kronic
4ac1ae5a2e
Realigns the skill with how sessions actually run now — it predated the Workflow tool and the ultracode two-review practice. Driven by an analysis of the full session history + the 26-file memory corpus + skill-authoring research, adversarially reviewed. Key changes: - Workflow-first orchestration: drop the dead manual "swarm (<=N agents)" model and the impossible 3-5-agent impl swarm; implementation is serial orchestrator MCP edits. New references/workflow-patterns.md (ground fan-out + design-review lens/critic) replaces agent-briefs.md. - Pre-code design-review + post-impl diff-review are now first-class gated phases (the spine that catches what green tests + a clean Play miss). - Size by blast-radius / netcode-heaviness, not time estimates. - Lean SKILL.md (222 -> 131 lines, -25% KB): leans on CLAUDE.md instead of duplicating its MCP cheat-sheet / anti-patterns / error-recovery. - ctx7 CLI / find-docs mechanism (the MCP verbs are gone); live-verified 6.5-era library pins; kill the dead NetCodeTestWorld test path. - Encode operator gates (no-time-estimates, present-forks, never-defer, tuning-autonomy) + the highest-recurrence MCP-edit / Workflow gotchas. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
49 lines
5.6 KiB
Markdown
49 lines
5.6 KiB
Markdown
# Workflow patterns for /dots-dev
|
|
|
|
The read-only fan-out phases run as deterministic **Workflows** (the `Workflow` tool), not hand-spawned agents. This file holds the two reusable patterns SKILL.md names. **There is no implementation pattern** — implementation is always serial orchestrator MCP edits (SKILL.md Phase 5), never a Workflow.
|
|
|
|
Two rules apply to every pattern here:
|
|
- **Check the failures list.** A Workflow whose agents die on a session/quota limit returns empty findings that look like a clean pass. If any agent died, wait for the reset and re-run with `resumeFromRunId` — never read an empty result as "no findings."
|
|
- **Scope guardrails positively.** Tell each agent what it MAY do (read-only; these tools), never "do NOT use tool X" — a blanket prohibition gets read as a *user* rule by the auto-approver and can deny the orchestrator's own later calls. Every brief here is read-only; pass only the upstream summary downstream.
|
|
|
|
---
|
|
|
|
## Pattern A — Ground fan-out (Phase 1, Feature scope)
|
|
|
|
Three read-only lens agents in `parallel()`; the orchestrator synthesizes the ≤400-word Project Brief and discards the raw returns.
|
|
|
|
**A1 — Code/asset graph (read-only).**
|
|
> Map the code surface relevant to `{goal}`. Prefer serena `find_symbol`/`find_referencing_symbols`; if serena's C# LSP stalls on the Unity solution, use `Glob`/`Grep` on `Assets/_Project/**/*.cs` + `**/*.asmdef`. Identify the existing components/systems/bakers touching `{surface}`, the asmdef each lives in (Simulation / Client / Server / Authoring), and any `*Math.cs` helper or system you'd otherwise duplicate. `manage_scene get_hierarchy` for relevant subscenes. Return a terse symbol/asmdef map (names + paths), not file dumps.
|
|
|
|
**A2 — Knowledge state (read-only).**
|
|
> Search the in-repo vault (obsidian-cli) and `basic-memory` for design docs, decision records, and notes touching `{goal}`. Return: relevant doc paths, any **locked** decisions that constrain `{goal}`, open questions already recorded. Quote ≤2 load-bearing lines each; link by path otherwise.
|
|
|
|
**A3 — Live state (read-only).**
|
|
> Read `mcpforunity://editor/state`, `project/info`, `instances`. `read_console(types=["error","warning"], count=20)` for a baseline. Report the **installed** Entities + Netcode versions (`packages-lock.json`) so ctx7 lookups pin correctly. Flag any pre-existing compile/Burst error.
|
|
|
|
---
|
|
|
|
## Pattern B — Design-review (lens + adversarial-critic)
|
|
|
|
Used **pre-code** over the plan (Phase 4) and **post-impl** over the git diff (Phase 7) — identical lenses, two inputs. Stage 1 = three lens agents in `parallel()` over ground-truth code; stage 2 = synthesize; stage 3 = adversarial critics that try to *refute* each finding against the actual code. **Findings default-refuted** — a finding stands only if a critic confirms it against the code. Post-impl additionally: **flag any simplification of a reviewed design as a likely regression** (this is the catch green tests + a clean Play miss). The synthesized output is a **durable Build Spec** — keep it (it feeds the build and the DR), unlike research which is discarded.
|
|
|
|
**B1 — Netcode / relevancy lens (read-only).**
|
|
> Review `{plan|diff}` against ground-truth code for netcode correctness: ghost ownership (predicted vs interpolated vs ownerless — move ownerless interpolated ghosts server-only); `[GhostField]` vs server-only fields; `GhostRelevancy` region splits (`SetIsIrrelevant`, untagged-stays-relevant); singleton-collision (a second `StorageEntry`/buffer of the same type → resolve by distinct tag, never `GetSingleton<T>`); command-vs-input choice (one-shot shared actions → `IRpcCommand`, not a dropped `InputEvent`); RPC wire-hash stability (unconditional struct, re-mean-don't-rename). Return findings: claim · file:line · why it's wrong · fix.
|
|
|
|
**B2 — Determinism / prediction lens (read-only).**
|
|
> Review for rollback-safety: `PredictedSimulationSystemGroup` runs multiple times per frame → idempotent, `.WithAll<Simulate>()`, no wall-clock/`Time.deltaTime`/`System.Random`; tick sentinels routed through `TickUtil.NonZero` + compared with `NetworkTick.IsNewerThan` (never raw `uint <`); predicted-physics group ordering (OrderFirst quirks); swept-not-point hit tests (tunnelling); ECB at-most-once destroy. Cross-assembly generics + enums in a Bursted system → ICE (store ops as `byte`). Return findings as in B1.
|
|
|
|
**B3 — Reuse / scope lens (read-only).**
|
|
> Does `{plan|diff}` duplicate an existing system or `*Math` helper instead of reusing it? Is the scope minimal — anything over-built, or a feature that could reuse a replicated component already on the wire? Any MonoBehaviour-world shortcut (MonoBehaviour singleton, ScriptableObject "service" for runtime sim, coroutine for sim) where ECS is correct? Any new `IAspect` (deprecated)? Return findings as in B1.
|
|
|
|
**Critic (stage 3, read-only).**
|
|
> Here is a candidate finding: `{finding}`. Try to REFUTE it by reading the actual code at `{paths}`. Default to "refuted" unless the code confirms the problem is real and reachable. Return: refuted | confirmed · the code evidence · (if confirmed) severity + the minimal fix.
|
|
|
|
---
|
|
|
|
## Orchestration notes
|
|
|
|
- Sizes scale to the ask: a small ghost change → one lens + a single critic pass; "thoroughly review this slice" / ultracode → all three lenses + 3-vote adversarial critics + a completeness pass.
|
|
- Keep agent context lean — each gets the plan/diff slice + the relevant ground-truth paths, not the whole session.
|
|
- The Build Spec is line-numbered and lives at `Docs/Vault/.../<Slice>_Build_Spec.md`; the run-ID (`wf_…`) is recorded in the session log / DR for `resumeFromRunId`.
|