Repository navigation
Add the Parsar Agent workspace with Web and Feishu on the public Agents API - #700
Merged
Merged
Conversation
SaladDay
changed the title
This comment has been minimized.
This comment has been minimized.
P0 keeps two entry points: the Web chat and Feishu bots. Agents are the reusable thing people talk to, and each conversation is one public Session. - Remove tasks, projects, schedules, the external MCP endpoint, frontend extensions and personal OAuth connections. - Agents require an environment template; Duplicate replaces templates. - A Feishu bot binds one Agent and is its identity. Environment templates map bot fields to variables, and the Feishu case configures lark-cli with them so the CLI mints and refreshes tenant tokens itself. - Feishu creates Sessions directly: one per private sender until /new, one shared conversation per group topic. Group chat is a switch; joining a group is enough. Earlier group messages are never read. - Feishu progress cards show the current step and keep thinking and the step record in collapsed panels with success and failure colors. - Redesign the Web UI as a chat product with readable type and spacing. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
- MiniMax Code now uses the selected Provider and supports HTTP MCP and functions, so Parsar drops its MiniMax-only restrictions. - Use randomUuid from @oac/agents-client everywhere so IDs work without a secure context. - Keep main's docs layout and describe the Web chat and Feishu product. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Core now passes official execution options through to each Harness, so Agents can turn on live web search (Codex and Claude Code) and choose a reasoning effort. Searches show as readable steps on the Web and in Feishu records. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
- Save on the Agent page points to the missing required field instead of staying silently disabled; required model and environment come first. - Unsaved Agent edits are restored after navigating away in the same tab. - People can keep typing while the Agent works; the message reaches the running task. - Hide the turn rail in short conversations, use one disclosure style for step groups, mark the selected Agent on the home page and make template placeholders read as examples. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
An execution profile is the cloud counterpart of a local Codex profile: Harness, model, reasoning effort and summary, web search, verbosity and service tier. Agents reference one profile and one environment template; conversations never override either. - One shared declaration lists the values each Harness admits; the server validates against it and the editor offers only those values, resetting any the newly selected Harness does not accept. - The resource library groups building blocks by what uses them, and removal is disabled while something still references an item. - Feishu's case adds profile.json, and configure.mjs creates the profile. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Measured on a hosted Core, a new Feishu workspace conversation waited about 25 seconds for setup before its first reply; a Session created without input is prepared by Core immediately. With the template switch on, Parsar keeps one such Session per Agent and hands it to the next Web or Feishu conversation with its first message. - The match is a hash of the frozen request without input and metadata, so any change to the Agent, profile, template, bot or provider replaces it. - Taking a prepared Session refills it; uncertain first-input delivery keeps the conversation's key and retries the same input. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
OAC_EXAMPLE_HOST selects one IPv4 address of the machine. It is both the listen address and the only accepted Host, so DNS-rebinding protection stays exact; loopback remains the default. Parsar has no sign-in, which the README states next to the setting. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Inside a topic the bot takes part in, replies without a mention continue the conversation. A reply that mentions only other people is meant for them, so the bot no longer answers it; mentioning the bot as well still reaches it. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Agents allowed at most eight Skills, eight MCP servers and eight instruction fragments, Plugins and knowledge items combined. None of this came from Core, which admits 50 Skills, 50 Plugins and 50 initial files per Environment. Parsar now applies exactly those limits from one shared definition, counts knowledge with the template's own files, and leaves instruction fragments and MCP servers uncounted. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Execution profiles added a template entity that Agents only referenced: three Agents shared two profiles, conversations cannot override them, and reading an Agent's model took an extra hop. Copying an Agent already covers reuse, as for every other Agent setting. The Harness, model and options return to the Agent, keeping the same Harness-declared option set, the reset notice on engine changes and the 执行与环境 section. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
On mx, three of seven Feishu workspace conversations failed while initializing and two turns waited about 180 s for a first output. Parsar now keeps a ledger of every Web and Feishu turn and a monitor that only acts while nothing has run: a failed Environment is replaced once with the identical frozen request, a turn silent for 45 s is cancelled and resent once, and a turn failing before output is resent once. Replaced turns stay hidden in Web and unreported in Feishu. - The Web composer sends through Parsar so its turns are watched too. - Users see the stage, waiting time and usual first-reply time, and final failures in plain language with a retry. - Models get a one-shot test, environment templates a real trial run, Agents seven-day health, and a page lists all conversation outputs. - The Feishu bot ignores messages from external tenants. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Core now reuses an identical earlier preparation for up to 24 hours and keeps only /environment. The Feishu workspace's setup wrote lark-cli's configuration to HOME and its encrypted credentials to ~/.local/share, so every reused preparation started with an unconfigured lark-cli; the same happened after a conversation moved to a new sandbox. The preset and the Feishu example now point LARKSUITE_CLI_CONFIG_DIR and LARKSUITE_CLI_DATA_DIR at /workspace/.lark-cli, and the README states that setup must keep what it needs under /workspace. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
- The Feishu workspace preset is the shipped example: the editor imports examples/feishu/environment.json and setup.sh instead of a second copy. - Web and Feishu cards read the waiting copy from shared/turn-status.mjs. - Every follow-up, from Web, Feishu or a manual retry, is recorded and sent through one ledger path that keeps its baseline across send retries, and an input Core never admitted settles instead of staying "running". - Environment trial runs move out of Session creation into trials.mjs. - Resource actions are one route table; the frozen request is hidden only at the response boundary; one helper downloads Artifacts. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
The first input of a claimed prepared Session was sent before Parsar recorded it, so while it waited for a free sandbox slot neither Web nor Feishu showed a stage and the monitor could not see it. It now takes the same record-then-send path as every follow-up, and any input that Core never admits settles once Core's five-minute reservation has expired. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
The monitor took the newest Core Turn after an entry's baseline. When a follow-up arrived before the first input was settled, both entries followed the follow-up's Turn and the first one was settled with its outcome. An input's Turn is the oldest one after its baseline, because later inputs only start later Turns. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Core answers input to an idle Session only once its Turn starts. Claiming a prepared Session awaited that answer, so a prepared Session that never started (one stayed "transitioning" on mx for 26 minutes) kept the new conversation spinning for five minutes and then left it without a reply. Claiming now records the Turn and sends the input in the background, and keeps the frozen request; when the input cannot be sent or the Turn has not started within 45 seconds, the monitor creates the Session normally through the same path that replaces a failed Environment. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Finding a stuck or slow turn meant opening conversations one by one or reading Core logs. The 运行 view now reads the Turn ledger: running turns from Web and Feishu with their stage, waiting time and the Agent's usual first-reply time, recently finished turns with duration, automatic recoveries and failure, and counts of running, waiting-for-you and recent failures with a five-minute concurrency line. Running turns can be stopped and a conversation's latest failed turn retried from there. It replaces the per-conversation run list that polled Core for every Session's Turns. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
The 运行 view was reachable only through 全部对话. It now has its own entry below 新对话 that shows how many conversations are running, read from the same Turn overview query as the page. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
SaladDay
changed the title
SaladDay
marked this pull request as ready for review
October 25, 2026 13:24
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
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.
Summary
Parsar is an optional example product, started independently of Core, that reaches Core only through the public
/v1Agents API and a dedicated Project key. It turns reusable resources into conversational Agents available on the Web and in Feishu./workspace, as Core's preparation reuse requires./new; each group topic is one shared Session. Topic replies addressed only to other people,@alland external tenants are ignored. Feishu runs with the bot's application identity.cli.mjsis a dependency-free JSON client of the same/appand/v1routes the UI uses; the manager example assembles an ordinary Agent with it, optionally behind an authenticated HTTPS gateway (Caddy example).@oac/agents-clientprojects reasoning summary stream events instead of passing them through as unknown events.Core is unchanged: no privileged endpoint, no shared tables, no Core key in the product.
Test plan
make check-example: typecheck, 83 server tests, build, 25 browser testsmake check-web-unit:@oac/agents-client742 tests, console 344 tests, console buildmake check-docs check-ciGenerated with Devin