Port goclaw context pruning to match upstream TS design in
openclaw/src/agents/pi-hooks/context-pruning/:
- Opt-in default: prune only when mode="cache-ttl" (was opt-out)
- Remove Pass 0 per-result 30% guard (duplicated Pass 1 with different
suffix, caused wobble)
- Dedupe double prune call per iteration: PruneStage owns the single
entry point; loop_history only runs limitHistoryTurns + sanitizeHistory
- Add cache-TTL gate for Anthropic prompt cache: skip prune while cache
is live, scoped per-session via sync.Map
- Add context.pruned event emission for observability
- Configurable TTL as Go duration string ("5m", "30s")
BREAKING CHANGE: context pruning now opt-in. Add
contextPruning.mode: "cache-ttl" to config.agents.defaults to restore.
Migration 51 / SQLite v19 backfills mode="cache-ttl" for agents with
existing custom context_pruning config missing the mode field, so
previously-configured agents keep pruning after the opt-in flip.
NULL configs stay NULL (new opt-in default applies).
Web UI adds Cache TTL input + toggle wiring mode to cache-ttl/off.
- Replace char-based heuristic (chars/4) with tiktoken BPE for accurate
token counting, especially for non-ASCII content (Vietnamese/Chinese)
- Add pruningEstimator wrapper with tiktoken/fallback dual-path
- Raise default soft trim budget from 3K to 6K chars (3K head + 3K tail)
- Media tools (read_image, read_document, read_audio, read_video) get
higher soft trim budget (8K: 4K head + 4K tail) and skip hard clear
entirely — their vision/audio descriptions are irreplaceable
- Add per-result context guard (Pass 0): force-trim any single tool
result exceeding 30% of context window
sanitizeHistory dedup rewrites ToolCalls IDs but not RawAssistantContent.
Anthropic provider prefers RawAssistantContent → sends stale IDs → HTTP 400.
Fix: clear RawAssistantContent when dedup fires. Also merge consecutive
same-role messages (user↔user, assistant↔assistant) to fix role alternation
errors from session corruption.
Closes#532
- Replace prefix truncation with SHA-256 hash-based shortening for oversized tool call IDs (40-char OpenAI/Azure limit)
- Normalize provider-prefixed model IDs (e.g. openai/o3-mini) before capability checks for temperature and max_completion_tokens
- Add regression tests for ID collision, correlation, and prefixed model routing
Root cause: compaction threshold compared total prompt tokens (including
system prompt, tool definitions, context files) against budget. Sessions
with only 50-60k of conversation could trigger compaction when overhead
pushed total to 120k+.
Changes:
- Enable context pruning by default for all providers (was Anthropic-only)
- Add EstimateHistoryTokens() that excludes system messages from count
- Fix EstimateTokens() to include ToolCall argument sizes
- Calibrate overhead on first LLM response, subtract from threshold checks
- Mid-loop: prune old tool results first (at 70% budget), compact only if
still over budget after pruning
- maybeSummarize: subtract overhead from calibration base
- Remove applyContextPruningDefaults() (Anthropic-only auto-enable)
- Simplify context pruning UI: show only Keep Last Assistants by default,
hide advanced params behind collapsible, remove Mode selector
Replace single-value idRemap with queue-based idQueue so multiple tool
results sharing the same original ID pair correctly in encounter order.
Previously the map overwrote on duplicates, causing the second result to
be dropped and a synthetic placeholder to be emitted instead.
Simplified approach to fix duplicate tool_call_ids that cause HTTP 400 on OpenAI-compatible APIs (OpenRouter, vLLM, DeepSeek).
- uniquifyToolCallIDs: appends runID+iteration+index to all IDs at response time, guaranteeing cross-turn uniqueness via UUID
- sanitizeHistory: simple seen-set dedup for legacy sessions with pre-existing duplicate IDs
- Anthropic guard: skips ID rewriting when RawAssistantContent is present
- Unit tests for both paths + edge cases
Closes#283
sanitizeHistory now returns dropped count so callers know when orphaned
tool_use/tool_result messages were removed. When orphans are found in
buildMessages, the full session history is sanitized and persisted,
preventing repeated warnings on every request.
- Add SetHistory() to SessionStore interface and both implementations
- Adapt memoryflush caller to new two-return signature
- Change sanitize log level from Warn to Debug
Multi-agent AI gateway with WebSocket RPC, HTTP API, and messaging channel integrations.
Go port of OpenClaw with multi-tenant PostgreSQL, per-user isolation, security hardening,
and production observability.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
viettranx