@@ -4,7 +4,9 @@ import (
 	"context"
 	"encoding/json"
 	"errors"
+	"strconv"
 	"strings"
+	"unicode/utf8"
 
 	"github.com/Gitlawb/zero/internal/trace"
 	"github.com/Gitlawb/zero/internal/zeroruntime"
@@ -390,6 +392,27 @@ type compactionState struct {
 	// starts at 1.0 and converges via an EMA as each turn reports actual usage, so
 	// later turns compact nearer to real capacity. Zero is treated as 1.0.
 	calibrationRatio float64
+
+	// summarizerFactory lazily builds the dedicated (cheap) summarizer provider
+	// (Options.Summarizer) for the current main model. summarizerProvider
+	// memoizes the built provider and summarizerResolved records that the
+	// factory ran (it may legitimately return no provider: the main model is
+	// already the cheap one). summarizerBroken is the sticky fallback: once a
+	// build or call fails, the run stays on the main provider until the next
+	// model switch — a misconfigured cheap model must never break (or
+	// repeatedly delay) compaction.
+	summarizerFactory  func(ctx context.Context, mainModelID string) (Provider, error)
+	summarizerProvider Provider
+	summarizerResolved bool
+	summarizerBroken   bool
+	// window is the context window the threshold was derived from. The
+	// reactive free-prune stage compares against it (the limit the provider
+	// enforces), not against the 70% proactive threshold — a pruned history
+	// between the two would still succeed on retry.
+	window int
+	// mainModel is the model the run currently streams on; the summarizer
+	// choice is made against it and remade after a mid-run switch.
+	mainModel string
 }
 
 // calibrate folds one turn's (rawEstimate, actualPromptTokens) sample into the
@@ -424,17 +447,88 @@ func (state *compactionState) calibratedTokens(raw int) int {
 
 func newCompactionState(options Options, task *taskState) *compactionState {
 	state := &compactionState{
-		enabled:      options.ContextWindow > 0,
-		threshold:    compactionThreshold(options.ContextWindow),
-		preserveLast: options.CompactionPreserveLast,
-		onUsage:      options.OnUsage,
-		task:         task,
-		planner:      newContextPlanner(contextPlannerConfig{contextWindow: options.ContextWindow, serviceTier: options.ServiceTier}),
-		trace:        options.Trace,
+		enabled:           options.ContextWindow > 0,
+		threshold:         compactionThreshold(options.ContextWindow),
+		window:            options.ContextWindow,
+		mainModel:         options.Model,
+		preserveLast:      options.CompactionPreserveLast,
+		onUsage:           options.OnUsage,
+		task:              task,
+		planner:           newContextPlanner(contextPlannerConfig{contextWindow: options.ContextWindow, serviceTier: options.ServiceTier}),
+		trace:             options.Trace,
+		summarizerFactory: options.Summarizer,
 	}
 	return state
 }
 
+// updateWindow re-derives the compaction threshold after a mid-run model
+// switch. The low-water mark is reset: it was measured against the old
+// threshold, and keeping it could suppress a compaction the smaller window
+// now requires. <= 0 (unknown model) leaves everything unchanged.
+func (state *compactionState) updateWindow(window int) {
+	if window <= 0 {
+		return
+	}
+	state.enabled = true
+	state.window = window
+	state.threshold = compactionThreshold(window)
+	state.lowWaterMark = 0
+	if state.planner != nil {
+		state.planner.config.contextWindow = window
+	}
+}
+
+// switchModel records a mid-run model switch: the window is re-derived
+// (window <= 0 keeps the old one) and the summarizer choice is forgotten so
+// the next compaction re-resolves it against the new main model. A run that
+// started on the cheap model (no dedicated summarizer) and escalated to an
+// expensive one therefore regains cheap summaries — the exact workflow a
+// dedicated summarizer exists for.
+func (state *compactionState) switchModel(modelID string, window int) {
+	state.mainModel = modelID
+	state.updateWindow(window)
+	state.summarizerProvider = nil
+	state.summarizerResolved = false
+	state.summarizerBroken = false
+}
+
+// summarizeProvider returns the provider summarization calls should use and
+// whether it is the dedicated summarizer: that one when it applies to the
+// current main model and is healthy, else the main provider. The factory runs
+// at most once per main model; a build error marks the summarizer broken for
+// that model. Callers branch on the boolean rather than comparing Provider
+// values — interface comparison panics on a non-comparable dynamic type.
+func (state *compactionState) summarizeProvider(ctx context.Context, main Provider) (Provider, bool) {
+	if state.summarizerFactory == nil || state.summarizerBroken {
+		return main, false
+	}
+	if !state.summarizerResolved {
+		state.summarizerResolved = true
+		built, err := state.summarizerFactory(ctx, state.mainModel)
+		if err != nil {
+			state.summarizerBroken = true
+			return main, false
+		}
+		state.summarizerProvider = built
+	}
+	if state.summarizerProvider == nil {
+		return main, false
+	}
+	return state.summarizerProvider, true
+}
+
+// fitsWindow reports whether a calibrated size is safely under the limit the
+// provider enforces. The proactive threshold (70% of the window) is a trigger
+// for the paid summarizer, not the hard limit; the reactive path compares
+// against the window itself, with a small margin for estimate error, so a
+// history pruned to, say, 85% of the window takes the free retry.
+func (state *compactionState) fitsWindow(size int) bool {
+	if state.window <= 0 {
+		return size <= state.threshold
+	}
+	return size <= state.window-state.window/20
+}
+
 // maybeCompact runs proactive compaction at the top of a turn. It returns the
 // (possibly compacted) message slice. It is a no-op when compaction is disabled,
 // when the history is under threshold, or when the history has not grown past
@@ -528,10 +622,34 @@ func (state *compactionState) recover(
 		// context-limit error never triggers an unexpected summarization call.
 		return messages, false, nil
 	}
-	if state.reactiveAttempted {
+	if !isContextLimitError(errorMessage) {
 		return messages, false, nil
 	}
-	if !isContextLimitError(errorMessage) {
+
+	// CHEAP FIRST STAGE (mirrors maybeCompact): a context-limit failure can
+	// often be fixed for free by dropping superseded reads and pruning old
+	// tool-result bodies. It runs regardless of the one-shot reactive budget
+	// (it costs nothing and cannot loop: a second pass finds nothing left to
+	// prune). The free retry is taken only when the pruned history fits the
+	// window the provider enforces — otherwise a retry would fail the same
+	// way, so fall through to the paid summarizer in this same call.
+	toolTokens := estimateToolDefTokens(tools)
+	prunedAny := false
+	if pruned, reclaimed := pruneSupersededReadResults(messages); reclaimed > 0 {
+		messages, prunedAny = pruned, true
+	}
+	if pruned, reclaimed := pruneStaleToolOutput(messages, state.preserveLast); reclaimed > 0 {
+		messages, prunedAny = pruned, true
+	}
+	if prunedAny {
+		size := state.calibratedTokens(estimateTokens(messages) + toolTokens)
+		if state.fitsWindow(size) {
+			state.lowWaterMark = size
+			return messages, true, nil
+		}
+	}
+
+	if state.reactiveAttempted {
 		return messages, false, nil
 	}
 
@@ -571,11 +689,21 @@ func (state *compactionState) recover(
 // provider call. The summary stream intentionally does NOT forward OnText (so
 // compaction stays invisible on the user-facing surface), but it DOES forward
 // OnUsage so the summarizer's token cost is still counted by usage/budgeting.
+// The call goes to the dedicated (cheap) summarizer when one applies; if that
+// call fails the summarizer is marked broken for the current model and the
+// same slice is retried once on the main provider, so the caller sees at most
+// one (main-provider-quality) error.
 func (state *compactionState) summarizeClosure(ctx context.Context, provider Provider) func([]zeroruntime.Message) (string, error) {
 	if state.planner == nil {
 		state.planner = newContextPlanner(contextPlannerConfig{})
 	}
 	return func(toSummarize []zeroruntime.Message) (string, error) {
+		chosen, dedicated := state.summarizeProvider(ctx, provider)
+		summary, err := summarizeWithPlanner(ctx, chosen, toSummarize, state.onUsage, state.planner, state.trace)
+		if err == nil || !dedicated {
+			return summary, err
+		}
+		state.summarizerBroken = true
 		return summarizeWithPlanner(ctx, provider, toSummarize, state.onUsage, state.planner, state.trace)
 	}
 }
@@ -659,8 +787,41 @@ func summarizeMessagesOnce(ctx context.Context, provider Provider, messages []ze
 	return summary, nil
 }
 
+// summaryToolResultClamp bounds each tool-result body in the summarizer input.
+// Tool outputs dominate the elided middle's bulk (build logs, file reads,
+// search results), yet the summary instructions only need their gist — the
+// model that ACTED on the full output already turned it into decisions the
+// surrounding messages record. 2,000 chars (~500 tokens) keeps the useful
+// head+tail of a result while cutting the summarizer input by 5-10x on
+// tool-heavy sessions; the summarizer's own cost is the price being cut.
+const summaryToolResultClamp = 2000
+
+// summaryToolArgsClamp bounds each tool-call argument blob in the transcript:
+// a write_file call carries the whole file body in its arguments.
+const summaryToolArgsClamp = 500
+
+// clampForSummary head+tail-truncates text for the summarizer transcript,
+// snapping to rune boundaries so the clamp never splits a UTF-8 sequence.
+func clampForSummary(text string, max int) string {
+	if max <= 0 || len(text) <= max {
+		return text
+	}
+	head := max * 3 / 4
+	tail := max - head
+	for head > 0 && !utf8.RuneStart(text[head]) {
+		head--
+	}
+	start := len(text) - tail
+	for start < len(text) && !utf8.RuneStart(text[start]) {
+		start++
+	}
+	return text[:head] + "\n…[" + strconv.Itoa(len(text)-head-(len(text)-start)) + " chars omitted for summarization]…\n" + text[start:]
+}
+
 // renderTranscript flattens messages into a plain-text transcript for the
 // summarizer. Secret scrubbing already happened upstream at the tool boundary.
+// Tool-result bodies and tool-call arguments are clamped (see the constants
+// above): the summary needs what was done and decided, not the raw payloads.
 func renderTranscript(messages []zeroruntime.Message) string {
 	lines := make([]string, 0, len(messages))
 	for _, message := range messages {
@@ -670,13 +831,13 @@ func renderTranscript(messages []zeroruntime.Message) string {
 			if len(message.ToolCalls) > 0 {
 				calls := make([]string, 0, len(message.ToolCalls))
 				for _, call := range message.ToolCalls {
-					calls = append(calls, call.Name+"("+call.Arguments+")")
+					calls = append(calls, call.Name+"("+clampForSummary(call.Arguments, summaryToolArgsClamp)+")")
 				}
 				line += "\n[tool calls: " + strings.Join(calls, "; ") + "]"
 			}
 			lines = append(lines, line)
 		case zeroruntime.MessageRoleTool:
-			lines = append(lines, "tool result: "+message.Content)
+			lines = append(lines, "tool result: "+clampForSummary(message.Content, summaryToolResultClamp))
 		default:
 			lines = append(lines, string(message.Role)+": "+message.Content)
 		}