Skip to content

Releases

v16.7.7

Choose a tag to compare

@github-actions github-actions released this 22 Nov 08:51

jeopi-agent-core

Added

  • Added WorkflowGuard and adaptive StepBudget to agentLoop/Agent with bounded model steps, consecutive-failure stop, non-user outer-cycle cap (including stop-boundary non-user steering), total soft-tool escalation cap, length-truncated tool-call continuation cap, and progress-resetting stall watchdog timeout.

Fixed

  • Preserved already-dequeued genuine user messages in emitted events and loop results when step_hard_cap halts the run instead of silently dropping them.

jeopi-cli

Fixed

  • Bounded automatic AgentSession re-entry continuations (unexpected-stop retry, thinking-loop retry, yield-queue idle injection, TTSR resume, and length-compaction retries) per user prompt and surfaced WorkflowGuard stops as structured session notices.
  • Ignored agent-attributed queued messages when deciding whether to override a no-progress compaction dead-end pause, preventing synthetic follow-ups from looping compaction indefinitely.

Full Changelog: v16.7.6...v16.7.7

v16.7.6

Choose a tag to compare

@github-actions github-actions released this 21 Nov 16:08

jeopi-ai

Fixed

  • Keep guardThinkingLoopStream armed across multi-block thinking and text parts separated by whitespace-only text deltas (such as Gemini Cloud Code Assist \n separators), disarming only after visible non-whitespace text or a tool call, and trigger ThinkingLoopDetector after 3 consecutive identical segments.
  • Normalize volatile error metadata (ISO timestamps, UUIDs, request/trace IDs, durations) in ToolCallLoopGuard and stall consecutive isError streaks on the same call signature even when error details vary.

jeopi-cli

Fixed

  • Preserve GeminiHeaderRunDetector across multi-block thinking parts separated by whitespace-only text within the same assistant message instead of resetting it on every thinking_start.
  • Inject the loop-guard redirect before retrying transient empty-response errors when the failed turn already emitted thinking content, preventing blind re-sampling into the same reasoning loop.

Full Changelog: v16.7.5...v16.7.6

v16.7.5

Choose a tag to compare

@github-actions github-actions released this 21 Nov 00:43

jeopi-cli

Added

  • Index UTF-8 project source by content and line range in the project database. Jeox can now search wiki passages paired with referenced project paths and use explicit relevance feedback to rank those pairs.

Fixed

  • Verify the .jeopi/db default in the Jeox tool integration test, restoring release validation for the new project index.

Full Changelog: v16.7.4...v16.7.5

v16.7.3

Choose a tag to compare

@github-actions github-actions released this 20 Nov 23:37

jeopi-cli

Fixed

  • Send one redirect per repeated tool-call streak. If the same single call keeps returning identical text after the redirect, pause the run with a notice instead of requesting another model turn.
  • Stop retrying an advisor batch when Cloud Code Assist reports a missing endpoint or model (404); later updates can still recover.

Full Changelog: v16.7.2...v16.7.3

v16.7.2

Choose a tag to compare

@github-actions github-actions released this 20 Nov 17:18

jeopi-ai

Fixed

  • Detect when consecutive reads of the same file return identical text even if the requested page changes, while allowing reads that return new content.

Full Changelog: v16.7.1...v16.7.2

v16.7.1

Choose a tag to compare

@github-actions github-actions released this 20 Nov 02:48

jeopi-ai

Fixed

  • Enhanced ToolCallLoopGuard to normalize Python/JS subprocess array arguments (e.g. ["git", "status"]), strip bash redirect and pipeline filters (head, tail, cat), and detect repetitive read offset patterns across turns to halt unproductive tool loops.

Full Changelog: v16.7.0...v16.7.1

v16.7.0

Choose a tag to compare

@github-actions github-actions released this 18 Nov 16:23

jeopi-cli

Fixed

  • Disposed JS eval VM contexts per owner on AgentSession disposal and SDK creation failure so finished subagents and sessions do not leak live JS worker contexts.
  • Bounded AgentLifecycleManager with LRU limits (maxLiveIdle = 2, maxAdopted = 16), dropped in-memory revive closures on park() when a persisted reviver factory is available, and detached aborted subagent sessions in finalizeSubagentLifecycle().
  • Shared a process-global subagent spawn semaphore (with separate top-level and nested pools) across TaskTool and eval's agent() bridge, lowered default task.maxConcurrency (8) and task.agentIdleTtlMs (60_000), and enabled a default 5 * 60 * 1000ms idle timeout for LSP clients.

jeopi-tui

Changed

  • Added incremental streaming Markdown lexing in Markdown component to avoid re-lexing unchanged prefixes during streaming updates.

Full Changelog: v16.6.29...v16.7.0

v16.6.29

Choose a tag to compare

@github-actions github-actions released this 18 Nov 02:49

jeopi-cli

Fixed

  • Isolated third-party skill provider toggles (claude, claude-plugins, codex, opencode, gemini, etc.) from native/agents toggles and filtered pre-dedup skill entries so disabled higher-priority providers cannot shadow enabled ones.

Full Changelog: v16.6.28...v16.6.29

v16.6.28

Choose a tag to compare

@github-actions github-actions released this 18 Nov 01:13

jeopi-agent-core

Fixed

  • Applied a 120_000ms SSE idle timeout to ProxyMessageEventStream so stalled proxy responses fail cleanly instead of hanging.

jeopi-cli

Fixed

  • Applied a 120_000ms SSE idle timeout to unguarded readSseJson consumers (image-gen, Codex/Perplexity web search, and MCP HTTP POST responses) to prevent hangs on stalled streams.

jeopi-utils

Fixed

  • Added IdleTimeoutError and bounded idleTimeoutMs handling to abortableSource, readLines, readJsonl, readSseEvents, and readSseJson, bounding stream cancel() cleanup so stalled streams never hang indefinitely.

Full Changelog: v16.6.27...v16.6.28

v16.6.27

Choose a tag to compare

@github-actions github-actions released this 17 Nov 04:30

jeopi-ai

Added

  • Shared cross-runtime OAuth 429 cooldown synchronization via mode-0600 ~/.agents/auth-pool/cooldowns.json (shared-cooldown-bridge.ts, wired into AuthStorage), with stat-gated in-memory caching for zero-copy repeat lookups and hermetic test isolation.

jeopi-cli

Added

  • Added fail-open 500ms local qev:8017 Noul rerank support (rerank parameter / JEOX_JEV_RERANK=1) with a 15s negative-health circuit breaker on jeox search/scope.

Changed

  • Pointed JeoxTool to the shared <cwd>/.jeo/db/jeo.db graph DB (auto-migrating .jeopi/db/jeo.db) with @wiki/ project cross-links and fast hasIndexedFiles() auto-index checks.
  • Pointed JeoxTool to the shared <cwd>/.jeo/db/jeo.db graph DB (auto-migrating .jeopi/db/jeo.db) with @wiki/ project cross-links and fast hasIndexedFiles() auto-index checks.(previously this shipped a GitHub Release with an empty body, e.g. v16.2.22); pass--allow-empty` for an intentional no-notes rebuild.

jeopi-mnemopi

Added

  • Bridged project-relevant ~/vaults/llm-wiki pages into @wiki/... virtual files/chunks and backticked <wiki> code_nodes/code_refs (references edges surfaced in whoCalls) while isolating temp workspaces (discoverProjectWikiFiles).

Changed

  • Unified defaultCodeDbPath() from .jeopi/db/jeo.db to .jeo/db/jeo.db (matching jeo-code and jeo-skills/scripts/jeox.py) with automatic legacy .jeopi/db/jeo.db migration on first access.
  • Enabled PRAGMA synchronous=NORMAL, temp_store=MEMORY, cache_size=-32000, and mmap_size=134217728 in enablePragmas(), added fast hasIndexedFiles() leaf check, and skipped rebuildEdges() on no-op incremental indexDir() runs.

Full Changelog: v16.6.26...v16.6.27