Skipping problem-space modeling turns codebases into accidental data models. Running full architectural ceremonies on every dependency bump or one-line fix stalls delivery.
We partition Left-Hand Side (LHS) planning—problem models and invariants before Right-Hand Side (RHS) code—across three intake tracks.
1. Three-Track Planning Matrix
| Dimension | Track A: Greenfield, Major & Audits | Track B: Active Features | Track C: Fixes, Dependabot & Triage |
|---|---|---|---|
| Complexity | High / Very High |
Low / Medium (High if escalated) |
Low / Medium (narrow scope) |
| Main Tier | Carbon (Med/High) / Hydrogen (XHigh) |
Hydrogen (Med/High; XHigh on High) |
Hydrogen (Low/Med) / Flash |
| Subagents | pro (coach, research) |
flash_lite/flash (pro on High) |
flash_lite/flash (no pro/inherit) |
| Skills / Flow | /boost, /plan, semantic design |
/plan, domain Tier 2 skills |
Repro/changelog check or triage |
/grill-me |
2–3 deep loops (forces, failures) |
1–2 loops (0–1 if trivial) |
0 loops (1 if ambiguous/breaking) |
| Artifacts | Canvas, Parti, Matrix, ADR, Runbook, scratch/ |
2-part spec (4-part + Parti if High) |
Minimal 2-part spec or triage |
Track-specific convergence keeps planning proportional to risk:
- Track A (Greenfield, Major & Audits): Stays on LHS until forces settle. Throwaway
scratch/spikes (zero git footprint) calibrate Weighted Trade-off Scorecards (ADR-*.md) before decomposing into an EpicRunbookand<= 4-checkbox sub-issues. - Track B (Active Features): Uses a
2-partspec (Overview+MVP Scope), escalating to a4-partspec (Baseline,Out of Scope),Architectural Parti, andEntity-State Matrixfor state machines or security-sensitive paths. - Track C (Fixes, Dependabot & Triage): Optimizes for fast disposition. Deterministic fixes and minor dependency bumps skip
/grill-me; inbound requests triage toReady,Backlog, orClose/Merge.
Across all tracks, inherit and omitted model tiers remain banned so subagents never clone parent quota.
2. Hard LHS/RHS Boundaries
Two explicit contracts prevent premature drift and silent architectural decay:
Exploration != Sign-Off (Forward Gate)
Conversational steering (“option 1 looks good”) narrows the LHS design space without authorizing RHS work. Agents emit zero issues, branches, or edits until explicit promotion (“proceed” or “create the issue”).
Mechanical Demotion (Reverse Bridge)
When RHS implementation hits a runtime limit, deadlock, or platform constraint falsifying an LHS invariant or state guard, the agent halts rather than adding an ad-hoc workaround:
- Mark
Blocked: Halt the branch and cite the unmet invariant (“Platform X fails Guard Y due to Z”). - Demote & Amend: Return to LHS Concept Stage to revise the Architectural Parti, Entity-State Matrix, and
ADRbefore re-promoting.
3. Pros & Cons
Pros
- Proportional overhead: Matches
/grill-medepth, model burn, and artifacts to blast radius. - Protected problem space:
Exploration != Sign-Offandscratch/spikes prevent premature code collapse. - Self-healing invariants: Mechanical Demotion turns runtime blocks into durable ADR and matrix updates.
Cons
- Misclassification risk: Routing a structural Track A problem as Track B/C forces rework.
- Interview friction: Multi-loop
/grill-mesessions delay the first RHS commit. - Demotion pauses: Halting an active RHS branch for LHS revision trades speed for coherence.