可配置CodeGraph启动路径和参数，修复在会话中显示大段内容

lib/index.js

@@ -20,6 +20,12 @@
 // <codegraph_context> — the agent's grep/read reflex has nothing left to
 // find. Every failure path is a silent no-op.
 //
+// The steered message keeps `role: 'user'` (so the model reads it as prompt
+// context) but MUST NOT use `source.kind: 'user'` — that kind means "a human
+// typed this" and the Chat UI prints it as a full user bubble. It carries
+// `plugin:dsh-codegraph` instead, which the UI files as a collapsed
+// injected-context row. See FRONTLOAD_SOURCE_KIND.
+//
 // Why a native plugin instead of a plain MCP server: codegraph's MCP server
 // exposes ZERO tools while the workspace is unindexed (and tells the agent it
 // should not index itself). This plugin always exposes the tools, including
@@ -62,8 +68,11 @@
 // in the wild: intermittent "neither subprocess nor shell mounted" boot
 // crashes). Lazy resolution also lets the tools start working if an executor
 // mounts after this plugin. The `codegraph` executable must be on PATH (e.g.
-// global `npm i -g @colbymchenry/codegraph` or the npm thin shim); a clearer
-// error is thrown when it is missing.
+// global `npm i -g @colbymchenry/codegraph` or the npm thin shim), or its
+// full path given via the `codegraphPath` config option (cordis.patch.yml),
+// with optional extra CLI arguments via `codegraphArgs` (an array of strings,
+// prepended to every invocation); a clearer error is thrown when it is
+// missing.
 
 import { existsSync } from 'node:fs'
 import { dirname, join } from 'node:path'
@@ -91,10 +100,20 @@ export const inject = ['tools', 'systemPrompt']
 //   steers the result into the turn as <codegraph_context> — so the agent's
 //   grep/read reflex has nothing left to find. Every failure path is a
 //   silent no-op; it never breaks or blocks the user's prompt.
+// - codegraphPath (default ''): full path to the `codegraph` executable. When
+//   set (e.g. in cordis.patch.yml), the tools invoke that executable directly
+//   (both the subprocess and shell execution paths). When empty, the
+//   executable is resolved from the system PATH as before.
+// - codegraphArgs (default []): extra CLI arguments passed to the codegraph
+//   executable on every invocation, e.g. ['--no-color', '--threads', '4'].
+//   Only meaningful together with `codegraphPath` (with a PATH-resolved
+//   executable, pass such flags through the tool arguments instead).
 export const Config = z.object({
   guideSearch: z.boolean().default(true),
   surface: z.union([z.const('core'), z.const('full')]).default('core'),
-  frontload: z.boolean().default(true)
+  frontload: z.boolean().default(true),
+  codegraphPath: z.string().default(''),
+  codegraphArgs: z.array(z.string()).default([])
 })
 
 // System-prompt guidance that makes the codegraph tools preferred for code
@@ -137,7 +156,7 @@ function shQuote(value) {
 }
 
 function executableHint() {
-  return '`codegraph` was not found on PATH. Install it once, e.g. `npm i -g @colbymchenry/codegraph` (or use the official install.sh / npm thin shim), then retry.'
+  return '`codegraph` was not found on PATH. Install it once, e.g. `npm i -g @colbymchenry/codegraph` (or use the official install.sh / npm thin shim), or set the `codegraphPath` option in cordis.patch.yml to the full path of the codegraph executable, then retry.'
 }
 
 // The core tool surface (see Config.surface): the index bootstrap/maintenance
@@ -157,6 +176,42 @@ const CORE_TOOLS = new Set([
 // own output.
 const FRONTLOAD_MARKER = '<codegraph_context'
 
+// LOAD-BEARING: the durable `source.kind` of an injected message decides how
+// the Chat UI files it, and `'user'` is the wrong answer.
+//
+// The client projects every durable `user/message` to a Chat node by
+// `source.kind` (see @deepseek-ai/dsh-client-ui-chat):
+//   kind !== 'user'  -> a "context" node: a COLLAPSED one-line row
+//                       (icon + "injected context" + producer label), expanded
+//                       only on demand, and folded into the turn's process
+//                       disclosure. `source.form`/`source.summary` only shape
+//                       how that row reads.
+//   kind === 'user'  -> a real user/steering BUBBLE: the message body is
+//                       printed in full in the transcript.
+//
+// So an injection that claims to be a user message renders as if the human had
+// pasted the whole 12 KB explore dump — the exact complaint this fixes. Every
+// DSH-native injection uses a non-'user' kind (`plugin:<name>`,
+// `runtime-context`, `skill-catalog`, `agent-instructions`, `subagent-settled`,
+// `tool-jobs`, ...), and so do the other plugins
+// (`plugin:graph-memory`, `plugin:dsh-mcp-manager`). This plugin must too.
+const FRONTLOAD_SOURCE_KIND = `plugin:${name}`
+
+// The `notice` form gives the collapsed row a one-line account of what was
+// injected, so the reader can tell an explore dump from any other injection
+// without expanding it. The 120-character bound mirrors DSH's own
+// CONTEXT_SUMMARY_MAX_CHARS, clamped here rather than imported so the plugin
+// keeps working on host versions that predate that export.
+const SUMMARY_MAX_CHARS = 120
+function summarizeExplore(body) {
+  const found = /Found (\d+) symbols? across (\d+) files?/i.exec(body)
+  if (found) return `explore: ${found[1]} symbols in ${found[2]} files`
+  const firstLine = body.split('\n').map((line) => line.trim()).find((line) => line.length > 0)
+  const summary = firstLine ? firstLine.replace(/[*#>]/g, '').trim() : 'explore result'
+  if (!summary) return 'explore result'
+  return summary.length <= SUMMARY_MAX_CHARS ? summary : summary.slice(0, SUMMARY_MAX_CHARS - 1) + '…'
+}
+
 // Structural keywords (EN + zh) signalling a code-structure/flow question —
 // the HIGH-confidence gate tier: fire explore without further verification.
 const STRUCTURAL_KEYWORDS = [
@@ -201,7 +256,7 @@ function findIndexRoot(cwd) {
 
 export function apply(ctx, config = {}) {
   const tools = ctx.tools
-  const { guideSearch = true, surface = 'core', frontload = true } = config
+  const { guideSearch = true, surface = 'core', frontload = true, codegraphPath = '', codegraphArgs = [] } = config
   const toolEnabled = (name) => surface === 'full' || CORE_TOOLS.has(name)
 
   // Make the codegraph_* tools the preferred route for code search: instruct
@@ -242,14 +297,25 @@ export function apply(ctx, config = {}) {
     }
     if (sub !== undefined) {
       if (resolvedExe.value === null) {
-        try {
-          resolvedExe.value = await sub.resolveExecutable('codegraph')
-        } catch {
-          throw new Error('dsh-codegraph: ' + executableHint())
+        if (codegraphPath) {
+          // Explicit full path from config (cordis.patch.yml) — use it as-is.
+          if (!existsSync(codegraphPath)) {
+            throw new Error(
+              `dsh-codegraph: configured codegraphPath "${codegraphPath}" does not exist. ` +
+              'Fix the path in cordis.patch.yml, or remove the option to resolve `codegraph` from PATH.'
+            )
+          }
+          resolvedExe.value = codegraphPath
+        } else {
+          try {
+            resolvedExe.value = await sub.resolveExecutable('codegraph')
+          } catch {
+            throw new Error('dsh-codegraph: ' + executableHint())
+          }
         }
       }
       const proc = sub.spawn({
-        argv: [resolvedExe.value].concat(argv),
+        argv: [resolvedExe.value].concat(codegraphArgs).concat(argv),
         cwd: cwd || '/',
         stdio: {
           stdin: 'ignore',
@@ -270,7 +336,8 @@ export function apply(ctx, config = {}) {
       }
       return text
     }
-    const command = 'codegraph ' + argv.map(shQuote).join(' ')
+    const executable = codegraphPath || 'codegraph'
+    const command = executable + ' ' + [...codegraphArgs, ...argv].map(shQuote).join(' ')
     const spec = shell.resolve({
       command,
       workdir: cwd || '/',
@@ -622,7 +689,15 @@ export function apply(ctx, config = {}) {
           type: 'text',
           text: `${FRONTLOAD_MARKER} note="Structural context from CodeGraph for this prompt — treat the returned source as already read; call codegraph_explore for more.">\n${body}\n</codegraph_context>`
         }],
-        source: { kind: 'user' }
+        // NOT `{ kind: 'user' }`: a user-role injection must still carry a
+        // non-'user' source kind, or the Chat UI renders it as a human bubble
+        // instead of a collapsed injected-context row. See FRONTLOAD_SOURCE_KIND.
+        source: {
+          kind: FRONTLOAD_SOURCE_KIND,
+          plugin: name,
+          form: 'notice',
+          summary: summarizeExplore(body0)
+        }
       }))
     }