not separate the crate's layer from its height. - **Axis-aligned test fixtures: try `rotation=64`, then check the pose** (added 2026-10-02, same run): `rotation=0` leaves the bridge's `RF_DEFAULT` (512, the config's placement), which yawed an `Inventory_Base` probe about 10 degrees (a side-face reading moved 0.22 m across 1.2 m of the face); `rotation=64` spawned the same box with its faces on the world axes (equal readings at both offsets), on flat concrete. `scripts/3_game/ce/centraleconomy.c` gives 64 two names, `RF_IGNORE` ("object will spawn as model was created", :57) and `RF_RANDOMROT` (:63); the result agrees with the first comment and does not settle how the engine reads the value. Before trusting face coordinates, check the pose: two parallel rays at one height on a vertical face catch a yaw (unequal readings) but not a tilt, so also read `scene_raycast(method="bullet")` normals on two faces that are not parallel, such as a side and the top (a tilt about one face's normal leaves that normal unchanged), and check that each lies along its axis, allowing for the ground: the statics below read up to 0.31 degrees off the axis, the terrain about as much, the items on it. Read the fixture again a few seconds later. It may change an item's configured resting side; slopes and settling were not measured. The `normal` of a default `rvproxy` reply is not a face normal (added 2026-10-02, two runs) [EXACT][CLAIM-MCPV-RVPROXY-NORMAL]: the bridge copies the engine's `RaycastRVResult.dir` (`MCPBridge.c:2958`), for a ray the "direction and size of the intersection" (`scripts/3_game/global/dayzphysics.c:104`), as `enforce-script-reference` ("Surface normals") found on house walls. In this run and in run cf2d6bb3 it lay along the ray in all 190 `rvproxy` hits on an object, the box yawed about 10 degrees included ((-1.963, 0, 0) and (-1.966, 0, 0)), and its length is the ray's path through the object: (-2.1, 0, 0) across a 2 m box is its width plus the 0.05 m sphere at each side. `bullet` returns the hit face's normal (`MCPBridge.c:2998`): a unit vector in all 40 of its hits, up to 0.31 degrees off the axis on the `HouseNoDestruct` statics and on the axis on the items. *(Corrected 2026-10-02: this entry read the pose from the returned `normal` lying along the axis, citing (-2.1, 0, 0); with the default method it always does.)* An item can come out with its sides swapped, and none of these checks shows it (measured 2026-10-02 on DayZDiag 1.29.163709, dayz-mcp run cf2d6bb3) [EXACT][CLAIM-MCPV-ITEM-YAW]: one MLOD held two 2 m boxes, A (x -3..-1) in `Component01` and B (x 1..3) in `Component02`, and a twin had B in no component; each model was spawned, binarized and as MLOD, with `flags=8389668` and `rotation=64` as `HouseNoDestruct` and as `Inventory_Base` (`item_large`). On the four statics the west box was A, as modelled: it answered as component 0 (`Component01`; `Component02` reads 1), and on the twins it was the only box that took rays. On the four items the boxes had swapped sides, as a turn of 180 degrees about Y leaves them: the west box answered as component 1, and on the twins every hit was component 0 on the east box while the west box took no ray. The boxes are symmetric in z, so these rays cannot tell that half-turn from a mirror in x. Parallel rays read equal on the items too, and the normals pointed the same way as on the statics, since the swap keeps every face on the axes; a fixture symmetric about its origin, like the single box above, cannot show it, and `object_inspect` returns memory points in model space (`GetMemoryPointPos`, `MCPBridge.c:2240`), not the pose. So give the fixture sides a ray can tell apart and check which answers where: one component per side, read from the `component` of `rvproxy` hits (`geom`, `view` or `fire`; the bridge leaves it 0 on `bullet` hits, `MCPBridge.c:2983-3004`), or a collider on one side only. Put the parts off-centre in z as well to tell a half-turn from a mirror, and probe along z too: a quarter-turn moves parts that lie along x onto the z axis. `telemetry_read(mode="object_at")` returns the object's `GetOrientation()` (`MCPBridge.c:3137`); that run did not read it. Not measured: whether the swap is fixed or random (all four items swapped; 64 is also `RF_RANDOMROT`), other `rotation` and `flags` values, base classes and layers, and the cause. - **Standing on top: move the player before reading its height** (added 2026-10-02, measured on DayZDiag 1.29.163709 in a later run, on 2 m boxes) [EXACT][CLAIM-MCPV-STAND-PROBE]: a second reading of the object's physics body next to the walk-into probe above, on its top instead of a side face. An idle player is not a probe. After `player_teleport(pos=[x, <top + 0.07>, z], skip_clearance_check=true)` the player kept the teleport height (y 340.6000) and did not fall, on a box top and above open ground alike: the server read it twice over 3-4 s (`query_player_state`), and above open ground the client's physics position (the `start_pos` that `player_move` reports) still read it 15 s after the teleport (the worn-items section below saw the same after a teleport to y=160). Every `query_player_state` of these probes read godmode on, the DayZ_MCP default; the run did not separate the bridge, godmode and the engine as the cause. A short move applies the fall: the run teleported the player above the centre of each top, called `player_move(angle_deg=0, speed="walk", hold_s=0.5)`, then `query_player_state`. Above open ground the player dropped to the ground (y 340.6000 to 338.459); on the boxes, three `Inventory_Base` items at least 4 m from their neighbours (two outward-wound, MLOD and binarized, and one inward-wound), the server's y matched the top to within 0.0001 m (spawn y + 2.000 m, where `scene_raycast(method="bullet")` hit it) and its x, z lay 0.70-0.75 m from the centre, inside the 2 m top. Read the server's position after the move: with x, z still inside the top's footprint, y at the top's height there means the player stands on it and y at the ground means it fell through; an endpoint off the footprint, or any other height, is INCONCLUSIVE. Start above the centre of an isolated top (no other support at that height within the walk's reach, or the player can end on it), and run the same teleport and move above open ground as the falling reference. Keep the move short: 0.5 s took the player 0.70-0.75 m by the server's position read after the move, with the edge 1 m away; a top whose nearest edge is closer than that can be walked off in 0.5 s, and the endpoint check then reads INCONCLUSIVE. - **Placement**: `telemetry_read(mode="object_at", type=<classname>, pos=<spawn_pos>, radius=2)` → `found=true`, `pos` ~ spawn, reasonable `orientation`. PASS = not buried or floating (cross-reference `pos.y` with visuals). ### Item / arma - Load + visible/textured/winding as above, with **emphasis on proportions vs real-world reference** and on post-import geometry (orientation, winding) — where generated/imported meshes tend to fail. - `telemetry_read object_at` → `attachment_count`, `health01`. Telemetry PASS = found + healthy stats. - Collision: as for a static object, the ray battery with its two controls. Picking the item up, by hand or with `action_use`, does not test it (static-object playbook, "Picking an item up proves nothing about its collision"). ### Vehicle (placement/structure only) - Load + visible + collision + telemetry. `vehicle_enter(pos)` → `seated=true` confirms seat. `telemetry_read` → `engine_on_server`, `wheel_count`, `fuel_fraction`. - `vehicle_enter` SEATS player (placement/seat). To DRIVE: drivability ladder (§DRIVABILITY + `references/acceptance-ladder.md`) covers autonomous driving via owner-side verbs (`vehicle_get_in_client`/`engine_set`/`vehicle_control`/`vehicle_telemetry`). ## CAPTURAR ARMA ALZADA / ADS (raise client-side) [VERIFIED-SR2M gate iter36→37] To validate RAISED/aimed weapon pose (e.g. support hand grip) there is no player input tool. Weapon raise is forced IN THE MOD, **client-side** — captured character is client's LOCAL player, so server `init.c` override DOES NOT alter the pose client renders (exact symptom: server log states `raised=1` but capture shows weapon LOWERED). Drop-in: a `modded class MissionGameplay` in a gate mod (declares a `missionScriptModule` in its `class defs`, `files[]={"Mod/Scripts/5_Mission"}`; build `-PackOnly` so `.c` survives packing): ```c modded class MissionGameplay { override void OnUpdate(float t) { super.OnUpdate(t); PlayerBase p = PlayerBase.Cast(GetGame().GetPlayer()); if (!p) return; HumanInputController h = p.GetInputController(); if (!h) return; h.OverrideRaise(HumanInputControllerOverrideType.ENABLED, true); // ENABLED persiste; ONE_FRAME parpadea -> idle } }; ``` Key takeaways: - 3rd-person aim pose depends SOLELY on `IsRaised()` (`dayzplayerimplement.c:1742`, AimingModel) → sustained raise is enough. DO NOT call `SetIronsights()`: forces ironsight camera which fights MCP free-cam, and server-side left weapon untextured. - `WeaponADS()` (`human.c:86`) is INPUT flag without script override (`human.c:253-274`) → always 0 even if ADS works. Success signal = `IsRaised()` / client log, NOT `WeaponADS()`. - Keep `OverrideRaise(ENABLED)` server-side (init.c) as well so server agrees. - Judge grip by hand at NATIVE resolution (crop orbit frame), never rescaled contact-sheet. Grip mechanism + geometric parity: skill `dayz-animation-pipeline` (`references/weapon-in-hands.md`). ## TOOL → WHAT IT VERIFIES | Tool | Verifies | Failure signal | |---|---|---| | `world_spawn` | classname loads | `unknown_type` / `spawn_failed` → mod not mounted | | `camera_set` + `capture_screenshot` | render: visible, textures, winding, proportions | invisible / magenta / holes | | `scene_raycast` (default `rvproxy`) | LOD collision (Geometry/ViewGeo/FireGeo, by `intersect`) | no hit where it should hit | | `scene_raycast(method="bullet")` | a physics shape on the masked layers (static-object playbook) | no hit on the target where it should hit; a miss alone is not proof | | `player_move` walk + `query_player_state` | the player stopped by the object (static-object playbook) | `arrived:true` through it; a walk-through alone is not proof | | `telemetry_read` (object_at) | placement, orientation, attachments, health | `found=false` / buried pos | | `bridge_status` | peer liveness (gate) | `last_poll_age_s=null` → bridge down | | `world_time_set` / `world_weather_set` | reproducible scene for captures | — | ## WHAT IT DOES NOT COVER (ALWAYS declare in report) - **Player and UI actions**: open/close doors, interactive inventory, shooting, reloading, menus. There is no generic player input tool. (Driving IS covered — ladder §DRIVABILITY / `references/acceptance-ladder.md` with owner-side verbs.) For a building, geometry/collision YES; **doors NO** → manual test. - **`exec_enforce`** does not execute on the headless diag server (GATE4B-LIM, engine limitation like MakeScreenshot) — do not rely on it to "execute arbitrary verification logic". - **`telemetry_read`** is exposed as-is (BUG-010/011/012, hardening pending): does not certify large JSONL fixtures or extreme ranges. - **CONTINUOUS actions (with progress bar): `action_use` STARTS them but does NOT COMPLETE them** (measured 2026-09-07, LFPowerGrid, ledger entry `fb-20260907-184749-3fc1`). Returns `ok=1, started=1` and the server effect **never arrives**: `OnFinishProgressServer` does not fire and the target remains in the world (`entities_query` at 0.003 m after two attempts with 60 s and 45 s waits). `key_press` is **not** the way out: it documents itself "not OS input, key-up, hold", and a continuous action needs SUSTAINED input. This immediately kills **dismantling, deploying/deploy, and crafting**, which is precisely how a mod moves money and persistent objects. - **What CAN be affirmed, and it is not little**: the CONDITION is evaluated for real, so `started=1` means "valid candidate" and serves as a measure that the mod **offers** the action on that target in that state. Always verify it with a negative control that must come out red (an excluded target returns `condition_failed`), or `started=1` proves nothing. - **Corollary that bites separately**: since deploying from a kit is a continuous action, **there is no way via MCP to create a truly persistent object** — `world_spawn` uses default flags and does not survive restart. Add that `dayz_test_stop` does not shut down gracefully (the next startup prints `... was not closed. Always shut down the server gracefully`) and **no test of `OnStoreSave`/`OnStoreLoad` between restarts is conclusive via MCP today**. If the result comes out "did not persist", that is the harness, NOT the mod: do not report it as a bug. - **`class Doors` (`Building`) doors do NOT open by MCP without `door_index`** (measured 2026-09-07, LFSecure I-0/I-4, PBO L4): `action_use(ActionOpenDoors, classname=<door>)` returns `condition_failed` even with the player 1.4 m away, because the bridge builds the `ActionTarget` with `componentIndex=-1` and `Building.GetDoorIndex(-1)` (native, `3_game/entities/building.c:28`, "index of the door based on the view geometry component index") resolves no door. `object_anim(source=<door source>, phase=1.0)` answers `phase=1`, but the engine's door controller returns it to its "wanted" phase (re-read `phase=0`) and `IsDoorOpen(index)` stays false, so every action that needs an open door also returns `condition_failed`. Without `door_index`, opening/closing, sound and sync were a MANUAL test (the user's F) and MCP could only verify "starts closed" (`object_anim` reads `phase=0`); see the update below. - Update (measured 2026-10-01, DayZDiag 1.29.163709, test building whose door component is a button on a moving piece) [EXACT]: `action_use` with `ActionOpenDoors`/`ActionCloseDoors` and `door_index` DOES open and close the door even there. `object_anim` (`SetAnimationPhaseNow`) on the door's source still cannot hold a static pose: the door settles at `ajar` with phase 0.07 instead of the requested phase — doors cannot be posed statically. For the verdict, use `object_doors` and rays. - **`setup_failed` is a FALSE NEGATIVE for local instantaneous actions** (`IsLocal() && IsInstant()`: `ActionTogglePlaceObject`, `ActionDropItemSimple`; measured 2026-09-07): the bridge checks `GetRunningAction()==null` right after `PerformActionStart`, and an instantaneous action has already finished. The hologram appeared and the object fell to the ground in both cases. Rule: with `setup_failed` on an instantaneous action, verify the EFFECT (capture, `entities_query` of dropped object) before marking the action as failed. - **`scene_raycast`: `entry=0` is TERRAIN and `normal` is NOT a unit normal** (measured 2026-09-07 with `intersect=fire` and `geom`): a vertical ray against the ground returns `hit=1`, `object_type=""`, `surface_type=cp_grass|cp_concrete2`, `entry=0`, `exit=0`; against an object it returns `entry=1`. The `normal` field is `RaycastRVResult.dir` as-is, which in line-object collision is "direction AND SIZE of the intersection" (`3_game/global/dayzphysics.c:104`): measured magnitudes 0.16-0.42. Its direction is the ray's own ("Axis-aligned test fixtures" above), so it says nothing about the face: never read it as a normal nor use it as a threshold (`dot >= 0.9` is never reached); a face normal comes from `method="bullet"`. *(Corrected 2026-10-02: this line said to use it only as direction (sign/axis).)* A mod that filters `!hit.entry` discards natural ground (hologram that never snaps to ground: LFSecure I-1). *(Qualified 2026-10-03: a ray that starts inside an object also reads `entry=0`: the inside rays of run 606a5dbb read `entry 0, exit 1` at distance 0 on the controls. Tell terrain by its empty `object_type`, not by `entry=0` alone; `entry=1` holds for hits from outside.)* - **Collision probe with vanilla CONTROL before blaming the mesh** (measured 2026-09-07, LFSecure L3 -> L4): if `scene_raycast` does not hit a mod static, repeat the SAME ray (view, fire, and geom; from outside AND from inside) against a vanilla static spawned next to it (`world_spawn(type="Land_Container_1Aoh")`: flat walls, `House`, `IsBuilding()` true, also useful as a wall for holograms). Control HIT + mod MISS in all modes and from both sides = reversed collision convex components (inverted winding in Geometry/View/Fire: ODOL preserves them, engine does not see them); control HIT + mod HIT only from inside = reversed faces only in visuals. With the source winding restored (L4) the same ray gave the front face at +0.079 m and a wall of 0.42 m. The correction rule lives in `dayz-p3d-audit` (0% agreement = discrepancy, not direction). *(Qualified 2026-10-03: the two readings above name the likely cause, not a measured one; only a pair that changes the winding alone decides it, as "Ray battery and its controls" in the static-object playbook says.)* ## REPORTING A report with: criteria table (criterion · PASS/FAIL/INFO · evidence), PNGs and telemetry/raycast JSONs as evidence, global verdict, and a **"to manual test"** section with what is not covered (doors, shooting, etc.). Context cost: each capture weighs ~25k tokens (~240-320 px) — **batch** orbit-pose captures and do NOT re-read a PNG except to verify something concrete (images inflate context quickly). Hard budget ~25k tokens/image: do not request higher resolution. ## TROUBLESHOOTING | Symptom | Cause | Fix | |---|---|---| | `claude mcp list` → `× Failed to connect` | `claude mcp list` is NOT reliable: gives false `Failed` with the server operational. E4 contention on port 8765 (`ExclusiveThreadingHTTPServer` lock, fail-closed) was the PRE-broker failure mode | verify FIRST with ToolSearch (`mcp__dayz-mcp__*` tools available? → all OK, ignore `Failed`). Broker registration `--client` (since 2026-06-24) overcomes E4 contention: N sessions share the server without fighting for 8765, and orphan-guard releases orphans alone. If tools are truly missing: register in broker mode (`--client`) and open NEW session (tools load at startup) | | tools do not appear in session | server was registered AFTER opening session | open a new session (MCPs load at startup) | | `bridge_status.server_peer.last_poll_age_s = null` | server bridge does not poll | check `dayz_mcp.json` in server_profiles + key; confirm `@DayZ_MCP` mounted (absolute paths `!Workshop`) | | `client_peer … null` (server ok) | client does not connect or `client_profiles\dayz_mcp.json` missing | BUG-009 (flaky autoconnection): larger `-ServerWait` / retry; seed client config | | `version_state = legacy_blocked` | `--require-version` ON against a bridge not sending `ver=` | deploy 4B PBO (sends `ver=4~…`), or register server without `--require-version` for that run | | byte-identical captures between poses | grab caught a stale desktop frame (not render) | [EXACT] `PrintWindow(hwnd, hdc, PW_RENDERFULLCONTENT=2)` on the client's main window DOES capture the D3D game area — full scene in >20 captures even with another window covering it and without taking focus, as long as the Windows session is unlocked and the display is on (measured in game, DayZDiag 1.29, 2026-09-28). A black game area means the display is off or the session is locked, not that the route cannot read D3D. `CopyFromScreen` only works with the window in front, and a covered DayZDiag also drops to ~20 FPS; validate on a crop of the game area, not whole-frame variance (corrects the LL-247 entry of 2026-08-12) | | tool timeout and then "zombie" commands on reconnect | BUG-024: timeout leaves command queued; bridge executes it on return | after a timeout, reconcile with `bridge_status` before continuing | ## REFERENCES - `dayz-test-ingame` — build/deploy/launch (this skill composes it with `-ExtraMods "@DayZ_MCP"`). - `DayZ_MCP_dev\tools\README-mcp.md` — startup order + bridge troubleshooting. - `DayZ_MCP_dev\HANDOFF.md` (LIVE-STATE) — server invariants, 11 tools, GATE4B-LIM, backlog. - `_shared\dayz-conventions.md` — L2 (LODs, ViewGeo/FireGeo, DayZ response format). - Precedent of camera+orbit capture: A6's `gate-mcp.ps1` (window-grab by orbit, dayz-test-ingame finding 2026-06-17) — alternative batch sweep to agent driving tools. ## VEHICLE: repeatable smoke recipe (added 2026-06-24) Visual smoke of a CarScript vehicle driven by MCP, verified end-to-end on MercedesAMGLF 2026-06-24. Repeatable without re-deriving: 1. **Mission = stock `dayzOffline.chernarusplus`** (from DayZServer install), NOT a Phase 0 mount-probe `void main()`: mount-probe spawns no player → client does not render the world and capture comes out black. Stock spawns player (`CreateCharacter`→`CreatePlayer`) → free-cam with rendered world. Pass it with `-Mission "<...>\dayzOffline.chernarusplus"`. 2. **Bridge seed in profiles generated by `dayz-test.ps1`**: `dayz_mcp.json` `{"url":"http://127.0.0.1:8765/","key":"<key>","pollHz":5}` (ASCII) in `<Mod>_dev\_server\profiles\` and `..\_client\profiles\`. The bridge reads it from `$profile:` (server `MCPBridge.c:125`, client `MCPClientBridge.c:211`); key = `DayZ_MCP_dev\tools\.dayz_mcp.key`. Seed BEFORE launch (the bridge reads config in its init, only once). Do not use `install-mcp.ps1 -Register` (re-registers broker mode `--client`). 3. **Launch**: `dayz-test.ps1 -Mod <Mod> -Mode all -Build -PackOnly -ExtraMods "@DayZ_MCP" -Mission "<...>\dayzOffline.chernarusplus" -ServerWait 240`. `-PackOnly` mandatory in mods with `.c` (binarize drops them → NO_IGNITER). Call the `.ps1` via absolute path (P:\ is subst). 4. **Readiness**: stock runs CE (`InitOffline`) ~1-3 min; during CE `bridge_status` comes out STALE (`last_poll_age_s` grows even though `version_state=ok`) — do NOT spawn there. Wait for `connected to server` in server's `script*.log` and for `bridge_status` to return fresh (<1 s). Poll log host-direct with PowerShell, NOT bash Monitor (inherits bindfs cache, LL-142). 5. **Smoke** (grouped, R5): `world_time_set(year,month,day,hour=12,minute=0)` (all 5 mandatory) + `world_weather_set overcast=1.0` → `world_spawn <Class> pos≈[player+~10]` → wait 30 s (survives without native crash = LL-099 ruled out) → `camera_set` (cam_mode **"lookat"**, `cam_pos`+`look_at`) + `capture_screenshot` + `scene_raycast` + `telemetry_read object_at`. Healthy telemetry = `found=1`, `health01=1`, `velocity 0`. 6. **Flaky capture**: grab sometimes catches loading screen or client's "Continue" menu overlay instead of render. If it happens: wait for settle (~30-40 s in background — foreground sleep is blocked in this environment) and recapture; use several angles. Top-down that clearly reveals proxy (mis)alignment: `cam_pos=[carX+3, 16, carZ+1]` `look_at=[carX, ground+0.1, carZ]`. ~25k tokens/image — batch captures, do not re-read a PNG except to verify something concrete. A frame flat everywhere, vanilla terrain and sky included, is not a flaky grab of the model: withhold its texture verdict, check that the client window has focus and capture again after settling, since textured runs have opened with 4-11 s of flat frames; if the scene stays flat, its visual evidence is `SETUP_FAIL` and the run is relaunched (`dayz-test-ingame` `references/session-findings.md`, "H2R import (2026-10)"). 7. **User's eye > capture** when render is ambiguous: in MercedesAMGLF grab was flaky and user reading was reliable diagnosis of winding and proxy alignment. If there is a human in the loop, cross-check with them. (Case s3 2026-06-24: I was about to sign PASS on faint renders; user's eye caught proxy misalignment I could not resolve → correct FAIL.) 8. **SHARED SESSION.** Port 2302 and Steam client remain single resources, but exclusion is coordinated with the runbook's FIFO lease, not attributing processes nor evicting other sessions. `dayz_test_run` acquires and maintains exclusion; retain `run_id` and terminate that same run with `dayz_test_stop`. NEVER kill a process to unlock the box: if state does not reconcile, retain the process and declare degraded shutdown. 9. **"Wheels = sheet/flat disk" is NOT a proxy bug.** A `world_spawn` of a CarScript without attachments leaves `wheel_count=0`/`attachment_count=0` → only hub/disk is visible, without tires. Expected in visual smoke; real wheels require attachments (physics phase), out of scope for smoke. 10. **`legacy_blocked` ("poll did not include ver=") is usually INCOMPLETE INIT, not version mismatch (added 2026-06-28).** Refines the homonymous row in TROUBLESHOOTING: with 4B PBO deployed, right after launching server `bridge_status` can come out `legacy_blocked`/`last_poll_age_s=null` because the bridge has not yet completed the handshake; transitions to `ok` with `ver=4~…` when server mission LOADS completely (~1-2 min). Do NOT redeploy nor re-register for that — wait and re-check. It is only a real mismatch if it stays `legacy_blocked` with mission already loaded (client connected, world rendered). Origin: SUB_BRZ Phase 5 2026-06-28. 11. **Historical pattern superseded 2026-07-15.** `cmd start`/`.bat` avoided a background job losing its child, but created a process outside the registered lifecycle. For any current smoke, use exclusively `dayz_test_run`, retain its `run_id` and terminate only that run with `dayz_test_stop`. No unmanaged fallback exists to bypass the lifecycle guard. ## DRIVABILITY + ACCEPTANCE LADDER (summary — detail in reference) Phase 5 of DayZ-MCP project added and gated in-game owner-side verbs that DO **drive** the car (client takes ownership and handles throttle/steer): `vehicle_get_in_client(pos)` (seats + ownership), `engine_set("start"/"stop")`, `vehicle_control(throttle, steer, brake, handbrake, hold_ttl_s)` (SUSTAINED control, fail-closed), `vehicle_telemetry()`, `vehicle_release()`, and `query_get_in_condition(pos, component)` (server peer, diagnoses which of the 7 gates of `ActionGetInTransport` blocks). On top of them runs the **acceptance ladder** rip→drivable: ordered rungs **R1 spawns → R2 render → R3 get-in available → R4 seated → R5 drives → R6 wheel direction**, each reading in-game ground-truth and mapping its failure to a known fix in the SUB_BRZ taxonomy (`dayz-vehicles/references/`). Reference orchestrator `references/drive_ladder.py` (drives R1→R6, outputs `verdict.json`; does NOT apply fixes nor rebuild) + fixtures `tests/test_drive_ladder.py`. **Full detail** (verbs, verified owner-side mechanism, preconditions, spawn placement, R2.5 restore-gameplay, anti-false-green guardrails, ladder table, and failure→fix mapping) → `references/acceptance-ladder.md`. ## (added 2026-07-14) No restore-gameplay tool: for user MANUAL test, relaunch without @DayZ_MCP `camera_set` ALWAYS suppresses player control (`SuppressGameplay()` -> `PlayerControlDisable`, see §R2.5) and **there is NO MCP tool that fires `RestoreGameplay()`** — only internal gate `drive_probe_client` does it, and it is not exposed as a verb/tool. Practical consequence: after an MCP smoke that used `camera_set` (free-cam or static-cam lookat), the USER cannot take player control to test the car by hand — their input remains suppressed and there is no verb to revert it on the fly. Rule: when workflow passes from MCP smoke (automatic) to **user manual test** (driving, judging feel/aesthetics), **relaunch the game WITHOUT `@DayZ_MCP`** (or at least with a client that never received `camera_set` in that session). A clean client has normal control. If the client was already neutered: reconnecting (Esc -> Disconnect -> Reconnect) recreates the player with clean camera/control; there is no tool shortcut. Origin: SUB_BRZ s35 — time was lost looking for a non-existent restore tool and user had to close the game and ask to relaunch normally. ## (added 2026-07-14) Autonomous smoke: world_spawn takes engine [x, y_up, z], and capture with sleeping display = black frame (SP-060) Dos caveats de smoke MCP autonomo verificados in-vivo (SUB_BRZ s32): 1. **`world_spawn` takes vector in ENGINE order `[x, y_up, z_north]`** (`MCPBridge.c:1638` `Vector(x,y,z)`), but bridge connect-log prints player position as `<x, z_north, y_up>`. Passing the log triplet VERBATIM spawns the object at ~6 km altitude -> `IsSpawnReady` (2.0 m radius) is never satisfied -> job timeout + `found=0`, and engine auto-deletes the orphan (`NETWORK (E): Will delete object ... outside world coords`). Cost 3 timeouts in a row. Rule: convert `<x,z,y>` -> `[x,y,z]` before every `world_spawn`/`camera_set`; after a spawn timeout, grep RPT for `outside world coords` BEFORE retrying (distinguishes bad-coords from slow-spawn). Verifiable with `scene_raycast` to terrain (gives real y_up). 2. **Capturing with the display asleep = BLACK frame** even though the client runs (on 2026-07-11 a locked session with LockApp also gave black frames; since 2026-09-14 the capture backend fails outright on a locked session instead, see point 4). Fix: bring the client window to the foreground + input wiggle (mouse / F15) before EVERY `capture_screenshot`. Misleading symptom: it looks like "the mod's render is broken" and it is the compositor. 3. (minor) Spawn adjacent to player can fall INSIDE a building -> probe 3-4 `scene_raycast` to clear terrain before choosing pos. 4. [EXACT] A locked Windows session is a distinct capture failure (added 2026-09-14, SUB_BRZ s93): with `LogonUI` alive, `capture_screenshot` does not return a black frame — it fails outright with `capture_backend_failed` on every attempt while the game client stays alive and probing. Before budgeting captures in an unattended run, run `Get-Process LogonUI` first: with a locked session there is no host capture at all, for the MCP capture and for any window-grab script. (measured in game, DayZ 1.30.164014 Exp) Origin: SUB_BRZ s32 MCP smoke (2026-07-11): 3 `world_spawn` timeouts with raw log pos + 2 black captures with LockApp; both resolved with the above. ## (added 2026-07-14) Telemetry numeric gates: calibrate against MEASURED PHYSICAL FLOOR, not ideal fixtures (SP-061) Invariant for any acceptance-ladder / numeric gate on in-game telemetry (car drive_ladder, spikes, future harnesses): 1. **A threshold calibrated with ideal mathematical fixtures is a gate that no real physics passes.** The engine has a native noise floor (body dither between SetVelocity and reading; client network interpolator). Before setting a threshold: MEASURE the floor in a real run (metric p95 on real data), threshold = floor x margin (>=2x), and document the measured numbers next to the constant. Two
Comments (0)
Sign in to join the conversation.
Reviews (0)
No reviews yet.
No comments yet.