Pass each parent-visible child log into direct parallel worker execution so tool before_execution reuses it instead of creating a second generic tool log.
This keeps native badge types such as wait/progress (HDL) intact through execution, updates the parallel helper DOX contract, and adds a regression test for log reuse.
Resolve tool-specific log objects for all parallel child jobs, removing the code_execution_tool-only special case and using native get_log_object() when available, with generic fallback.
Update parallel_tools docs to reflect the shared logging contract and add coverage for wait/ fallback behavior in parallel-tool tests.
Skip usr/.time_travel during self-update user-data backups and add the same Time Travel shadow-history exclusion to default BackupService metadata.\n\nAdd focused regressions proving regular user files stay in backup coverage while Time Travel shadow repositories are omitted.
Load active-project AGENTS.md path-chain guidance into the protocol area without duplicating the project root instructions.
Move the AGENTS.md protocol wording into a prompt template, reuse existing file/path helpers, and cover direct-path discovery plus prompt assembly with focused tests.
Search the raw user message when recalling relevant skills instead of the rendered history wrapper.
Replace the stopword catalog with structural matching: names use normal terms, tags/triggers use long terms or phrase matches, and descriptions require phrase matches.
Switch the Skills catalog UI/API from scope-wide prompt pins to current-chat history loading, and remove user-facing removal of loaded skills.
Disable legacy active-skill prompt injection so forgotten scope config no longer inflates new-chat prompts.
Clear the active Responses provider continuation when automatic history compression or manual chat compaction rewrites local history, while preserving stored response IDs for cleanup.
Add focused regressions for both compression paths so compacted chats do not keep stale provider-side context.
Use the code_exe log type for code_execution_tool children launched through parallel so the WebUI reuses the existing code execution renderer.
Add a focused regression test and document the parallel child-log contract.
Accept shell-style local MCP command values from the manager before spawning stdio clients, so configs like uvx workspace-mcp resolve to the correct executable and args.
Split collapsed option/value argument lines in the MCP manager, document the input contract, and add a regression for the google_workspace MCP shape.
Detect LiteLLM JSONDecodeError failures caused by the Responses mock streaming iterator trying to parse a real SSE stream.
Fall back to Chat Completions before any output is emitted, keeping custom OpenAI-compatible proxy prefixes provider-neutral.
Add a regression test for SSE-shaped JSON decode failures and update the transport DOX contract.
Verified with: PYTHONPATH="/home/eclypso/a0/agent-zero" conda run -n a0 pytest tests/test_stream_tool_early_stop.py tests/test_responses_architecture.py -q; PYTHONPATH="/home/eclypso/a0/agent-zero" conda run -n a0 python -m py_compile helpers/litellm_transport.py; docker exec -i -w /a0 9c436228a4a2 /opt/venv-a0/bin/python -m py_compile helpers/litellm_transport.py; git diff --check.
Only treat top-level JSON objects as tool roots during streaming, so a complete nested tool_calls item cannot end the stream before the parallel wrapper closes.
Add regression coverage for partial parallel wrapper snapshots.
Mark chats after successful persistence and reconcile saved chat IDs while building WebUI state snapshots so stale in-memory contexts disappear after their chat files are removed.
Keep fresh unsaved chats visible, skip running contexts, and cover the regression with snapshot and persistence tests.
Fall back to Chat Completions for provider/proxy Responses endpoint failures before any output is emitted.
Preserve streamed Chat Completions tool-call deltas as structured LLMResult function-call items, so fallback providers can still drive tools.
Document Docker host-gateway addressing for local model servers, where container localhost does not reach host loopback.
Verified with: PYTHONPATH="/home/eclypso/a0/agent-zero" conda run -n a0 pytest tests/test_stream_tool_early_stop.py tests/test_responses_architecture.py -q; PYTHONPATH="/home/eclypso/a0/agent-zero" conda run -n a0 python -m py_compile helpers/litellm_transport.py helpers/llm_result.py; git diff --check.
Ensure function parameter schemas include an explicit properties object before Responses requests are dispatched through LiteLLM.
Keep prompt/MCP permissive schemas compatible with stricter OpenAI-compatible chat validators and cover chat, legacy function, and native Responses tool inputs with regressions.
Let backup creation and restore cleanup use unlimited pattern scans so archives for agents with more than 50,000 files are complete. Keep UI preview and dry-run paths bounded for responsiveness while making the dry-run truncated flag safe for optional limits.
Add regression coverage for unlimited scans, backup creation, clean-before-restore, and restoring an archive entry past the old 50,000-file boundary.
Restore parallel normalization tolerance for provider/model outputs that stringify the tool_calls array inside tool_args.
The wrapper still expects normal tool-call objects after decoding, so nested parallel and document_query guards continue to run through the same validation path.
Keep fallback for endpoint-specific and Venice-style Responses payload rejections, but stop treating generic Bad Request validation text as proof that Responses is unsupported.
Add a regression that generic pre-output Responses 400s raise without trying Chat Completions.
Keep the valid-tool-root preference added for streamed snapshots, but avoid collecting and re-walking parsed candidates.
Remember the first parsed object/root while scanning and use a small predicate for tool-request recognition.
Keep the malformed SKILL.md warning behavior, but reduce the diagnostic machinery to a once-per-path warning with line numbers only for parser-owned structural errors.
Update the regression to assert the stable warning message without depending on the previous path/line/error dedupe key.
Keep the streaming early-stop caller untouched by making extract_json_root_string prefer the first complete JSON root that normalizes as a tool request while preserving its first-root fallback for non-tool JSON.
Add regression coverage for incidental JSON before a streamed tool-call envelope.
Default empty skills_tool calls to list, accept legacy method as a deprecated action alias, and warn when malformed SKILL.md frontmatter causes a skill to be skipped. Update the helper/tool DOX notes and focused regressions.
Save MCP image and image-resource payloads as scoped artifacts, return their paths in tool text and attachment metadata, and keep them model-visible through raw image history content. This gives downstream media delivery a real file path instead of only an inline data URL.
Treat pre-output Responses API 400 validation failures as unsupported so OpenAI-compatible providers that reject /v1/responses payloads can retry through Chat Completions. Also prefer a valid tool-call JSON object after leading prose or incidental JSON to reduce false misformat warnings.
Store explicitly loaded skill IDs as chat-wide context data while keeping full skill bodies in normal tool-result history. Reattach any loaded skill body that is no longer visible after compaction by reimporting the current skill from its source, without revision hashes or protocol reinjection.
Update compaction and summary prompts to preserve loaded skill names only, refresh DOX contracts, and add focused coverage for context-data persistence, legacy agent-data migration, duplicate suppression, and post-compaction reattachment.
Introduce generic project edit-data extension hooks so plugins can contribute named project settings sections without coupling helpers.projects to plugin-specific storage.
Move _model_config project LLM payload handling behind those hooks and preserve inherited global model settings unless a project override already exists or is explicitly selected.
Cover the inheritance regression, existing scoped overrides, multiple plugin sections, core-field collision protection, and extension payload filtering in focused project/model-config tests.
Use the active settings dictionary when scheduling MCP config updates so applying global MCP servers no longer references an undefined config variable.
Add a focused regression test covering the settings apply path and document the deferred MCP update contract in settings.py DOX.
Skip persisted tool-result files for BACKGROUND contexts, remove ephemeral direct parallel worker chat folders during cleanup, and ignore chat directories without chat.json during startup loading.
Add focused regressions for background output persistence, parallel worker cleanup, and orphan directory loading.
Release MCP config locks before awaited initialization or tool calls, isolate MCP session operations in disposable DeferredTask workers, and bound session cleanup so wedged transports cannot freeze later agent work.
Add deterministic MCP regression coverage for lock scope, config update initialization, cleanup timeouts, and isolated operation timeouts. Update the helper DOX contract for the new concurrency and cleanup behavior.
Add a Settings > Skills scanner section and modal backed by Snyk Agent Scan-oriented checks and prompts. Wire uploaded skill zips from Import Skills into the scanner before import, and add a preparation API that extracts archives into temporary scan roots without installing or executing them.
Harden shared skill zip extraction against traversal, backslash paths, and symlink entries, and cover the new flow with focused scanner tests.
Validate call_subordinate profile arguments against available agent profiles so missing profiles fail as repairable errors. Persist each agent's profile in saved chats and avoid flattening existing subordinate profiles during profile switches, settings refresh, or restart reload.
Create the .a0proj/skills directory during project metadata setup so Project Settings can open the skills folder for new projects. Repair the folder when loading existing project settings, and cover both creation and repair with regression tests.
Raises the Socket.IO heartbeat interval and timeout defaults so long context and prompt work do not trip python-engineio's empty packet queue timeout.
Adds positive-integer environment overrides and updates the runtime configuration regression test.
When compression hides an explicitly loaded skill body, reattach the current missing revision as a normal skills_tool history result under one fixed budget. Preserve skill name and revision metadata in automatic and manual compaction summaries without copying full skill bodies.
Reject document_query during parallel tool-call normalization before worker jobs are started. Add prompt guidance so the model avoids batching document_query, and update DOX plus regressions for the new sequential-only contract.
Prefer explicit tool_name examples and first prompt headings when deriving native Responses function tool names, falling back to the prompt filename only when no callable name is declared.
Add regression coverage for code_execution_tool, memory_load, call_subordinate, behaviour_adjustment, and filename-only fallback, and document the contract in responses_tools DOX.
Keep running parallel jobs alive when an await call reaches its timeout so agents can await the same job ids again instead of cancelling child work. Distinguish direct background tool workers from call_subordinate child chats so nested subordinate chats can use parallel normally while true worker recursion remains blocked. Update the parallel prompt, DOX notes, and regressions for non-destructive timeout and non-blocking collect semantics.
Hide A0 connector remote tool prompts unless a connected CLI advertises the matching capability. Remote file access enables text_editor_remote, F4-enabled remote execution enables code_execution_remote, and supported enabled Computer Use that is not in rearm-required state enables computer_use_remote.
Apply the same gate to Responses API function-tool generation, move the prompt hook to the active tool-prompt extension path, and update connector prompt wording, DOX, and regression coverage.
Verified with:
- conda run -n a0 pytest tests/test_a0_connector_prompt_gating.py tests/test_default_prompt_budget.py tests/test_responses_architecture.py -q
- document that parallel accepts full reply-shaped tool call objects and ignores planning-only fields
- steer prompt guidance toward one mixed batch for ready independent calls instead of splitting by tool type
- add normalization and prompt regression coverage plus matching DOX notes
- add a parallel wrapper and runtime for concurrent background tool calls
- run parallel call_subordinate jobs as child chats with visible subagent steps that match normal subordinate args
- render parallel child tool steps with normal tool-call args while keeping job handles in wrapper results and prompt extras
- group parallel child chats in the sidebar with persistent accordion and caret behavior
- add prompt, extension, DOX, and regression coverage
Wrap FastA2A agent card generation so Agent Zero advertises streaming support on A2A endpoints by default.
Add focused regression coverage for the agent-card capability rewrite and proxy wiring, and update the helper DOX profile for the new wrapper.
Teach the Responses fallback classifier to inspect status codes, wrapped exception types, and response bodies so LiteLLM NotFoundError wrappers that hide the /v1/responses URL still fall back to chat completions.
Keep rate-limit errors non-fallback and add a regression test for the OpenAIException detail-only 404 shape observed with providers that do not expose the Responses API.
Route Agent Zero turns through a LiteLLM transport layer that prefers the Responses API while preserving chat-completions fallback for providers without compatible endpoints.
Persist Responses metadata in history and agent state so provider-state continuation, local replay, native function-call execution, and stored-response cleanup survive normal chat workflows.
Normalize prompt caching by provider: OpenAI and Azure use prompt_cache_key and prompt_cache_retention, while Anthropic, Gemini, Bedrock, OpenRouter, and compatible chat providers keep block-level cache_control breakpoints and cached tool definitions.
Revamp the global and project MCP manager surfaces with list-first layout, clearer examples, a dedicated scanner modal, manager/raw toolbar parity, and local-command-first server creation.\n\nAdd server and tool search, plugin-style enable toggles, per-tool disabled_tools handling in the MCP backend, internal A0 MCP tool search, regression coverage, and updated DOX contracts.
Add project-scoped MCP server configuration with global/project merge semantics, a richer settings UI, and chat composer access. Introduce MCP config scanning plus project-aware status/detail/log/apply APIs while preserving the raw JSON editor. Strengthen MCP runtime handling for dotted tool names, timeouts, status accuracy, and project-aware tool execution, with focused regression coverage.
Add comprehensive file-level DOX documentation across the repo and update directory AGENTS.md indexes. Many new `*.py.dox.md` files were added under api, helpers, tools, plugins, extensions, webui, and other dirs to document endpoint purpose, ownership, runtime contracts, work guidance, and verification. Several AGENTS.md files were created or updated (agents profiles, api, docker, extensions, helpers, plugins, skills, webui components, etc.) to list child DOX files and clarify documentation/work guidance. Also add example and bundled profile DOX files (agent0, default, developer, hacker, researcher) and minor updates to helpers/dirty_json.py and its tests. These changes improve on-disk documentation coverage and establish the convention that each direct runtime file should have a matching `*.dox.md` describing its contracts and verification steps.
