[Enhancement] Long-horizon tasks enhancement - #38
Merged
Conversation
When LLM calls code_search without pattern (intending to browse files), the error now explicitly guides it to use file_list instead. Tool description also clarified to prevent tool-selection confusion.
session_search (and other dynamically registered tools) were correctly disclosed to the LLM in TOOL_DEFINITIONS but failed name resolution because TOOL_REGISTRY was frozen at import time. Now _default_tool_registry() cache-invalidates and rebuilds when TOOL_DEFINITIONS or TOOL_HANDLERS grow, ensuring runtime-registered tools resolve correctly.
LLM frequently calls shell_list/shell_open/shell_send/shell_read/shell_close when it means terminal_list/terminal_open etc. Add static aliases so these resolve correctly without hitting the 'Unknown tool' path.
…s entirely _is_bypass_active() now checks both config-level (approval_bypass) and session-level (SessionAwareGate._bypass_all from 'Allow ALL for this session'). When either is true, workspace boundary checks AND dangerous command approval are skipped entirely — no orchestrator round-trip needed. Hardline patterns (rm -rf /, mkfs, etc.) remain blocked unconditionally upstream.
LLM repeatedly calls code_search with only path (intending to browse files). Instead of returning an error that the LLM retries or gives up on, now gracefully redirect to file_list(path, depth=2). Also removed pattern from required in schema so engine-level validation doesn't reject it first.
_normalize_decision() had a hardcoded allowed-set that predated the ALLOW_ALL_SESSION enum value. The user's selection was being silently converted to 'deny', so SessionAwareGate._bypass_all was never set.
…th spiral P0: Fix false progress detection + hard-gate repeated reads - evidence_count only increments for genuinely new source paths - _task_progress_marker includes repeated_reads count as negative signal - File reads beyond repeated_read_limit return summary instead of content - Write-then-read exception: edit resets read count for that path P1: Periodic checkpoint forcing + research ledger capacity expansion - New agent_checkpoint_interval config (default 15 rounds) - checkpoint_notice() forces intermediate deliverable in research posture - Research ledger: 200→600 chars/note, 24→32 items, 16K total cap with FIFO P2: Large-task strategy guidance + pre-compression knowledge extraction - _LARGE_TASK_PROTOCOL injected for complex tasks (decompose, search, record) - _auto_extract_findings() captures file paths and key content before compression - Extracted findings auto-injected into research ledger post-compression
- Fix runtime NameError: _research_protocol_for and _auto_extract_findings referenced non-existent 'LeapFlowEngine' class name (should be AgentEngine) - Fix _is_bypass_active: penetrate SmartApprovalGate._delegate wrapper to detect session-level bypass correctly in all configurations - Fix research ledger: _enforce_total_capacity truncates _next_step as last resort when no list buckets remain but total chars exceeds limit
Per AGENTS.md anti-hardcoding principle: 1. _research_protocol_for(): remove architecture_tokens and _large_task_keywords keyword tuples. Selection now driven by input length (structural numeric signal) with configurable threshold, not keyword scanning. 2. _tool_focus_metadata(): prefer ToolSpec.context_plane from registry manifest; name-based inference demoted to deprecated fallback with log warning. 3. _auto_extract_findings(): replace '"ok": false' substring sniffing with proper JSON parsing for robustness. 4. Error classification: prefer structured failure_code over string matching; substring check retained as compat fallback with TODO for removal.
P0: Enable cross-session memory recall - Remove session_scope filter from _prefetch_and_freeze_memory() so BM25 search can retrieve entries from any past session - Upgrade PCD CORE level from MemoryDisclosure.NONE to SESSION_SUMMARY so even simple queries get basic memory context P1: Automatic session summary generation + proactive history injection - _generate_session_summary(): heuristic extraction of goal/tools/outcome at session end, persisted as SESSION_SUMMARY kind in DuckDB - _prefetch_and_freeze_memory() Layer 0: injects last 5 session summaries into memory context at session start - SemanticMemoryProvider.query_recent_summaries() for efficient retrieval - Session title updated to descriptive summary on end P2: System prompt guidance + session title improvement - Guidelines added: use session_search for past task recall - end_session() accepts optional title parameter for descriptive updates
- Fix session summary not reaching DuckDB: MemoryManager.insert() routes SESSION_SUMMARY to narrative provider (MEMORY.md), but query_recent_summaries() reads from DuckDB. Added explicit semantic.insert() to ensure cross-session retrieval works. - Increase message fetch limit to 200 for better last_assistant extraction - Deserialize metadata JSON in query_recent_summaries() return values - Add logger.debug for Layer 0 task history injection failures
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.