`tools/screen_designs.py judge` sends each capture together with its concept
mockup to a vision model and asks whether the capture realises the same screen and
layout intent. This covers what pixel/perceptual metrics cannot judge.
- Prompt explicitly ignores art style, fidelity, fonts, and placeholder text, since
the references are stylised mockups.
- Verdicts are classed match / review (mismatch below --fail-confidence, default
0.6) / diff (confident mismatch, exits non-zero) / unknown; --json, --only,
--dry-run, --model supported.
- Reads the repo .env for OPENROUTER_API_KEY without overriding real env vars.
- Default judge model google/gemini-2.5-flash-lite; openai/gpt-4o-mini verified too.
Tests/docs
- 7 new offline tests in tools/test_screen_designs.py (request payload embeds both
images, JSON/fenced/parts/non-JSON parsing, threshold classification, .env
loading); 30 Python tests pass.
- Live run: 3 captures match, 3 review, 0 confident diffs.
- README, TODO section moved to DONE, MEMORY. Version 0.2.15.
Host (applied with root@192.168.88.30)
- [service] REQUIRE_SIGNIN_VIEW = false in /etc/gitea/app.ini (backup written
alongside), Gitea restarted. Verified anonymously: repo page 200, API 200, and a
release asset downloads at full size (50.8 MB AppImage).
Localization safety (Open Polish)
- Strings.placeholder_count() counts conversions and ignores %% escapes.
- Strings.fmt() validates the count and fails safe (warns, returns the template)
instead of raising an engine error when a key's placeholders drift.
- Strings.validate() lints the table for a trailing %, unsupported conversions, and
mixed positional/sequential placeholders.
- Tests cover the counter, the fail-safe path, a clean table, and a formatting pass
over every placeholder key with type-correct arguments.
Docs
- TODO: distribution, release-pipeline, and open-polish sections are complete and
moved to DONE; the v0.2.4-v0.2.8 backfill is closed as superseded.
- README/MEMORY updated. Version 0.2.14.
Joke Shop
- Reservation list without ID (guest list valid, ID missing) — denied.
- Borrowed/mismatched ID (ID valid but mismatched) — denied.
- Guest who refuses the wand screen: new `refuses_screening` event flag, copied
onto NPCController, handled in _toggle_wand_screening (logs wand.refused, records
a wand_refused event, small prop hit, panel stays closed).
- Late arrival during the performance (ID + ticket, "arrive" intent).
Studio
- Filming-boundary event and a later repeat, both using the new `film_boundary`
intent which maps to `steps_end` so the influencer stands past the posted line.
- Celebrity arrival now pulls in a small crowd via a new `extra_spawns` event field
(two tourists + an influencer), spawned through the normal path.
Building
- Filming-boundary event whose hint points at the body camera and request-reason
verbs, plus an unclear-intent event that hints at asking for a reason.
Engine support
- Events may carry `hint` (string key logged on spawn), `extra_spawns`, and
`refuses_screening`; 9 new localized hint strings.
Tests/docs
- Headless suite 139 -> 149 (test_scenario_content.gd) and a new
smoke_scenario_content.gd that spawns the refusal and celebrity events through the
real signal path.
- README, TODO section moved to DONE, MEMORY. Version 0.2.13.
- Host root filesystem was expanded from 7.8G to 16G (~7.8G free), so the disk
blocker is gone.
- Re-published v0.2.0 (4 artifacts) and v0.2.9 (5 artifacts) from the local build/
outputs; both releases verify.
- Root SSH from the workstation is not available (publickey denied, password auth
closes the banner); the host is administered via the zwitschi login.
- Version 0.2.12 exists to verify the tag pipeline end to end: the workflow should
create the release via the checkout-credential fallback and upload all five
artifacts.
The v0.2.10 tag build still failed to publish: its env dump showed no token at all,
because this Gitea/act runner does not provide `github.token`, and the repo has no
GITEA_TOKEN secret.
publish_release.sh now resolves credentials in order: GITEA_TOKEN,
GITEA_AUTH_HEADER, ~/.git-credentials, then the local git extraheader that
actions/checkout writes for the job. That header carries the runner's basic-auth
credentials for the same server, so tag builds publish with no repo secret.
Verified by adding the extraheader locally and publishing v0.2.10 (all five
artifacts: Linux/Windows/macOS ports, NSIS installer, AppImage), then removing it.
Docs: RELEASING describes the four credential sources and warns that github.token is
not available here; MEMORY records both failed attempts and the final fix; TODO keeps
the "verify on the next tag" follow-up. Version 0.2.11.
Tag builds since v0.2.2 built all artifacts but never published a release:
- tools/publish_release.sh only assembled AUTH in the GITEA_TOKEN branch; the
~/.git-credentials branch set CRED and then used an unbound ${AUTH[@]} under
`set -u`, so every API call failed and the job ended "could not create or find
release". It now sets AUTH=(-u "$CRED").
- The repo has no GITEA_TOKEN secret and the runner has no git credentials, so the
workflow now falls back to the built-in workflow token:
GITEA_TOKEN: ${{ secrets.GITEA_TOKEN || github.token }}.
v0.2.9 was then built and published locally with all five artifacts (portable
Linux/Windows/macOS, NSIS installer, AppImage) and the release page verified via
the API. v0.2.4-v0.2.8 remain unpublished; they are superseded by the latest.
Docs
- docs/RELEASING.md records the incident and the local-publish fallback.
- TODO gains a release-pipeline follow-up (verify CI on the next tag).
- MEMORY notes the root cause, the missing secret, and which releases exist.
- Version 0.2.10.
- Escalation cue now includes a shape meter and numeric level
("Escalation: Blocking [###--] L3/5") in addition to the colour and name, so it
reads in grayscale. Hud.escalation_meter(level) is pure and tested.
- HUD font scale from settings actually works: set_font_scale() resizes fonts via
theme overrides instead of Control.scale, which pushed the right-anchored clock
label off-screen. It now covers the top, log, target, hint, time, mode,
escalation, cooldown, and debug labels.
- Settings load before the HUD exists, so _ready() re-applies player settings after
_build_ui(); previously a saved font scale was ignored until the next apply.
Tests/docs
- Headless suite 134 -> 139 (test_accessibility.gd).
- Verified visually at 1.4x scale: no overflow, meter readable.
- README, TODO section moved to DONE, MEMORY. Version 0.2.9.
- AudioFeedback.start_crowd_bed() builds a 4 s looping AudioStreamWAV at runtime
(one-pole filtered noise plus a 68/103 Hz drone, seeded) and set_crowd_level()
maps crowd mood 0..1 to -34..-12 dB, so the street gets louder as it gets
agitated or panics. The bed starts with the shift and stops at shift end.
- play_radio_open()/play_radio_close() push short filtered-noise bursts; wired to
opening the radio panel, submitting a report, cancelling, and closing the panel.
- Still asset-free and headless-safe: with a dummy audio server the calls no-op.
_free_player() returns null before _ready and callers handle it.
Tests/docs
- Headless suite 127 -> 134 (test_audio_feedback.gd).
- New smoke_audio.gd: bed playing, level follows mood events, radio crackles once
per open/close, bed stops at shift end. Verified windowed (real audio server).
- README, TODO section moved to DONE, MEMORY. Version 0.2.8.
Once an NPC reaches its intent waypoint it now does something instead of standing
in an idle T-pose:
- sit: the rig settles 0.34 m into a seat and the legs swing forward on the steps
- photo/film: influencers and tourists raise their camera prop briefly (a sine
pulse, so it reads as taking a shot rather than a permanently raised arm)
- perform: performers sway the sign board they carry
Pose selection lives in NpcAvatar.theater_pose() (intent + arrived); the math is in
pure helpers (crouch_for_pose, prop_lift_for_pose, sway_for_pose) so it is
headless-testable. Prop/leg nodes are captured at build time and animated from
their rest offsets. Performers also get a hand-held board and post.
Tests/docs
- Headless suite 120 -> 127 (test_npc_theater.gd).
- Verified visually with a staged sit/photo/perform lineup in front of the entrance.
- README, TODO section moved to DONE, MEMORY. Version 0.2.7.
- New src/game/ui/event_timeline.gd flattens every recorded incident event
(ADR-0005) into one time-ordered list and shows time, person, and a one-line
summary. Step with Left/Right or the buttons; Esc returns to the review. No live
3D replay in this first version.
- Shift review gains a "Body camera timeline" button that hands the review to the
timeline; the review dim is hidden while it is open (stacked overlays made the
screen nearly black) and restored on close.
- ShiftCoordinator records started_at_ms so the review can stamp events as mm:ss.
- 8 new string keys for the timeline UI and the review button.
Tests/docs
- Headless suite 113 -> 120 (test_event_timeline.gd), new smoke_timeline.gd that
opens the timeline from the review, steps through every event, and closes back.
- README, TODO section moved to DONE, MEMORY notes. Version 0.2.6.
- Responder outcomes now reach the shift review
(ShiftCoordinator.responder_outcomes), and the review screen names them.
- "evacuation" (emergency call, low danger, no weapon) actually clears the entrance:
the door is locked, every NPC flees and is marked as leaving, and a
security_failure reputation hit applies.
- An emergency call the situation did not justify (no weapon, danger <= 1) appends
an investigation note, records investigation_opened, and applies the existing
false_report hit. The review lists the investigations.
- Fix dispatch dying after the first shift: coordinator.setup() builds a new
RadioSystem, so the responder_arrived connection made in _ready was lost on every
venue switch/restart. _connect_radio() re-wires it beside _connect_crowd().
Tests/docs
- Headless suite 105 -> 113 (test_emergency_aftermath.gd), new smoke_emergency.gd
covering the investigation, evacuation aftermath, and review copy.
- README controls/status/smoke checklist, TODO section moved to DONE, MEMORY notes.
- Version 0.2.5.
NPC navigation (previously dead in the live game)
- Only assign NavigationAgent3D.target_position when it changes: re-assigning every
frame invalidated the path, so get_next_path_position() kept returning the current
point and avatars never moved.
- Follow the first path point that is closer to the destination than the avatars are.
Nav points sit on the navmesh plane (y=0) while bodies stand at y~0.5, so the old
3D distance test always looked "already there"; path[0] is also the old start point,
so a naive "far enough away" pick steered backwards and oscillated.
- Fix fsm.state -> fsm.current_state in the idle-lifetime check.
Crowd presentation
- Aggressive behavior, visible weapons, and responder arrivals make idle bystanders
stop and face the incident for 12 s; someone waiting at the entrance keeps focus.
- Panic (mood >= 0.75) walks people away from the door at 2.4 m/s, then they despawn.
- Wire reactions to EventBus.crowd_changed: CrowdSystem.apply_event() emits on the bus,
not on its own `changed` signal (the first wiring received nothing).
- Measured a 16-NPC entrance at ~1.2 ms/frame (824 fps, vsync off), so no LOD/culling.
World and HUD
- Skyline is now a ring of four alpha-scissored planes at 200 m instead of a single
plane; PlaneMesh defaults to horizontal, which is why the backdrop rendered as a band
overhead. Larger ground plane, mipmapped texture, scissor threshold 0.5 (0.3 left
mip dashes in the sky), lifted albedo so the city is not pure black.
- Flashlight aims straight ahead and lights the person/door instead of the ground.
- Shorter HUD controls line so it stops overlapping itself.
Tests, tooling, docs
- Headless suite 98 -> 105 (new test_crowd_presentation.gd, 7 tests).
- New smoke_crowd.gd: rubbernecking, panic, and flee-target sanity.
- capture_screenshots.gd starts a shift from the venue menu and stages the night set
properly; all six README screenshots regenerated.
- README, TODO/DONE (crowd section moved), MEMORY updated. Version 0.2.4.
Verbs (PLAN levels 1-3)
- Eye contact: own Level 1 radial slice, logged as "eye contact".
- Recognition is a player decision: "Recognize" sets NPCController.player_recognized,
writes access evidence, and GameManager only then appends a RECOGNITION artifact.
It is no longer implied by NPC type.
- Body camera: real flag (off by default) driving the REC overlay; B toggles it and
a formal warning arms it automatically, writing a body_camera incident event.
- Blocking stations the guard 0.9 m from the person toward the entrance and turns
to face them, instead of only emitting the escalation level.
- Request reason: Level 3 ask answered from the NPC's seeded dialogue_lines.
- Radial grows to 12 slices (174 px outer radius, centered labels).
Incident recording fix
- IncidentRecorder.start_incident() was never called, so add_event /
add_access_evidence / add_escalation silently no-opped and the shift review
always reported zero incidents. Incidents now open on focus (guarded by
has_open_incident), verbs call _ensure_incident, and ShiftCoordinator closes
leftovers as "unresolved" before building the review.
Copy and docs
- 9 new string keys (eye contact, body camera, recognition, reason, station).
- README controls, 26-step smoke checklist, test count, status lists.
- TODO section moved to DONE; MEMORY records the verb design and the recorder bug.
Verification: headless 98/98, smoke_plan_verbs.gd + smoke_session_shell.gd +
windowed smoke_focus_radial.gd all pass. Version bumped to 0.2.3.
- Venue-correct access artifacts (AccessLoadout) and screened-aware verify
- Cone focus on NPC heads with billboarded ring + display name (ADR-0008)
- Action wheel with de-escalation-first slices (ADR-0009); Esc/RMB cancel
- Procedural NPC face/hair/neck/shoes (ADR-0010); named input actions
- Fix GameManager parse error (radial typed as ActionRadial)
- Fix first-person camera height and guard spawning inside the building
- Move building volume back so the door/awning/marquee are visible
- Pull the event scheduler once per simulated second; cap live NPCs
- Regenerate README screenshots; update TODO/DONE/MEMORY/README; 77 tests
Adds src/test/smoke_focus_radial.gd for the windowed focus/radial smoke.
- Add a 'Verify Windows installer contents' step (7z l shows Step Line.exe +
uninstall.exe) so the TODO verification items are proven by CI output
- Rebuild 0.2.1 binaries (embedded PCK) and AppImage locally; smoke clean;
57/57 tests
- Phase-based lighting: sun direction/energy/color, ProceduralSkyMaterial
gradient, and ambient per shift phase (GameManager.PHASE_LIGHTING),
applied on phase transitions and shift start
- WorldBuilder: street lamp poles with emissive bulbs + OmniLights,
set_night_mode(), glowing STEP LINE marquee (Label3D + emissive bars),
crosswalk stripes, planters, low-poly street trees
- NPC avatars: per-seed height/brightness variation (deterministic LCG)
- Denser routine traffic probabilities
- capture_screenshots.gd: fast-forward to LATE phase and capture night set;
README gallery is now 2x3 day+night; docs/screenshots regenerated
- Version bumped to 0.2.0; build/ artifacts rebuilt; Linux binary smoke-tested
- New headless tests for phased lighting config and street dressing (56 total)
- Install Godot 4.3 export templates locally and fix macOS export
(enable rendering/textures/vram_compression/import_etc2_astc)
- Build Linux/Windows/macOS artifacts into build/ via tools/build_releases.sh;
Linux binary passes headless smoke test
- Add tools/publish_release.sh (creates Gitea release + uploads assets with a token)
- TODO: mark build/install steps done; publishing v0.1.0 + link verification
remain as the single action blocker (needs GITEA_TOKEN)
zwitschi