feat(mcp): surface scheduled mcp_tool tasks before they break (#390) - #557
Merged
Conversation
Phase 2 of the MCP removal, part 1 of 2: find the tasks, do not rewrite them. `agent --type mcp_tool` is removed in v0.85.0. It is the one deprecated surface whose users are absent when it breaks: an interactive `tool call` warns on every invocation right up to removal, but a scheduled task was warned once -- when it was created -- and then runs unattended. At removal it simply starts failing on its next cron tick. - `kbagent doctor` gains an `mcp_tool_tasks` check listing each affected task, the tool it calls and the native command that replaces it. A tool the parity map does not know reports `native_command: null` rather than inventing one. Filesystem only -- no API call, no MCP spawn -- and skips when there is no agents.json. - `kbagent agent list` notes the affected tasks under the table and points at doctor. The marker is deliberately NOT in the Type cell: Rich truncated it to "DEPRECA…" there, which is worse than not flagging it. `--json` gains an additive per-task `deprecation` key, present only on affected tasks so every existing consumer sees a byte-identical payload. Deliberately NOT included: automatic rewriting. ParityEntry carries only the command name, no argument mapping, while mcp_tool params are the MCP tool's own `input` dict -- different key naming and, for several tools, different structure. Guessing the argv of a scheduled WRITE task that then runs unattended is the worst place to be wrong; a bad rewrite beats a clear failure in no scenario. Part 2 will emit a reviewable migration plan with explicit TODOs instead.
padak
force-pushed
the
feat/mcp-tool-task-detection
branch
from
August 11, 2026 23:02
0230a74 to
d5a3ff7
Compare
This was referenced Aug 12, 2026
padak
added a commit
that referenced
this pull request
Aug 12, 2026
…#390) (#581) Decision: `kbagent agent migrate-mcp-tasks` will not be built. Migrating a scheduled `--type mcp_tool` task before v0.85.0 is manual, or AI-assisted. The reasoning that ruled out automatic rewriting also argues against a plan-generator: the parity map knows which native command replaces each tool but not how to map arguments -- mcp_tool params are the MCP tool's own `input` dict (`componentId`) against CLI flags (`--component-id`), with different shapes for several tools. An agent reading `--help` next to the task's actual input does that mapping better than any generator we could ship, with a human confirming. Detection already landed in #557, which is the part that could not be improvised. gotchas.md now states plainly that no command exists and why, then gives the four-step recipe. keboola-expert.md gets a one-line trigger pointing at it, kept short because that file is at 61.6 KB of its 62 KB budget. Two corrections from review are folded in. The recipe originally ended with `agent update --type cli_command --argv ...`, which does not work -- that command cannot change a task's action, so the documented path exited 2. It now says create-new, verify with one `agent run`, delete-old. That path changes the task ID, and `Trigger.task_id` chains tasks by ID, so the recipe also says to repoint any chained task BEFORE deleting or the chain breaks silently. tests/test_mcp_migration_recipe.py pins the recipe to the CLI's real flags by reading `params` off the command object rather than scraping `--help` -- rendered help is wrapped and truncated by terminal width, so the first version of that test passed locally and failed on all three CI runners. Residual risk, stated rather than hidden: there is no telemetry, so "probably nobody uses this" is a judgement about the user base, not a measurement. If it is wrong the failure mode is a silently dead cron task -- which `kbagent doctor` now surfaces, and that is why the decision holds without the command.
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.
Phase 2 of the MCP removal, part 1 of 2: find the tasks, do not rewrite them.
agent --type mcp_toolis removed in v0.85.0. It is the one deprecated surfacewhose users are absent when it breaks: an interactive
tool callwarns onevery invocation right up to removal, but a scheduled task was warned once --
when it was created -- and then runs unattended. At removal it simply starts
failing on its next cron tick.
kbagent doctorgains anmcp_tool_taskscheck listing each affected task,the tool it calls and the native command that replaces it. A tool the parity
map does not know reports
native_command: nullrather than inventing one.Filesystem only -- no API call, no MCP spawn -- and skips when there is no
agents.json.
kbagent agent listnotes the affected tasks under the table and points atdoctor. The marker is deliberately NOT in the Type cell: Rich truncated it to
"DEPRECA…" there, which is worse than not flagging it.
--jsongains anadditive per-task
deprecationkey, present only on affected tasks so everyexisting consumer sees a byte-identical payload.
Deliberately NOT included: automatic rewriting. ParityEntry carries only the
command name, no argument mapping, while mcp_tool params are the MCP tool's own
inputdict -- different key naming and, for several tools, differentstructure. Guessing the argv of a scheduled WRITE task that then runs
unattended is the worst place to be wrong; a bad rewrite beats a clear failure
in no scenario. Part 2 will emit a reviewable migration plan with explicit
TODOs instead.