freefrank/lostodysseyrecomp/lo-render-flicker

Script-first render investigation

Work from the game repository root. Follow project rules; use tools/README.md to locate tools and read only the relevant subdirectory documentation when needed. Private inputs/reports belong in ignored out/; reusable code belongs in tools/.

Default: one agent, existing scripts

• The current agent owns the task. Fixer/reviewer agents are optional roles, not mandatory stages. Do not launch them unless the user requests delegation. Routine scans, sorting, identity checks, builds, tests, and Git operations use scripts/CLI.
• Select the route below; do not execute every tool for every report. Existing coverage CLIs print a compact summary and save the full JSON. Read that summary first, then use Python to select at most five relevant rows for the initial LLM context. Never print entire coverage reports, HLSL directories, or feedback archives into context.
• Reuse evidence when input identity, program bytes, PS pairings, relevant pass/binding state, tool semantics, and tested code are unchanged. Compare the archive checkpoint before scanning again. Reconcile saved mapping classifications with the current map; a new map does not require rereading all raw observations.
• Consult tools/shader_analysis/reviews/feedback_mapping_20260925.json before reopening known candidates. Reuse unchanged implemented/held decisions. New source, PS pairing, binding evidence or a contradictory capture reopens only affected cases. Translate only previously untranslated program bytes.
• Do not broaden a single capture into a whole-archive audit unless requested. Batch repetitive operations in scripts; parallel shell jobs do not require parallel LLM agents. Keep reports brief: change, evidence, remaining gap.

Choose one route

Commands and arguments are in tools/capture_analysis/README.md, tools/feedback_archive/README.md, and tools/shader_analysis/README.md. Read --help only for an unfamiliar command. Missing evidence goes into a precise held reason; repeated speculation is not progress.

LLM decision gate

Inspect only new/contradictory source slices or unresolved visual/pass relationships. Verify actual run revision, consumer, frames and drops; configured AA or zero vendor evaluations do not prove raster jitter was off. Locate the earliest bad image stage when investigating a visible symptom.

If a depth preview looks unchanged, compare selected raw float depth pixels before concluding that depth was stable. Check the active depth compare and polygon offset direction before attributing a missing color layer to depth rejection. F1 readbacks can stall long enough to reset temporal history; use capture traces for diagnosis and a separate same-scene comparison for visible behavior.

Before mapping, prove the exact four VP constants and their position/clip dependencies, preserved UV/light/skin data, and independent camera/depth/pass agreement. Review all observed PS pairings, including weak pairs. Match geometry, world transform, unmodified camera words, viewport/VTE and ordered depth use; a register address alone is not allocation identity. guards=31 needs the unknown-slot/flags/depth context and is not sufficient by itself. Hold mixed-camera uses and clip-XY/W screen sampling until producer, extent and temporal alignment are proven. Dynamic indices/control flow and unsupported audit results remain unresolved; no script candidate flag authorizes mapping.

Fix and finish

Edit only proven position-slot cases; preserve runtime guards. A VS used with mixed cameras may need a narrowly reviewed DrawPositionVPSlot VS/PS pairing and an independent earlier scene anchor instead of a VS-wide rule; retain the motion-vector fallback unless it is separately proven. Reuse an equivalent reviewed projection evaluator or add the smallest independent arithmetic regression, with old-behavior/rejection controls and unchanged Z/W and unrelated constants. Label synthetic inputs. Run only the affected build/test once; reuse success for docs, commit and push. CPU evidence is not GPU/visual acceptance.

Use the same agent for routine scoped documentation updates in this economical workflow. Do not create a second review or documentation round for bookkeeping. Retain exact commands/results and held reasons in one case report. Keep the changelog/current status accurate. When game testing is authorized, compare the same save, scene, camera, consumer and capture method in isolated working directories, and report screenshot-readback timing limits. Captures alone do not authorize game launch, executable replacement, gameplay control or publication; preserve saves/configs. Commit/push only within the user's current authorization.