mirror of
https://github.com/coder/coder.git
synced 2026-10-31 11:02:50 +08:00
## Summary When the `mcp-tool-search` experiment is enabled, chatd stops inlining connected MCP tool schemas into every generation. It instead exposes a built-in `find_tools` tool whose description carries a compact catalog of the deferred tools, and only ships full JSON schemas for tools the model has activated by searching or by calling them directly. Closes [CODAGT-760](https://linear.app/coder/issue/CODAGT-760). ## Problem Tool-heavy agent configurations (GitHub, Linear, Notion, and dev-tooling MCP servers) inline over 100k tokens of tool schema definitions into every generation. Initial uncached requests reached ~216k tokens with time-to-first-token close to nine minutes, while the model typically invokes only a handful of tools per turn. ## How it works - `decideMCPToolSearch` defers external and workspace `.mcp.json` MCP tools whenever the experiment is enabled. Native, dynamic, provider, skill, and transport tools are never deferred. - `find_tools` embeds a server-grouped catalog in its tool description (degrading to names-only, then counts-only, then a constant-size summary past a context-scaled size cap) and scores keyword matches across tool names, descriptions, parameter schemas, and server metadata. Queries can scope to one server with a `server:` prefix, and exact `names` arguments always activate. - Activation state is ephemeral: it is re-derived each generation from surviving chat history (`find_tools` results and direct calls to deferred tools), so activations naturally lapse when compaction summarizes them away. Aggregate activated schema weight is capped at 10% of the context window, shedding the least recently activated schemas first; `find_tools` shares that budget across parallel calls in one step. No new persistence. - Deferred tools stay registered for execution, so the model can call a cataloged tool directly without searching first; the schema is activated for subsequent steps. - Fail-open: the experiment being disabled, an empty candidate set, or an MCP tool named `find_tools` all disable deferral, leaving today's behavior byte-identical on the wire. - Prometheus counters/histograms track `find_tools` calls, matches, activations, and deferred token weight. - The conversation timeline renders `find_tools` calls with a collapsed search summary and expandable match list, falling back to the generic renderer on malformed payloads. ## Validation - Unit tests for the catalog, matcher, experiment-gated decision, and activation derivation; end-to-end chatd generation tests covering search-then-call, direct-call activation, experiment-off wire parity, compaction lapse, and subagent tool gating. - Storybook interaction tests for the timeline rendering and malformed-payload fallback. - Remote dogfood UAT on dev.coder.com passed: deferral with a real MCP server and Anthropic model, direct calls without prior search, activation persistence across turns, experiment-off parity, and clean UI/console. > Disclosure: Mux (AI agent) authored this PR on Mike's behalf.
package chattool
import (
"bufio"
"strings"
"unicode"
"github.com/coder/coder/v2/coderd/render"
coderstrings "github.com/coder/coder/v2/coderd/util/strings"
)
// readmeInputMaxBytes caps how many README bytes are parsed so a giant README
// can't OOM coderd. It sits well above the output rune cap, so it never trims a
// real excerpt.
const readmeInputMaxBytes = 64 * 1024
// readmeText returns the README as bounded, frontmatter-stripped plain text
// truncated to maxRunes, or "" when the README is blank or conversion fails.
func readmeText(readme string, maxRunes int) string {
// Cap the parse input first (see readmeInputMaxBytes); goldmark and the
// tokenizer tolerate a mid-line or mid-rune cut.
bounded := readme[:min(len(readme), readmeInputMaxBytes)]
text, err := render.InnerTextFromMarkdown(stripReadmeFrontmatter(bounded))
if err != nil {
return ""
}
return coderstrings.Truncate(text, maxRunes, coderstrings.TruncateWithEllipsis)
}
// stripReadmeFrontmatter strips a leading frontmatter block from the README if it
// exists. An unterminated frontmatter block is treated as a regular body section.
// UTF-8 BOMs are stripped if present, and CRLF is normalized to LF. Leading
// whitespace is also stripped.
func stripReadmeFrontmatter(readme string) string {
trimmed := strings.TrimLeftFunc(readme, func(r rune) bool {
return unicode.IsSpace(r) || r == '\ufeff'
})
var out strings.Builder
scn := bufio.NewScanner(strings.NewReader(trimmed))
scn.Buffer(nil, readmeInputMaxBytes+1) // headroom
var lineNumber int
var fences int
for scn.Scan() {
line := scn.Text()
lineNumber++
// Only handle fences if we haven't already found an
// opening and closing fence.
if fences < 2 {
isFence := strings.TrimRight(line, " \t\r") == "---"
if isFence {
fences++
continue
} else if lineNumber == 1 {
// No leading fence -> no frontmatter. Return the entire document.
return trimmed
}
// We are still in the frontmatter block. Skip writing this line.
continue
}
_, _ = out.WriteString(line)
_, _ = out.WriteString("\n")
}
// Can err if input is too big. Shouldn't happen normally.
if scn.Err() != nil {
return trimmed
}
// Scanner did not scan any lines. No frontmatter to strip.
if lineNumber == 0 {
return trimmed
}
// If we are still fenced but we reached the end of the document
// we have an unterminated fence.
if fences == 1 {
return trimmed
}
return out.String()
}