Do

/do

Orchestrate end-to-end. Follow shared/instructions/gate-flow.md.

Procedure

1. Prep (deterministic) — scripts/Invoke-DoPrep.ps1 (MCP snapshot, minimal, context pack, living matrix start). If memory/.context-compact.md exists from a prior compact, continue from that pack only.
2. Gate 1 — Clarify — goal, scope, done, constraints.
3. Gate 2 — Research — delegate /research (shared/fixtures/delegatesTo-research.json); depth one; native fork if parallel.
4. Re-clarify if scope changed.
5. Gate 3 — ShortPlan — user confirm yes. New-ShortPlan / Confirm-Plan.
6. Gate 4 — Implement — Prefer Copilot Auto (10% discount). Use Invoke-LadderCascadePlan for the rung list. After each attempt score quality 0..1 (Get-NormalizedQualityScore / dimensions). Escalate when fail or qualityBelow — matrix picks next effort or family per escalatePolicy until Done (ok + quality ≥ min) or maxSteps. Always attach synth pack. Log with Save-MatrixEvidence.ps1 -QualityScore.
7. Context thrift (any time) — If Test-ShouldCompact / soft token warn: ask user, then scripts/Invoke-ContextCompact.ps1 (lean pack + prune ladder/matrix dumps + restore blank inject). Hard stop → handoff + new chat. Never dump full tip cards or chat history back into context after compact.
8. Gate 5 — /2080 — ≤ five recommendations.
9. Finish — scripts/Invoke-DoFinish.ps1. Stage /learn matrix-cell when a better start cell is evidenced.

Native parallel

After confirm only. Independent steps → native fork / context: fork. Sequential for gates, secrets, promote.

Session tokens

Test-SessionTokenThreshold — soft warn → compact (or handoff); hard stop → New-HandoffPack + new chat. No fake IDE compact API — we compact SkillsForge artifacts only (ADR O4).

delegatesTo

research, 2080