dsh-gitbash-shell 0.34.3 → 0.35.1

package/src/inline-tools.js

@@ -0,0 +1,434 @@
+/**
+ * dsh-gitbash-shell/inline-tools — runtime shell-tool patching for the OFFICIAL
+ * presets on Windows (v0.35.0).
+ *
+ * WHY THIS EXISTS. The official preset rows compose their tool set at boot
+ * (`tool-bash` disabled on win32, `tool-pwsh` enabled), and the patch layer
+ * CANNOT reach them: `applyEntryPatches`' id map only descends into rows that
+ * carry `group: true` (vendor/include/src/index.ts buildMap), while a preset
+ * declaration is a plain row whose `config.plugins` is opaque data. Rewriting
+ * the declaration's whole `config` would freeze the official row list at our
+ * restatement instant — every future official row addition would be invisible
+ * until we hand-synced. So the fix runs at the tool layer instead, where the
+ * official contract explicitly invites per-agent variants:
+ *
+ *   - the tools registry REJECTS same-scope duplicate names but points at this
+ *     exact path for variants ("for a per-agent variant, register through that
+ *     agent's `agent.ctx` instead", packages/core/tools/src/index.ts);
+ *   - `agent/created` fires once per agent with the composed tool face
+ *     reachable through `ctx.get('tools').schemas(agent)`.
+ *
+ * One reconcile per agent reads the face and acts on three facts:
+ *
+ *   1. `pwsh` present + `bash` absent  -> activate the OFFICIAL
+ *      @deepseek-ai/dsh-tool-bash module inside an agent-owned scope (full
+ *      official behaviour: jobs, approval escalation, rendering — zero
+ *      behaviour is re-implemented here) and deny the fake `pwsh` (its
+ *      executor is ctx.shell, which on this host IS Git Bash — a pwsh tool
+ *      that runs bash is a lie to the model);
+ *   2. `ops_*` tools present (dsh-ops installed) -> our FastCtx bridge and
+ *      gitbash_pwsh stand down; the official bash activation still runs, the
+ *      two command surfaces are complements, not duplicates;
+ *   3. `bash` already present (variant presets on, or a future official
+ *      change) -> touch nothing, so the variantPresets switch and this path
+ *      are mutually exclusive BY OBSERVATION, with no explicit lock.
+ *
+ * Everything here is Windows-only by caller gate; on other hosts the module
+ * loads but mounts nothing.
+ */
+
+/** Answer for one observed tool face. Pure. */
+export function inlineNeeds(names) {
+  const set = new Set(Array.isArray(names) ? names : [])
+  const has = (name) => set.has(name)
+  return {
+    // A pwsh tool without a bash tool is the official win32 composition on
+    // this host: bash must be patched in, and the pwsh tool would execute
+    // Git Bash through ctx.shell (a lying name) — deny it.
+    needsBash: has('pwsh') && !has('bash'),
+    denyPwsh: has('pwsh') && !has('bash'),
+    // dsh-ops publishes file/search tools under the ops_ prefix. When they are
+    // here, our FastCtx bridge and pwsh replacement would be duplicates.
+    opsPresent: has('ops_inspect_local_file') || has('ops_grep') || has('ops_bash'),
+  }
+}
+
+/**
+ * The routing addendum appended to the `bash` tool's description through the
+ * system-prompt assemble waterfall. ALWAYS applied — it states shell routing,
+ * which holds regardless of the path dialect a delegate sees.
+ */
+export const GITBASH_ROUTE_ADDENDUM = [
+  'Prefer this bash tool for ALL commands; use a PowerShell tool ONLY when a Windows-native capability (cmdlets, registry, services, WMI) is truly required.',
+  'Fix failed bash commands in bash — never switch shells mid-task, never embed `pwsh -Command` inside a bash command (variable expansion and output encoding break there).',
+].join('\n')
+
+/**
+ * The dialect addendum (MSYS facts) — appended ONLY while this plugin's path
+ * dialect is on for the assembling agent, because it describes the /c/ world
+ * the dialect creates.
+ */
+export const GITBASH_MSYS_ADDENDUM = 'This bash is Git for Windows (MSYS): paths use drive roots (/c/Users/...), and ~, /tmp, /dev/null and /usr/... resolve exactly as Git Bash resolves them.'
+
+/** @deprecated composed form kept for tests: route + MSYS sentences. */
+export const GITBASH_BASH_ADDENDUM = GITBASH_ROUTE_ADDENDUM + '\n' + GITBASH_MSYS_ADDENDUM
+
+/** The pwsh-replacement tool's model-facing description. */
+export function gitbashPwshDescription() {
+  return 'Run a command in Windows PowerShell 7 (pwsh). Use ONLY when a Windows-native capability '
+    + '(cmdlets, registry, services, WMI, scheduled tasks) is required — for everything else prefer the bash tool. '
+    + 'Each call uses a fresh shell; set workdir explicitly. Paths inside command must be Windows form (C:/... or C:\\\\...). '
+    + 'For scripts longer than about 1 KB, or containing quotes/backslashes/non-ASCII text, write the script to a file first '
+    + 'and pass script_path instead. Output is UTF-8; long output is truncated with the full text saved to a reported file. '
+    + 'Fix failed pwsh commands here; do not switch shells or mix PowerShell syntax into bash commands.'
+}
+
+/**
+ * Resolve the host's own PowerShell 7 — never a bundled payload, never
+ * Windows PowerShell 5.1. Mirrors the probe order the official pwsh resolver
+ * documents (well-known install locations, then PATH), WITHOUT its 5.1
+ * fallback: if no pwsh 7 exists the caller simply does not publish the tool.
+ * Pure I/O via the injected `exists` so the candidate matrix is testable.
+ * @returns {string} an absolute pwsh.exe path, or '' when absent.
+ */
+export function resolveHostPwsh7({ env = process.env, exists, platform = process.platform } = {}) {
+  if (platform !== 'win32') return ''
+  const programFiles = env.ProgramFiles ?? env.PROGRAMFILES ?? 'C:\\Program Files'
+  const programFilesX86 = env['ProgramFiles(x86)'] ?? 'C:\\Program Files (x86)'
+  const localAppData = env.LOCALAPPDATA ?? env.LocalAppData ?? ''
+  const userProfile = env.USERPROFILE ?? ''
+  const candidates = [
+    `${programFiles}\\PowerShell\\7\\pwsh.exe`,
+    `${programFiles}\\PowerShell\\7-preview\\pwsh.exe`,
+    `${programFilesX86}\\PowerShell\\7\\pwsh.exe`,
+    localAppData === '' ? '' : `${localAppData}\\Microsoft\\WindowsApps\\pwsh.exe`,
+    userProfile === '' ? '' : `${userProfile}\\scoop\\shims\\pwsh.exe`,
+    'C:\\ProgramData\\chocolatey\\bin\\pwsh.exe',
+  ].filter((entry) => entry !== '')
+  for (const candidate of candidates) {
+    if (exists(candidate)) return candidate
+  }
+  // PATH last: a PATH hit may be a store alias or shim, so it ranks below the
+  // well-known installs but still beats "no tool at all".
+  const pathValue = env.PATH ?? env.Path ?? ''
+  for (const dir of pathValue.split(';')) {
+    const clean = dir.trim().replace(/^"(.*)"$/, '$1')
+    if (clean === '') continue
+    const candidate = `${clean.replace(/[\\/]+$/, '')}\\pwsh.exe`
+    if (exists(candidate)) return candidate
+  }
+  return ''
+}
+
+/** Arguments `gitbash_pwsh` accepts — strict: anything else is rejected. */
+export const PWSH_ARGUMENTS = Object.freeze(['command', 'script_path', 'workdir', 'timeoutMs'])
+
+const PWSH_DEFAULT_TIMEOUT_MS = 120_000
+const PWSH_MAX_TIMEOUT_MS = 600_000
+
+/**
+ * Build the argv for one pwsh invocation. `command` and `script_path` are
+ * exclusive; a script path keeps the script's bytes OUT of the tool-call JSON
+ * entirely (the whole point of the dsh-ops issue #7 lesson: long heredoc-heavy
+ * commands inside one JSON string raise the malformed-tool-call rate and burn
+ * a whole model turn). Pure.
+ */
+export function pwshArgv(pwsh, args, { msysToWin } = {}) {
+  const hasCommand = typeof args.command === 'string' && args.command.trim() !== ''
+  const hasScript = typeof args.script_path === 'string' && args.script_path.trim() !== ''
+  if (hasCommand === hasScript) return { error: 'provide exactly one of command or script_path' }
+  if (hasScript) {
+    const raw = /** @type {string} */ (args.script_path)
+    // The model speaks MSYS (our prompt dialect); pwsh needs a Windows path.
+    const script = msysToWin ? msysToWin(raw) : raw
+    return { argv: [pwsh, '-NoLogo', '-NoProfile', '-NonInteractive', '-File', script] }
+  }
+  return { argv: [pwsh, '-NoLogo', '-NoProfile', '-NonInteractive', '-Command', /** @type {string} */ (args.command)] }
+}
+
+/**
+ * Validate one gitbash_pwsh call's arguments. Strict like ops_bash: an
+ * undeclared key is a hard error (the host bash/pwsh tools accept
+ * `run_in_background`; this tool does not, and a silent ignore would leave a
+ * server occupying the call until timeout). Pure.
+ * @returns {string|undefined} an error message, or undefined when valid.
+ */
+export function validatePwshArgs(args) {
+  if (args === undefined || args === null || typeof args !== 'object' || Array.isArray(args)) {
+    return 'invalid arguments: expected an object'
+  }
+  const unknown = Object.keys(args).filter((key) => !PWSH_ARGUMENTS.includes(key))
+  if (unknown.length > 0) {
+    return `invalid arguments: unknown key(s) ${unknown.map((key) => JSON.stringify(key)).join(', ')}; this tool accepts ${PWSH_ARGUMENTS.join(', ')}`
+  }
+  const hasCommand = typeof args.command === 'string' && args.command.trim() !== ''
+  const hasScript = typeof args.script_path === 'string' && args.script_path.trim() !== ''
+  if (hasCommand === hasScript) return 'invalid pwsh input: provide exactly one of command or script_path'
+  if (args.timeoutMs !== undefined && (!Number.isFinite(args.timeoutMs) || args.timeoutMs <= 0)) {
+    return 'invalid timeoutMs: expected a positive number'
+  }
+  if (args.timeoutMs !== undefined && /** @type {number} */ (args.timeoutMs) > PWSH_MAX_TIMEOUT_MS) {
+    return `invalid timeoutMs: must be no greater than ${PWSH_MAX_TIMEOUT_MS}`
+  }
+  return undefined
+}
+
+/**
+ * Resolve the OFFICIAL @deepseek-ai/dsh-tool-bash module entry from the
+ * host's own resolution base (the same seam probeHostExtras uses). Returns
+ * the absolute entry URL, or '' when the host does not ship the package.
+ */
+export async function resolveToolBashEntry(ctx) {
+  try {
+    const { createRequire } = await import('node:module')
+    const { pathToFileURL } = await import('node:url')
+    const req = createRequire(ctx?.baseUrl ?? import.meta.url)
+    return pathToFileURL(req.resolve('@deepseek-ai/dsh-tool-bash')).href
+  } catch {
+    return ''
+  }
+}
+
+/**
+ * Activate the OFFICIAL tool-bash inside one agent-owned scope: a child scope
+ * of the agent's context keeps every registration (tool, prompt section,
+ * effect) revertible at agent disposal, and the scoped-layer tool shadows the
+ * (absent) global one exactly the way the registry's own error message
+ * prescribes for per-agent variants. The official module is loaded ONCE per
+ * process; each agent gets a fresh scope and a fresh apply().
+ */
+export async function activateOfficialBash(agent, toolBashHref, createScopeFn, log) {
+  let mod
+  try {
+    mod = await import(toolBashHref)
+  } catch (error) {
+    log(`official tool-bash module import failed: ${error?.message ?? error}`)
+    return undefined
+  }
+  if (!mod || typeof mod.apply !== 'function') {
+    log('official tool-bash module has no apply(); activation skipped')
+    return undefined
+  }
+  const scope = createScopeFn(agent.ctx)
+  try {
+    await mod.apply(scope.ctx, {})
+    return scope
+  } catch (error) {
+    try { await scope.dispose() } catch { /* best effort */ }
+    log(`official tool-bash activation failed: ${error?.message ?? error}`)
+    return undefined
+  }
+}
+
+/**
+ * Mount the per-agent reconcile. `deps`:
+ *   - tools(ctx)       -> the host tool registry (or undefined)
+ *   - createScope(fn)  -> cordis createScope
+ *   - allowed(session) -> danger-full-access verdict for the pwsh channel
+ *   - echo(text)       -> result dialect echo (Windows -> MSYS), or undefined
+ *   - spawn(spec)      -> the host subprocess seam, or undefined
+ *   - pwshPath         -> resolved host pwsh 7, or '' (tool not published)
+ *   - pwshEnabled()    -> the row switch for publishing the pwsh tool
+ *   - log(message)     -> the plugin's tagged logger
+ *
+ * Registration decisions re-run on EVERY agent creation: a per-agent tool
+ * face is per-agent truth, and the observed state (variant presets on, a
+ * future official change, dsh-ops appearing) is absorbed automatically
+ * instead of being encoded as a boot-time snapshot.
+ */
+export function mountInlineTools(ctx, deps) {
+  const { tools, createScope, allowed, spawn, pwshPath, echo, log } = deps
+  /** agent -> { timer, scope, denyPwsh } */
+  const agents = new Map()
+  let pwshPublished = false
+  let pwshDisposer = undefined
+  let toolBashHrefPromise
+
+  const faceNames = (agent) => {
+    try {
+      const registry = tools(ctx)
+      return (registry?.schemas?.(agent) ?? []).map((schema) => schema.name).filter((name) => typeof name === 'string')
+    } catch {
+      return []
+    }
+  }
+
+  const unpublishPwsh = () => {
+    if (pwshDisposer) {
+      try { pwshDisposer() } catch { /* already gone */ }
+      pwshDisposer = undefined
+    }
+    pwshPublished = false
+  }
+
+  const publishPwsh = () => {
+    if (pwshPublished) return
+    if (typeof spawn !== 'function' || !pwshPath) return
+    if (deps.pwshEnabled && deps.pwshEnabled() !== true) return
+    const registry = tools(ctx)
+    if (!registry || typeof registry.register !== 'function') return
+    try {
+      const disposer = registry.register(gitbashPwshDefinition({ pwsh: pwshPath, spawn, allowed, msysToWin: deps.msysToWin, echo }))
+      pwshDisposer = typeof disposer === 'function' ? disposer : undefined
+      pwshPublished = true
+      log(`pwsh replacement published (pwsh 7 at ${pwshPath}; the host pwsh tool is denied on this host — its executor is Git Bash)`)
+    } catch (error) {
+      log(`pwsh replacement registration failed: ${error?.message ?? error}`)
+    }
+  }
+
+  const reconcile = (agent, { attempt = 0 } = {}) => {
+    if (!agent?.ctx) return
+    const names = faceNames(agent)
+    // The composed face may not be wired yet at created-time. Retry bounded:
+    // a face with NEITHER tool carries no verdict, and guessing would deny a
+    // real pwsh or duplicate a real bash. Five tries at 250ms, then give up
+    // quietly — a slower composition simply gets no inline patch this boot.
+    if (attempt < 5 && !names.includes('bash') && !names.includes('pwsh')) {
+      const timer = setTimeout(() => { try { reconcile(agent, { attempt: attempt + 1 }) } catch { /* never crash on a timer */ } }, 250)
+      timer.unref?.()
+      const state = agents.get(agent)
+      if (state) state.timer = timer
+      else agents.set(agent, { timer, scope: undefined, denyPwsh: undefined })
+      return
+    }
+    const needs = inlineNeeds(names)
+    let state = agents.get(agent)
+    if (state?.timer) { clearTimeout(state.timer); state.timer = undefined }
+    if (!state) {
+      state = { timer: undefined, scope: undefined, denyPwsh: undefined }
+      agents.set(agent, state)
+    }
+    if (needs.denyPwsh && !state.denyPwsh) {
+      // Deny the fake pwsh first: restrict is per-agent and idempotent, and
+      // the lying tool must never be callable, not even for the milliseconds
+      // before the bash tool lands.
+      try {
+        const deny = tools(ctx)?.restrict?.({ deny: ['pwsh'] })
+        if (typeof deny === 'function') state.denyPwsh = deny
+      } catch (error) {
+        log(`fake-pwsh deny failed: ${error?.message ?? error}`)
+      }
+    }
+    if (needs.needsBash && !state.scope) {
+      if (toolBashHrefPromise === undefined) toolBashHrefPromise = resolveToolBashEntry(ctx)
+      void toolBashHrefPromise.then(async (href) => {
+        if (href === '') {
+          log('official @deepseek-ai/dsh-tool-bash is not resolvable from the host base; bash stays absent for this agent')
+          return
+        }
+        const scope = await activateOfficialBash(agent, href, createScope, log)
+        if (scope && agents.get(agent) === state) {
+          state.scope = scope
+          log('official bash tool activated for this agent (Git Bash host; fake pwsh denied)')
+        }
+      }).catch(() => { /* logged inside activateOfficialBash */ })
+    }
+    // The pwsh replacement follows the ops_ presence: dsh-ops already ships a
+    // PowerShell story through the HOST pwsh tool — which our own deny would
+    // otherwise silence — so ours fills that gap instead of duplicating tools.
+    // The row switch is re-read per reconcile too, so a flip lands at the
+    // next agent (or on the next event after a hot toggle).
+    if (needs.opsPresent || (deps.pwshEnabled && deps.pwshEnabled() !== true)) unpublishPwsh()
+    else publishPwsh()
+  }
+
+  const teardown = (agent) => {
+    const state = agents.get(agent)
+    if (!state) return
+    agents.delete(agent)
+    if (state.timer) clearTimeout(state.timer)
+    if (state.denyPwsh) { try { state.denyPwsh() } catch { /* already lifted */ } }
+    if (state.scope) { void state.scope.dispose().catch(() => {}) }
+  }
+
+  try {
+    ctx.on('agent/created', ({ agent } = {}) => { try { reconcile(agent) } catch { /* never break agent creation */ } })
+    ctx.on('agent/disposed', ({ agent } = {}) => { try { teardown(agent) } catch { /* never break disposal */ } })
+  } catch (error) {
+    log(`agent lifecycle wiring failed: ${error?.message ?? error}`)
+    return
+  }
+
+  ctx.effect(() => () => {
+    for (const agent of [...agents.keys()]) teardown(agent)
+    agents.clear()
+    unpublishPwsh()
+  }, 'dsh-gitbash-shell: inline tool patching')
+}
+
+/**
+ * The pwsh-replacement tool definition, bound to a resolved pwsh 7.
+ *
+ * It spawns through the HOST's `ctx.subprocess` service (credential scrub +
+ * process governance + output spill are the service's own guarantees; the
+ * bash executors consume the same seam). There is deliberately NO file
+ * sandbox on this channel — the tool is published only to danger-full-access
+ * sessions and says so in the error it raises elsewhere.
+ * @param {object} options
+ * @param {string} options.pwsh - resolved pwsh.exe path.
+ * @param {(spec: object) => { done: Promise<object>, collected?: object }} options.spawn - the host subprocess seam.
+ * @param {(session: object|undefined) => boolean} options.allowed - per-session permission verdict.
+ * @param {(path: string) => string} [options.msysToWin] - argument dialect translation (MSYS -> Windows), the same mount table the file tools use.
+ * @param {(text: string) => string} [options.echo] - result dialect echo (Windows -> MSYS), gated by the caller.
+ */
+export function gitbashPwshDefinition({ pwsh, spawn, allowed, msysToWin, echo }) {
+  return {
+    name: 'gitbash_pwsh',
+    description: gitbashPwshDescription(),
+    parameters: {
+      type: 'object',
+      properties: {
+        command: { type: 'string', description: 'Short PowerShell command; for long or multi-line scripts use script_path.' },
+        script_path: { type: 'string', description: 'Existing script path; use instead of command for long or multi-line scripts.' },
+        workdir: { type: 'string', description: 'Working directory; defaults to the session workspace when omitted.' },
+        timeoutMs: { type: 'number', description: 'Timeout in milliseconds; the command is killed on expiry. Defaults to 120000.' },
+      },
+      required: ['command'],
+      additionalProperties: false,
+    },
+    async execute(args, exec) {
+      const invalid = validatePwshArgs(args)
+      if (invalid) throw new Error(invalid)
+      if (!allowed(exec?.agent?.session)) {
+        throw new Error('gitbash_pwsh requires danger-full-access for this session; this channel has no file sandbox.')
+      }
+      const built = pwshArgv(pwsh, args, { msysToWin })
+      if (built.error) throw new Error(built.error)
+      const timeoutMs = args.timeoutMs === undefined ? PWSH_DEFAULT_TIMEOUT_MS : args.timeoutMs
+      const controller = new AbortController()
+      const timer = setTimeout(() => controller.abort(new Error(`command timed out after ${timeoutMs}ms`)), timeoutMs)
+      timer.unref?.()
+      try {
+        const handle = spawn({
+          argv: built.argv,
+          cwd: typeof args.workdir === 'string' && args.workdir !== '' ? args.workdir : undefined,
+          stdio: {
+            stdin: 'ignore',
+            stdout: { maxBytes: 64_000, spill: { maxBytes: 64 * 1024 * 1024 } },
+            stderr: { maxBytes: 64_000, spill: { maxBytes: 64 * 1024 * 1024 } },
+          },
+          signal: controller.signal,
+        })
+        const outcome = await handle.done
+        const streamText = (reader) => {
+          if (reader === undefined || typeof reader.readFrom !== 'function') return ''
+          const read = reader.readFrom(0)
+          const text = read.text ?? ''
+          return echo ? echo(text) : text
+        }
+        const stdout = streamText(handle.collected?.stdout)
+        const stderr = streamText(handle.collected?.stderr)
+        let body = stdout
+        if (stderr !== '') body += (body !== '' && !body.endsWith('\n') ? '\n' : '') + '[stderr]\n' + stderr
+        if (body === '') body = '(no output)'
+        const markers = []
+        if (outcome?.signal !== undefined && outcome?.signal !== null) markers.push(`[signal: ${outcome.signal}]`)
+        if (outcome?.exitCode !== undefined && outcome?.exitCode !== null && outcome.exitCode !== 0) markers.push(`[exit code: ${outcome.exitCode}]`)
+        return { content: [{ type: 'text', text: [body, ...markers].join('\n') }] }
+      } finally {
+        clearTimeout(timer)
+      }
+    },
+  }
+}