Repository navigation
Show what tools did in the REPL: parallel call count, Ctrl+O arguments, MCP errors and stderr - #151
Merged
Merged
Conversation
The stream loop read only the first FunctionCallContent and the first FunctionResultContent in each chunk. A model calling tools in parallel sends them all in one chunk, so the footer said "1 tool" for six calls, and only the first of several parallel writes reached the file-change list, the failure count and step verification.
The viewer showed a call's result but not what it was asked to do, so checking the SQL an MCP tool ran, or the parameters it bound, meant digging through the session file. Each kept output now has the call's arguments above its result. An MCP tool's result is shown as its text content rather than the CallToolResult JSON around it, and one with isError set counts as a failed call: it gets the red line under the call and counts toward the consecutive-failure cutoff, as the built-in tools' [ERROR] results do.
The wizard had no way to set Env, so a server that reads a password from
the environment (sql-mcp's PasswordEnvVar, for one) started without it
unless fuseraft itself happened to have the variable. It now asks for
KEY=VALUE pairs, and Env values expand ${ENV_VAR} tokens as Url and
Headers already do, so a secret need not be written into the saved
config.
The wizard also says which transport and authentication were picked;
the full-screen REPL showed only the question.
A stdio server's stderr is where it logs, and its errors often say to look there: sql-mcp refuses a database with "The server log lists why." The REPL dropped those lines, so the reason couldn't be read without running the server by hand. The manager now keeps each stdio server's last 200 stderr lines, kept even when the connect fails, and /mcp log <name> prints them and keeps them for Ctrl+O.
…ries Memory extraction saved what the session had read: a table of every employee's salary from a test database, and the columns of each table. The prompt now says not to save tool-retrieved data, personal information about anyone but the user, or secrets, nor what a future session can cheaply look up again.
While /mcp add or any other question had the panel, the status line kept spinning "working 15s" with "Enter adds to the turn", though nothing was running and Enter answered the question. The spinner and the keys are now left out until the panel closes.
Saving memories at /exit takes a model call, and the status line kept saying "Enter adds to the turn", so what was typed meanwhile was echoed as "(sent when this ends)" and then dropped with the session. The status line now says the session is ending, and input is refused with a note instead of being queued.
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.
Found by driving the full-screen REPL against sql-mcp and four live databases (Postgres, MySQL, SQL Server, Oracle). The SQL work went fine. Checking what the agent actually did took the session file, and a few TUI states were misleading.
Changes
FunctionCallContentandFunctionResultContentin each chunk. A model that calls tools in parallel sends them all in one chunk, so six calls showed as "1 tool". Only the first of several parallel writes reached the file-change list, the failure count and step verification.CallToolResultJSON around it. A result withisErrorcounts as a failed call: it gets the red line and counts toward the consecutive-failure cutoff. A notice the loop guard appends after the JSON is kept, and doesn't stop the result from being recognized./mcp addasks for a stdio server's environment (KEY=VALUE).Envvalues now expand${ENV_VAR}, asUrlandHeadersalready do, so a secret need not be saved in the config. The wizard also says which transport and authentication were picked./mcp log <name>shows the last 200 lines a stdio server wrote to stderr, which is where servers log. They're kept even when the connect fails. Before, those lines were dropped, though errors like sql-mcp's "The server log lists why" point there./exitit had saved a test table of every employee's salary, and each table's columns. The prompt now rules out tool-retrieved data, personal details of anyone but the user, secrets, and what can be cheaply looked up again./exit, it says the session is ending, and input is refused with a note. Before, input was echoed as "(sent when this ends)" and then dropped.Tests
ParallelCalls_AreAllCountedAndReported(fails before the fix),McpToolResultTests,McpSessionManagerStderrTests(stderr kept,${ENV_VAR}expanded inEnv),ReplCommandsMcpTests,StatusViewTests, and viewer cases inTuiRendererTests./mcp logshows sql-mcp's reason for refusing a login; no memories saved after the SQL session; typing during/exitgets the note.McpSessionManagerStderrTestshas acmd.exebranch for Windows that CI (Linux) doesn't run.