Files
zero/internal/providers/compaction_model.go
Kevin CodexandClaude Fable 5 6645f9744b perf(agent): cheap compaction — dedicated summarizer model, input cap… (#532)
* perf(agent): cheap compaction — dedicated summarizer model, input caps, free reactive prune, window re-derive

Compaction summarization was the single most expensive recurring event
in a long run: the full elided middle — verbatim tool bodies included —
went to the session's main model at full price (a ~100k-token call near
a 200k window, recursively multiplied on overflow).

Four changes, each independently safe:

- Dedicated summarizer model. Compaction summaries route to a cheap
  model resolved as ZERO_COMPACTION_MODEL > preferences.compactionModel
  > a curated default on official endpoints (Haiku 4.5 / Gemini
  Flash-Lite / gpt-4.1-mini; custom-endpoint catalogs are unknowable so
  they get none). Built lazily on the first paid compaction; ANY
  failure permanently falls back to the main provider for the run, so a
  misconfigured summarizer can never break compaction. "main" at any
  level forces the old behavior.

- Summarizer input caps. renderTranscript clamps tool-result bodies to
  2,000 chars and tool-call arguments to 500 (head+tail, rune-safe) —
  the summary needs what was done and decided, not raw payloads. Cuts
  summarizer input 5-10x on tool-heavy sessions.

- Free reactive prune. recover() now runs the same prune-first stage as
  the proactive path: a context-limit error is first answered by
  pruning stale tool bodies at zero cost, without consuming the
  one-shot reactive budget; the paid summarizer only runs when a later
  error finds nothing left to prune.

- Window re-derivation on model switch. Options.ContextWindowFor
  (wired to the model registry in exec and the TUI) lets the loop
  re-derive the compaction threshold after escalate_model, closing the
  documented KNOWN LIMITATION where a switch to a smaller-window model
  could overflow and a larger one over-compacted.

## Rebase notes (2026-08-25, onto 2749447c)

Rebased across 214 commits. main now summarizes through the context planner
(#879) and switches models via turn sessions, so the agent side was reworked
rather than merged textually:

- Summarizer selection lives on compactionState and is resolved lazily per
  main model through summarizeWithPlanner (planner + trace preserved); a
  failing dedicated summarizer is marked broken and the same slice is
  retried once on the main provider.
- Options.Summarizer now receives the main model in force:
  `func(ctx, mainModelID string) (Provider, error)`, and may return
  (nil, nil) when no dedicated model applies. After a successful
  escalate_model switch the loop calls compactor.switchModel, which
  re-derives the window and re-resolves the summarizer — a run that started
  on the cheap model and escalated regains cheap summaries (review P2).
- recover()'s free prune stage runs before the one-shot reactive guard
  (pruning is free and cannot loop) and mirrors maybeCompact's two prune
  passes; the free retry is taken only when the pruned history fits the
  window the provider enforces (with a 5% estimate margin), otherwise the
  same call falls through to the paid summarizer. Gating on the 70%
  proactive threshold would have sent histories in the 70-95% band to the
  summarizer unnecessarily (review P1 and the follow-up question).
- ContextWindowFor is wired through the turn-session switch path; the
  planner's window is updated alongside the compactor's.

Tests: the PR's suite is ported to the new seams, plus insufficient-prune
fall-through, free prune after the budget is spent, window-vs-threshold
gating, and summarizer re-resolution across a model switch.

* fix(agent): branch on a dedicated-summarizer flag instead of comparing providers

Review follow-ups for #532:

- summarizeProvider now reports whether the dedicated summarizer was
  selected and summarizeClosure branches on that flag; the previous
  `chosen == provider` compared Provider interface values, which panics
  when the dynamic type is not comparable. Regression test uses a provider
  with a slice field for both the summarizer and the main provider.
- CompactionModelID's curated-default test covers the official OpenAI
  profile alongside Anthropic and Google.

Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_012JqvAf7zsP1qYFVZayKXHq

---------

Co-authored-by: Claude Fable 5 <[email protected]>
2026-08-25 21:18:04 +08:00

109 lines
4.1 KiB
Go

package providers
import (
"context"
"os"
"strings"
"github.com/Gitlawb/zero/internal/config"
"github.com/Gitlawb/zero/internal/zeroruntime"
)
// Compaction summarization model selection.
//
// Compaction summaries used to run on the session's main model — near a 200k
// window one summarization is a ~100k-token input call at full price, the
// single most expensive recurring event in a long run. A cheap model handles
// the "dense factual summary" task fine, so summarization is routed to one
// whenever the target is knowable.
//
// Resolution order: ZERO_COMPACTION_MODEL env > preferences.compactionModel
// config > a curated per-kind default (OFFICIAL endpoints only) > "" (use the
// main model). The literal value "main" at any level forces the main model.
// A wrong choice is never fatal: the agent falls back to the main provider on
// the first summarizer failure for the rest of the run.
// CompactionModelEnv overrides the compaction summarization model.
const CompactionModelEnv = "ZERO_COMPACTION_MODEL"
// compactionMainSentinel disables the dedicated summarizer explicitly.
const compactionMainSentinel = "main"
// Curated cheap summarizer defaults, applied only for provider kinds whose
// model catalog is guaranteed (official endpoints). Custom / *-compatible
// endpoints get NO default — their catalogs are unknowable, and a guaranteed
// failed call per run is worse than main-model prices.
const (
defaultAnthropicCompactionModel = "claude-haiku-4-5-20251001"
defaultGoogleCompactionModel = "gemini-2.5-flash-lite"
defaultOpenAICompactionModel = "gpt-4.1-mini"
)
// CompactionModelID returns the model that compaction summarization calls
// should use for the given profile, or "" to use the session's main model.
// preference is the resolved config value (preferences.compactionModel).
func CompactionModelID(profile config.ProviderProfile, preference string) string {
value := strings.TrimSpace(os.Getenv(CompactionModelEnv))
if value == "" {
value = strings.TrimSpace(preference)
}
if value != "" {
if strings.EqualFold(value, compactionMainSentinel) {
return ""
}
return value
}
return defaultCompactionModel(profile)
}
func defaultCompactionModel(profile config.ProviderProfile) string {
metadata, err := ResolveRuntimeMetadata(profile, Options{})
if err != nil {
return ""
}
var candidate string
switch metadata.ProviderKind {
case config.ProviderKindAnthropic:
candidate = defaultAnthropicCompactionModel
case config.ProviderKindGoogle:
candidate = defaultGoogleCompactionModel
case config.ProviderKindOpenAI:
candidate = defaultOpenAICompactionModel
default:
return ""
}
// Session already runs on the cheap model: a dedicated summarizer buys
// nothing and would only add a second provider instance.
if strings.EqualFold(strings.TrimSpace(metadata.APIModel), candidate) {
return ""
}
return candidate
}
// CompactionSummarizerFactory returns the lazy constructor the agent loop
// calls for its dedicated compaction summarizer (agent.Options.Summarizer).
// The compactor passes the main model in force at that moment — on the first
// paid compaction and again after a mid-run model switch — so the choice
// tracks escalations: a run that starts on the cheap model gets no dedicated
// summarizer, and gains one once it escalates to a pricier main model.
// Returning (nil, nil) means "summarize on the main provider". The factory
// itself is nil only when no provider builder exists. Shared by the exec and
// TUI front ends so the selection rules cannot drift between them.
func CompactionSummarizerFactory(base config.ProviderProfile, preference string, newProvider func(config.ProviderProfile) (zeroruntime.Provider, error)) func(context.Context, string) (zeroruntime.Provider, error) {
if newProvider == nil {
return nil
}
return func(_ context.Context, mainModelID string) (zeroruntime.Provider, error) {
profile := base
if strings.TrimSpace(mainModelID) != "" {
profile.Model = mainModelID
}
modelID := CompactionModelID(profile, preference)
if modelID == "" {
return nil, nil
}
profile.Model = modelID
return newProvider(profile)
}
}