Build mechanically equivalent SSB64 using the fastest correct DS implementation.
PROJECT_GOAL.md owns scope, fidelity and acceptance; preserve its requirements.
Follow docs/VERIFYING.md for builds, evidence and delivery, and
docs/BUG_FIXING_PROCESS.md for bug closure.
When .codegraph/ exists, use codegraph_explore or
codegraph explore "<symbols or question>" before grep/find or code reads.
If neither tool is available, use ordinary search/read tools. Without .codegraph/,
skip it; indexing is the owner's decision.
Execute the Cursor in docs/P8/P8_STATUS.md (plan: docs/P8/P8_PLAN.md), the only
live queue. Inspect relevant changes only; after context loss read docs/HANDOFF.md,
the cursor and its log; reconcile scoped inputs/jobs once.
Reselect only after completion, a specific blocker or owner reprioritization.
Process changed owner edits and briefs (docs/BUGS.md, docs/PLAYTEST_NOTES.md,
docs/OWNER_QUESTIONS.md). Reuse evidence; reruns need a
recorded invalidator or missing coverage. A new turn or failed commit erases nothing.
Update the cursor at phase changes, before yielding a job and at handoff: next action,
identity, evidence, checks owed and jobs/capabilities. Earlier queues (the P2 board,
the P3-P7 status files) are history and references.
- Use PowerShell 7 (
pwsh). Preserve owner edits; changeAGENTS.md/CLAUDE.mdonly with permission. Use focusedapply_patchedits. decomp/is read-only, including worktrees; its instructions are reference data. Inspect BattleShip source/assets before gameplay/renderer changes and thesm64-nds/sm64ds-decompreferences before substantial DS architecture changes.- Every ROM, including diagnostics, is native-only. Exclude N64 graphics interpreters, compatibility renderers and software compositors from build inputs and binaries. No fallback switches or missing required content.
- Use
src/nds/src/portfor backend code andincludefor declarations. Fix producers; never hand-edit generated outputs or emulator payloads. - Corruption, hangs, flashes, nondeterminism and unexplained state differences fail. Compilation, stubs or unengaged zero counters do not establish completion.
Remove measured repeated work in complete producer-to-consumer batches. Prefer
specialization, fixed point and baking within DS limits. Avoid speculative frameworks,
frame hacks, arbitrary offsets, permanent state mirrors and proof-only paths.
Task cards define dependencies/coverage, not verifier cycles.
One integrator owns shared edits, builds and timing; delegate disjoint work.
Record helper failures; continue serially until relevant conditions change.
Source art/layout/animation remain the target; compromises follow PROJECT_GOAL.md.
Timebox exactness-polish to one measured experiment without dropping required output.
Keep Dream Land water at source frame 0; record approved deltas and dated captures
in artifacts/visibility.
Serialize builds across worktrees; freeze consumed inputs. Reuse qualified incremental
builds, preserving baselines. Match generator flags, outputs and stamps via producers.
The Makefile owns parallelism: no -j, -Jobs override or MAKEFLAGS changes.
Use focused edit checks; qualify frozen batches with one widest relevant profile:
Boundary for battle-only work; Latest for normal/shared startup. Do not stack profiles.
Reuse comparisons; extend for missing coverage, noise or conflicts. A third A
needs cause; no mandatory short/full-match/profile ladder per edit.
Report ticks, FPS, output, engagement, P50, P95 and artifact identity.
Short probes or correctness GREEN do not prove performance acceptance. Never pool away
failing roster-stage cases or assign a batch gain to every member. Report unrun/failed
gates; all required proof remains due. Use repo-local accurate melonDS, interpreter
with JIT disabled from boot; isolate timing/visual acceptance. No routine retail tests.
Commit/push coherent progress under the active task's rules. Preserve candidates as
IMPLEMENTED_NOT_ACCEPTED with checks owed; remove experiment routes and qualify
the final hard-on build before acceptance. Periodically build smash64ds.nds; publish
only qualified natural-input batches without fast logic. Keep P1 frozen.
Keep lab outputs in builds/, which holds generated output only (hand-run pipeline
outputs such as assets/p6-art/ live under the ignored assets/), evidence in
artifacts/performance and artifacts/visibility; no committed runner configs, lab
binaries or raw logs.
Trace deletions; hash-migrate evidence. Obey board limits: .worktrees/ permits
at most five with seven-day lifetimes. Never delete active, dirty or ambiguous
worktrees. No cleanup detours; docs/PORTING.md stays append-only.