[erikarenhill]

Summary

Fixes a regression reported on planner_arbitrage: when a cloud cuts PV mid-charge-slot, the EMS imports from grid to maintain the plan's charge power instead of letting the battery absorb only what PV is actually delivering. Pre-energy-path behaviour was the opposite — battery soaked up only the available surplus, grid stayed near the plan.

Root cause. useEnergyPath executes BatteryEnergyWh as remainingWh × 3600 / remainingS without consulting live gridW. When forecast PV doesn't materialise, the formula still demands the planned charge power and grid fills the gap. The reactive replan in mpc/service.go:266-290 would eventually catch up, but it's gated by a 500 Wh PV-error integral and a 60 s cooldown — typically ≥10 min of grid import before a fresh plan re-decides.

Fix. Use the plan's own gridW forecast (already computed by the DP as mpc.Action.GridW) as a charge-direction-only soft reactive cap on targetTotalW. When targetTotalW > 0 (charge slot) AND live gridW − plannedGridW > 100 W (importing more than plan, outside deadband), back off the target by the gap, floored at zero so the cap can never flip dispatch direction.

Why charge-only. The mirror case (discharge slot, live gridW more negative than plan) is intentionally NOT clamped:

• The battery still delivers its planned discharge Wh — the "extra" export comes from load undershooting forecast, not over-discharging. Backing off would leave Wh in the battery for a future slot the DP already evaluated and rejected against this one.
• Economics are asymmetric: extra import during a charge slot costs real money (paying for energy the plan assumed PV would supply, often at peak prices the DP would never have chosen for charging); extra export during a discharge slot is bonus revenue at the slot's chosen export price.
• The discharge-direction divergence (live import > plan during a discharge slot, e.g. load surged) is left to the reactive replan + downstream clamps (fuse, SoC floor, EV-discharge cap).

The opposite-direction case for charge slots (live gridW more negative than plan, PV came in higher than forecast) is handled by the existing PV surplus absorber that opportunistically adds charge.

PlannedGridW is a pointer on control.SlotDirective so existing tests and legacy callers opt out via nil. Wired through main.go's SlotDirectiveFunc adapter.

Files

• internal/control/dispatch.go — PlannedGridW *float64 field + charge-only cap logic in useEnergyPath
• internal/mpc/service.go — SlotDirective.GridW populated from a.GridW
• internal/mpc/service_test.go — assertion that SlotDirectiveAt propagates GridW (silent-regression guard)
• cmd/forty-two-watts/main.go — bridges mpc.SlotDirective.GridW → control.SlotDirective.PlannedGridW
• internal/control/control_test.go — regression tests
• docs/safety.md — new §8 "Plan-grid soft reactive cap" documenting mechanics + asymmetry rationale

Test plan

• go test ./internal/control/... — green
• go test ./... — green including e2e
• Regression: cap fires when PV drops mid-charge-slot
• Regression: cap does NOT fire on discharge slot with extra export (guards against re-symmetrising)
• Regression: deadband (≤100 W divergence) keeps cap silent
• Regression: nil PlannedGridW (legacy callers) bypasses cap
• Wiring test: SlotDirectiveAt propagates GridW from Action.GridW
• Field-deployed to dev Pi running planner_arbitrage (v0.78.1-2-g64d07dc, container healthy)
• Field validation during a cloud-pass charge slot — confirm battery target backs off and grid stays near plan

Review history

• 0219cae — original feature (charge + discharge cap)
• 64d07dc — self-review caught the discharge branch was semantically wrong; removed it, flipped test to regression guard, added docs/safety.md §8
• 228b99c — Copilot review pass: docstring drift fixes, dropped duplicate ptr helper, added SlotDirectiveAt.GridW test

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR addresses a regression in the energy-allocation (“energy dispatch”) execution path where, during a planned charge slot, unexpected PV drop could be compensated by unintended grid import to maintain the planned charge power. The fix plumbs the plan’s per-slot grid power forecast into dispatch and uses it as a soft reactive cap to reduce battery charge when live grid import exceeds the plan.

Changes:

• Add a per-slot planned grid power forecast (GridW → PlannedGridW) and wire it from MPC → main adapter → control dispatch.
• Apply a soft cap in ComputeDispatch’s energy path to back off charging when live grid import exceeds the plan by a deadband.
• Add regression tests for the cap behavior and document the mechanism in docs/safety.md.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
go/internal/mpc/service.go	Adds SlotDirective.GridW and populates it from mpc.Action.GridW.
go/internal/control/dispatch.go	Introduces SlotDirective.PlannedGridW and applies the plan-grid soft cap in the energy-dispatch path.
go/cmd/forty-two-watts/main.go	Bridges mpc.SlotDirective.GridW into control.SlotDirective.PlannedGridW.
go/internal/control/control_test.go	Adds regression tests for the planned-grid cap behavior.
docs/safety.md	Documents the plan-grid cap rationale, mechanics, and intended asymmetry.

Comments suppressed due to low confidence (1)

go/internal/control/dispatch.go:988

• The cap logic is explicitly gated by targetTotalW > 0 (charge slots only). The PR description mentions a symmetric discharge/export cap as well; either update the PR description (if charge-only is intended) or extend the implementation/tests to match the described discharge behavior.

		if currentDirective.PlannedGridW != nil && targetTotalW > 0 {
			const planGridDeadband = 100.0
			gridErr := rawGridW - *currentDirective.PlannedGridW
			if gridErr > planGridDeadband {
				adjusted := targetTotalW - gridErr

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.