Agent Changelog

v0.2.0(2026/06/17)¶

This release focuses on beak-agent runtime diagnostics, AI Hub error visibility, Skill invocation flow, and obsycli interaction rendering. Key updates include:

New Features¶

• obsycli adds the local ~/.obsycli/config.toml configuration file. Users can customize footer status fields through footer_fields, including agent/task, model, token, context, permission, and plan status (b0ebe75)
• obsycli input editor now persists local input history, so users can recover original prompts from history after either successful or failed sends (b0ebe75)

Bug Fixes¶

• Fixed Agent treating structured AI Hub error codes as generic LLM HTTP errors. It now exposes stable stop_reason values such as INVALID_BODY, VALIDATION_ERROR, AUTH_INVALID, AGENT_API_KEY_DISABLED, DAILY_CREDIT_LIMIT_REACHED, MONTHLY_CREDIT_LIMIT_REACHED, PROTECTED_SECRET_INVALID, and NO_AUTO_MODEL, with matching guidance in obsycli (a09dcaa)
• Fixed OpenAI-compatible LLM responses wrapped in an AI Hub envelope not being parsed correctly. Agent can now unwrap the real chat completion from the data field and preserve response model information (b0bc4e2)
• Fixed query_session_events missing cursor-based continuation and returning unstable error classifications. It now adds a cursor parameter and returns missing-argument or history-read failures as structured tool errors (b0bc4e2)
• Fixed render_document potentially overwriting an existing target file. Document generation now creates files exclusively to avoid accidental overwrites (b0bc4e2)
• Fixed obsycli session lists still showing closed or closing sessions and unclear empty-list output. Closed states are now filtered by default and empty results render as an empty table (b0ebe75)

Improvements¶

• Improved Agent user-visible language behavior. Final answers, tool commentary, progress feedback, approval reasons, task insights, Agent UI copy, and close summaries now prefer the original user request language instead of being affected by Skill docs, schemas, or repair prompts (b0ebe75)
• Removed the Agent local use_skill tool call. Skills continue to be used through read_skill/read_skill_file and Prompt instructions, preventing models from mistaking ordinary Skills for executable tool calls (b0bc4e2)
• Improved Python tool error handling with structured errors for missing code, non-zero exits, timeouts, missing python3, and oversized output. Oversized output is truncated into a short output_too_large error to avoid polluting later context (b0bc4e2)
• Improved Agent runtime stop results by adding titles, summaries, and continuation actions for AI Hub errors, and by reducing redundant completed/blocker output when the maximum tool-call count is reached (a09dcaa)
• Improved token usage and observability events. The LLM response model is written into token metadata, and obsycli can show the actual response model in the footer to help diagnose automatic routing and model fallback (b0ebe75)
• Improved beak-agent --version output with branch, commit, release date, and Go version, making it easier to identify deployed agent/obsycli artifacts (b0ebe75)
• Improved Agent local tool registration by extracting registry construction and adding regression coverage for tools such as send_email, web_search, python, render_document, and read_skill (b0bc4e2)
• Added agent/obsycli integration cases for LLM error recovery, Plan Mode, Python tool execution, session event queries, document rendering, Skill reads, email sending, Web Search, stream reconnect, task insights, and session close summaries (b0bc4e2)

v0.1.9(2026/06/15)¶

This release focuses on beak-agent default model configuration, a more complete uninstall flow, and release artifact completeness. Key updates include:

New Features¶

• Agent can now omit LLM_MODEL; the runtime uses default as the model identifier, and both the installer and Beak install-config no longer require users to provide a model configuration (d096537)
• Agent releases now include uninstall.sh, allowing users to fetch the uninstall script from the release URL. Packaging, upload, and local validation flows now check that this uninstall script exists (1e2c13f)

Bug Fixes¶

• Fixed upgrades potentially preserving an old active LLM_MODEL entry in agent.env. The installer now removes that old assignment and keeps only the default comment, preventing local stale model values from overriding the runtime default (d096537)
• Fixed LLM Chat span response_model always matching the request model. It now prefers the upstream response body's model field and falls back to the request model only when the field is missing or unavailable (d096537)

Improvements¶

• Completed the Agent uninstall scope. uninstall.sh --yes now stops and removes the beak-agent-update-check service/timer, removes updater/update-check binaries and /etc/sudoers.d/beak-agent, and fixes test-root residual process checks from matching host processes by mistake (1e2c13f)
• Clarified uninstall boundaries. uninstall.sh does not remove skill runtime dependencies installed or checked by skill-dep.sh, such as system commands, language runtimes, Python/Node packages, fonts, or other shared host components (1e2c13f)
• Updated agent installation, AgentRunner configuration, self-hosted installation, and uninstall documentation to match the current install/uninstall flow (d096537, 1e2c13f)
• Added tests for this release's configuration, LLM span, Agent uninstall, release packaging, and OSS upload manifest changes. test-release-agent.sh no longer depends on python3, making it compatible with slimmer CI environments (d096537, 1e2c13f)

v0.1.8(2026/06/12)¶

This release focuses on beak-agent runtime diagnostics, tool loop stability, Skill workflows, and obsycli rendering fixes. Key updates include:

New Features¶

• Agent adds the collect_bug_report tool, which collects runtime information, recent logs, metrics status, redacted env/config data, and profile cache diagnostics, then packages and uploads them to the current task session to help troubleshoot Agent issues (7e073a1)

Bug Fixes¶

• Fixed missing stable stop reasons for LLM timeouts, rate limits, upstream errors, and network errors. Agent runtime now returns structured reasons such as llm_timeout/llm_rate_limited/llm_upstream_error/llm_network_error with suggestions for continuing (7e073a1)
• Fixed insufficient diagnostics when tool call arguments contain invalid JSON, the same failed action is repeated, or the same class of tool error repeats. Stop results now include recently completed items, blockers, and safe_next_actions (7e073a1)
• Fixed obsycli potentially double-counting token usage when streaming token usage and final token usage arrive for the same reply. Delta token events are now deduplicated by reply_to_message_uuid (7e073a1)
• Fixed obsycli potentially mis-parsing arbitrary strings as attach session references. It now accepts only numeric indexes or explicit session IDs prefixed with session_ or session- (7e073a1)

Improvements¶

• Improved Agent tool loop implementation by splitting commentary, result envelopes, error classification, LLM calls, Plan Mode, repair, and stop checkpoints into separate modules, reducing runtime maintenance complexity (7e073a1)
• Improved structured tool result envelopes for LLMs with ToolSuccess/ToolError, schema validation repair, routing repair, and cache-hit markers, reducing invalid retries and wrong tool choices (7e073a1)
• Improved the Skill workflow: skills index no longer exposes guessable .beak-skills filesystem paths, tool events now include skill_name/skill_file_path/tool_display_name, and obsycli renders read_skill/read_skill_file more clearly (7e073a1)
• Improved profile cache sync with a profile cache file lock, stale lock cleanup, and skills index generation from installed version information, reducing inconsistent cache states during concurrent syncs (7e073a1)
• Improved Agent identity and Prompt rules. Self-introductions prefer runtime agent_name and avoid exposing the underlying LLM provider, model ID, internal project names, or code package names (7e073a1)
• Improved obsycli tool call rendering with tool_display_name, limited long tool reason body lines, and replay of recent tool output only when a working message currently exists (7e073a1)
• Improved obsycli task/session creation metadata by explicitly writing session_type=task and source_type=user, helping the server identify CLI user task sources (57fb449)
• Improved manual test scripts to support running multiple manual test instances in parallel, avoiding local runtime directory and state conflicts during multi-instance debugging (bd73f2e)
• Added tests for bug reports, tool loop stop reasons, structured tool envelopes, schema/routing repair, Skill reads, profile cache, LLM error classification, obsycli token accounting, and manual test scripts (7e073a1, 57fb449, bd73f2e)

v0.1.7(2026/06/10)¶

This release focuses on the obsycli interaction experience, beak-agent GTrace observability, Skill workflows, and Plan Mode stability. Key updates include:

New Features¶

• obsycli adds the /save command to download internal asset attachments from Agent replies to ~/.obsyclis/attachments/, with a hint when attachments are available (01d75d9)
• obsycli can render task_insight and session_summary directly in the conversation stream, making Agent interim insights and task close summaries easier to view (316ecf2, 01d75d9)
• Agent adds read_skill and read_skill_file tools to read the explicitly requested Skill SKILL.md and referenced files before following the Skill workflow (01d75d9)
• Agent supports user @ entities and explicit Skill request context. Message metadata can carry requested_skills and inject them into the Prompt, avoiding guesswork based only on Skill names (d460e73)
• Agent adds automatic task title generation. When obsycli creates an unnamed task, Agent can generate the title from the first user message (6b5618a)

Bug Fixes¶

• Fixed obsycli working status disappearing too early in some LLM turns. Normal messages that are sent successfully and have a message UUID now continue to track running status (01d75d9)
• Fixed unstable obsycli error prompts and tool output rendering by unifying user-visible error formatting and preserving complete single-line tool output, so key errors are not hidden by truncation (01d75d9)
• Fixed compatibility with legacy beak-agent configuration OTEL_EXPORTER_OTLP_PROTOCOL=grpc; it is now normalized to http/protobuf to avoid startup failures from old installs (8d663c1)
• Fixed repeated original user input in the LLM Chat span input_preview during multi-round tool loops. It now records the latest non-runtime-prompt message in the current LLM request (01d75d9)
• Fixed Plan Mode structured interaction schemas that could still rely on unstable action labels. accept/reject/skip/revise are now handled only by stable action IDs (c55f653, 01d75d9)

Improvements¶

• Improved beak-agent GTrace/OTEL export: OTLP HTTP/protobuf is now the default, OTEL_EXPORTER_OTLP_HEADERS is supported, and logs/traces include Agent identity, Plan Mode state, and LLM call context (01d75d9, 8d663c1)
• Improved Skill dependency installation by moving async dependency installs, timeout handling, process-group cleanup, and changed-package filtering into the cmd/agent runtime; Profile Client now focuses on sync and indexing (c4be7f0)
• Improved Agent tool and Prompt rules for attached session files, uploaded attachments, Skill file reads, working-directory writes, and progress_feedback usage (01d75d9, d460e73)
• Improved Plan Mode runtime and obsycli rendering, including direct-answer repair, structured interaction events, history replay, plan state restoration, and compact tool output rendering (c55f653)
• Improved obsycli task creation and listing. /newtask supports custom task names, and unnamed tasks can use automatic title generation (6b5618a)
• Split Agent Service and obsycli runtime code, and added tests for attachment saving, Task Insight, Skill dependency installation, GTrace/OTEL, Plan Mode, and session titles (316ecf2, c4be7f0, 01d75d9, 6b5618a)

v0.1.6(2026/06/03)¶

This release focuses on Plan Mode stability, the task close flow, and obsycli runtime display. Key updates include:

New Features¶

• obsycli adds the /close <number|task_id> command to close the current or specified task. Closed tasks enter the closing/completed flow and no longer accept new messages (4efd12a)
• Agent adds the session_summary event. When a task is closed, Agent generates a session summary and the server completes the session state transition (4efd12a)
• Agent adds the agent_runtime_result event. obsycli can display the tool loop stop reason, blockers, recommended next steps, and recently completed work (845d571)

Bug Fixes¶

• Fixed Plan Mode structured interactions that could fail to attach to the triggering user message during history replay. Plan events now fall back to reply_to_message_uuid to associate historical messages (845d571, 713593e)
• Fixed Plan Mode terminal states being rolled back by stale snapshots, preventing completed/interrupted states from being overwritten by outdated planning states (845d571)
• Fixed unstable Plan Mode interaction confirmation handling by consistently using accept/reject/skip/revise actions and requiring a final plan submission after requirement confirmation (713593e)
• Fixed the session status API missing closing support. Closing a task now publishes session_close_requested context so Agent can generate a close summary (4efd12a)

Improvements¶

• Improved the Plan Mode interaction schema by removing global free_text and using field-level custom_input/comment_input plus text_entry, improving structured input consistency across CLI and Web UI (713593e)
• Improved obsycli Plan Mode rendering by hiding intermediate plan_execution_progress noise, avoiding duplicated structured/text content, and correctly ending the current working state when plan events arrive (845d571, 713593e)
• Improved Agent Prompt and Prompt Builder Plan Mode rules with new interaction actions, text input fields, and session context constraints (713593e)
• Reorganized documentation and Plan Mode Web UI contracts, adding descriptions for /close, session_summary, agent_runtime_result, and the plan interaction schema (845d571, 713593e, 4efd12a)
• Added tests for Plan Mode, session close, runtime result, history replay, WSS status update, and obsycli /close (845d571, 713593e, 4efd12a)

v0.1.5(2026/06/01)¶

This release focuses on Agent MCP Client, Plan Mode, obsycli interaction capabilities, and permission/approval flow improvements. Key updates include:

New Features¶

• Agent adds MCP Client support, including syncing MCP configuration and tool packages from profile cache and calling MCP tools through mcp_list_servers, mcp_list_tools, and mcp_call_tool (fcf17e3, 14aad95)
• Agent adds session/task-level Plan Mode, supporting plan drafts, structured interactions, final plan approval, execution progress, blockers, interruption, and completion plan events (c69d29c, d65661d)
• obsycli adds /plan, /plan-approve, /plan-reject, /plan-revise, and /plan-interrupt, and can display plan status and structured interaction content (c69d29c, d65661d)
• obsycli adds /permission-full and /permission-standard to switch the current task permission mode in CLI (d65661d)

Bug Fixes¶

• Fixed obsycli /tasks and /attach not being limited to CLI-created tasks, avoiding accidental attachment to sessions not created by obsycli (29faa63)
• Fixed unknown obsycli startup subcommands not reporting errors promptly, avoiding silently ignored invalid commands (5d9f923)
• Fixed compatibility issues when obsycli reads recent history messages by adding fallback parsing for content.messages and content.data (5d9f923)
• Fixed high-risk commands such as remote login and database clients in bash not entering runtime approval checks (26edc7a)

Improvements¶

• Improved MCP tool exposure: stable routing tools are exposed by default, small toolset servers can be expanded into direct tools, and large toolsets continue through mcp_call_tool, reducing tool schema growth (fcf17e3, 14aad95)
• Improved Agent Prompt and prompt cache structure with MCP, Plan Mode, permission mode, and session context rules, and included MCP toolset version in cache key calculation (41d0018, fcf17e3, c69d29c)
• Increased the default query_session_events history query count from 3 to 20, improving multi-turn context recovery (26edc7a)
• Improved obsycli input experience with fuzzy slash command matching, highlighting, Ctrl+L clear screen, double Esc cancel, unified command history, and a cleaner startup prompt (26edc7a, c69d29c)
• Improved Agent tool execution by introducing composite tool executor and execution plan abstractions to handle tool policy, schema validation, approval, repair, and execution decisions consistently (fcf17e3, c69d29c)
• Added tests for MCP Client, Plan Mode, obsycli plan/permission commands, profile sync, tool execution, and approval flows (41d0018, fcf17e3, 14aad95, c69d29c, d65661d)

v0.1.4(2026/05/27)¶

This release focuses on simplifying the Agent Prompt, stabilizing obsycli history rendering, and splitting obsycli code structure. Key updates include:

Bug Fixes¶

• Fixed obsycli potentially syncing early historical answers after startup or reconnect, which could overwrite the current interaction area with an old message (6e05c87)
• Fixed obsycli potentially printing raw JSON for historical web_search_evidence and other agent events in the interactive UI. These events are still kept in the obsycli log for troubleshooting (6e05c87)

Improvements¶

• Simplified duplicate rules in the Agent core prompt and moved path scope, required arguments, and usage conditions into tool schemas/descriptions, reducing the default system prompt size (6e05c87)
• Strengthened the exec tool description to clarify that exec runs a direct program. Shell syntax, redirection, pipes, variable expansion, globbing, loops, and conditionals should use bash (6e05c87)
• Split the large obsycli chat.go file by responsibility, including startup, stream, agent event, state, session, render, tool/approval, and other areas, reducing maintenance cost (6e05c87)
• Added tests for obsycli stream historical message suppression, agent event logging, Prompt construction, and the exec tool description (6e05c87)

v0.1.3(2026/05/26)¶

This release focuses on agent runtime, obsycli rendering, and install/upgrade flow improvements. Key updates include:

New Features¶

• Agent UI schema overlay is now driven by scripts/prepare.py and scripts/compile.py; at runtime, Agent selects an overlay by skill and compiles agent-ui@1 draft blocks (5a1e699)
• submit_agent_ui_blocks supports the tool_input_schema provided by the overlay. Agent only submits draft blocks and no longer needs to submit overlay_id or layout (5a1e699)
• beak-agent adds the --default-envs flag to print runtime default environment variable comments, allowing install scripts to use defaults directly from the binary (823c0bb)
• Added AGENT_REASONING_DISPLAY_MODE with hidden/raw modes. Raw reasoning is hidden by default (823c0bb)

Bug Fixes¶

• Fixed unclear layering of tool-call running events: human-visible commentary is now written to metadata.commentary, while running event content remains empty to avoid client misinterpretation (c2cb6e6, 823c0bb)
• Fixed user-customized active environment variables in agent.env potentially being lost during Agent upgrades. Upgrade now preserves old KEY=value and export KEY=value entries not overridden by the new config (823c0bb, 5a1e699)
• Fixed Task Insight event ownership to ensure insights are persisted and published by reply-message relationship (5a1e699)

Improvements¶

• Improved obsycli thinking and tool-call rendering by folding and truncating long reasoning, tool descriptions, and output previews, reducing terminal noise (823c0bb)
• Improved Prompt rules so final reply language follows the original user task and avoids leaking runtime bookkeeping, repair prompts, tool protocol details, and other internal information (823c0bb)
• Improved Task Insight submission by allowing multiple interim facts, actions, hypotheses, and conclusions during a task, using runtime time to generate timestamps (5a1e699)
• Adjusted the manual upgrade entry point. Documentation now recommends using beak-agent-updater directly to trigger latest or specified-version upgrades (5a1e699)
• Added tests for Agent UI schema scripts, install/upgrade, tool-call events, reasoning display, and Task Insight (c2cb6e6, 823c0bb, 5a1e699)

v0.1.2(2026/05/25)¶

This release focuses on agent and obsycli improvements and does not include beak server changes. Key updates include:

New Features¶

• Agent adds tool-call visualization, showing tool-call reasons, batch execution status, and output previews in CLI, with Ctrl+T to switch to full tool output (4bd18ed)
• Agent adds Prompt Cache support, splitting the system prompt into stable prefix, semi-stable prefix, and dynamic suffix, and adding explicit cache block support to the OpenAI-compatible provider (c9274e7)
• The query_session_events tool adds on-demand querying of recent user conversations and tool execution history, reducing the amount of context injected by default (c9274e7)
• The obsycli session page adds task, Agent, input/output/cache Token, and remaining context status display (c9274e7)
• Added Agent Prompt structure and example documentation explaining cache layering, dynamic context, and tool-call conventions (c9274e7)

Bug Fixes¶

• Fixed Agent install-config API not returning llm_max_tokens, ensuring self-hosted Agent installs can pass the max-token configuration correctly (92e776d)
• Fixed unclear prompts from the self-hosted install script when receiving an error endpoint, HTML/JSON response, or empty response, adding HTTP status, response summary, and env file format validation (92e776d)

Improvements¶

• Improved /agents and /tasks display by hiding raw UUIDs while keeping internal numeric selection, improving CLI readability (c9274e7)
• Improved Agent final reply language rules: default to the latest user message language while preserving commands, code identifiers, and proper nouns in their original form (c9274e7)
• Improved LLM observability by adding Prompt Cached Tokens, cache hit ratio, and context usage to logs and trace attributes (c9274e7)
• Improved tool-call event metadata with batch and output summary fields for unified frontend and CLI rendering (4bd18ed)
• Added tests for self-hosted install scripts, Prompt construction, tool-call visualization, and configuration APIs (4bd18ed, 92e776d, c9274e7)