🤖 refactor: consolidate bash execution + await/output tools

[ThomasK33]

This PR consolidates the await/output surface for background bash processes and agent tasks, while keeping the bash (shell execution) and task (subagent spawn) tools separate:

• task: spawn sub-agent workspaces only
• bash: execute shell commands only
• task_await / task_list / task_terminate: remain the unified surface for waiting / rediscovering / terminating both agent tasks and background bash processes

Key changes

• task remains agent-only (no kind="bash").
• bash(run_in_background=true) now returns taskId: "bash:<processId>" (in addition to backgroundProcessId) so callers can task_await / task_terminate without needing bash-specific wrappers.
• Frontend tool rendering + token accounting treat bash as its own tool (no more task(kind="bash") special-casing).
• Updated tests/fixtures to request/require bash instead of task(kind="bash").

Compatibility / behavior notes

• task(kind="bash") is intentionally not supported anymore; it will be rejected by the tool schema and backend.
• Existing history entries using task(kind="bash") are handled best-effort by the UI (converted/rendered as bash output when possible).

Verification

• make static-check

---

📋 Implementation Plan

Plan: Consolidate waiting/output for bash + agent tasks under task_await

Goal

Keep the CC-style split, but consolidate waiting/output:

• task: spawn sub-agent workspaces only (subagent_type, prompt, title, run_in_background).
• bash: execute shell commands only (script, timeout_secs, run_in_background, display_name).
• task_await: remain the single “wait/output” tool for both agent tasks and background bash processes.

Recommended approach net LoC (product code only): ~-60 to -110 LoC (mostly deleting the task(kind="bash") branch + schema fields, plus a small additive tweak to bash background results).

Why this is needed (context)

We want a single, obvious tool for shell execution (bash) and a single, obvious tool for waiting/output (task_await) across both agent tasks and background bash processes. This avoids models reaching for bash-specific wrappers when task_await is the unified output surface.

Should we prefix agent task IDs as agent:<workspaceId>?

Recommendation: don’t (for now).

• Agent task IDs are real workspace IDs used across the app. Prefixing would require pervasive parsing/rewiring (and it risks breaking persisted history).
• We only need a prefix for bash tasks because they’re not workspace IDs (they’re process IDs), and task_await needs a cheap routing signal.
• Adding taskId: "bash:<processId>" to bash background results gets us the consistency win (both task and bash return a taskId usable by task_await) without changing agent ID semantics.

If we ever want stronger typing, prefer adding an explicit kind: "agent" | "bash" field to task_list results (low-risk) rather than changing the agent task ID format.

---

Implementation Steps

1) Split tool input surface in TOOL_DEFINITIONS (agent-only task)

Files:

• src/common/utils/tools/toolDefinitions.ts
• src/common/utils/tools/tools.ts

Actions:

1. Make task agent-only again:
   • Set TaskToolArgsSchema = TaskToolAgentArgsSchema.
   • Delete TaskToolBashArgsSchema and the superRefine union enforcement wrapper.
2. Update tool descriptions to match the split:
   • TOOL_DEFINITIONS.task.description: remove “Unified … (bash)” wording and only document agent spawning.
   • TOOL_DEFINITIONS.bash.description: update background guidance to point at task_await / task_list / task_terminate (not bash_output / bash_background_*).
3. Tighten TaskToolResultSchema to agent-only:
   • Remove bash-specific fields from TaskToolCompletedResultSchema (bashResult, exitCode, note, truncated).

Back-compat note: old chat history may still contain toolName="task" + input.kind="bash" and/or bashResult in the result. Preserve UI rendering via unknown-shape type guards (see step 4) rather than keeping task schema “wide”.

2) Remove bash dispatch from backend task tool

File: src/node/services/tools/task.ts

Actions:

• Delete the task(kind="bash") execution branch.
• Remove now-unused helpers/imports:
  • createBashTool, resolveBashDisplayName, toBashTaskId, BashToolResult, and formatBashReport().
• Keep the agent-task logic + defensive input validation intact.

3) Return taskId from bash background results (directly awaitable)

Files:

• src/common/utils/tools/toolDefinitions.ts (result schema)
• src/node/services/tools/bash.ts
• src/node/services/tools/taskId.ts (import-only)

Actions:

1. Extend BashToolResultSchema:
   • Add taskId: z.string() to the background-success variant, where taskId is the prefixed form (bash:<processId>).
2. Populate the new field in bash.ts:
   • When returning background spawn success (run_in_background=true), include taskId: toBashTaskId(spawnResult.processId).
   • When returning “sent to background” (migration path), include taskId: toBashTaskId(processId).

4) Keep task_await unified (no behavior change), but fix UX affordances

Files:

• src/node/services/tools/task_await.ts (likely no code changes)
• src/common/utils/tools/taskToolTypeGuards.ts
• src/browser/stores/WorkspaceStore.ts
• (possible) src/browser/components/tools/* where task(kind="bash") is special-cased

Actions:

1. Maintain unified waiting semantics:
   • Keep routing on taskId prefix (bash: → backgroundProcessManager; else agent task).
2. Update/rename type guards for legacy task(kind="bash"):
   • Stop depending on TaskToolArgs having kind.
   • Prefer an unknown → LegacyTaskBashArgs guard for rendering old history.
3. Simplify WorkspaceStore bash detection:
   • Replace isTaskKindBashToolCall() with isBashLikeToolCall():
     • toolName === "bash" → bash
     • toolName === "task" && input.kind === "bash" → legacy bash-in-task
     • toolName === "task" && result.taskId startsWith("bash:") / bashResult → legacy fallback
   • Remove the special “task uses bash tool under the hood” live-output cleanup path for new flows; keep a minimal legacy fallback only for old sessions.

5) Update token accounting buckets

Files:

• src/common/utils/tokens/tokenStatsCalculator.ts
• src/common/utils/tokens/tokenStatsCalculator.test.ts

Actions:

• Update getConsumerInfoForToolCall():
  • New behavior: toolName === "bash" is its own consumer bucket.
  • Legacy behavior: if toolName === "task" and input.kind === "bash", label as task (bash legacy).

6) Update tests and fixtures to use bash, not task(kind="bash")

Targeted updates (non-exhaustive but high-signal):

• src/common/utils/tools/toolDefinitions.test.ts

  • Remove/replace the assertion that task accepts { kind: "bash" … }.
  • Add an assertion that task rejects bash-shaped input.

• src/node/services/tools/task.bash.test.ts

  • Remove the tests that call createTaskTool() with kind:"bash".
  • Keep (or move) the tests that ensure task_await, task_list, and task_terminate work with taskId="bash:<processId>".

• tests/ipc/runtimeExecuteBash.test.ts

  • Tool policy: require bash (instead of task).
  • Prompt strings + event assertions: verify toolName === "bash" (no args.kind).

• tests/ipc/sendMessage.context.test.ts, tests/ipc/ollama.test.ts

  • Update prompts from “Use task(kind="bash") …” to “Use bash() …”.

• src/node/services/taskService.test.ts

  • Replace parent message part toolName:"task" input:{kind:"bash",…} with toolName:"bash" input:{script,…}.

• Update any strict equality assertions for background bash results to allow / expect the new taskId field.

7) Verification

Run a tight set of checks after changes:

• Unit: bun test src/node/services/tools/task.test.ts
• Unit: bun test src/node/services/tools/task.bash.test.ts
• Unit: bun test src/common/utils/tools/toolDefinitions.test.ts
• Unit: bun test src/common/utils/tokens/tokenStatsCalculator.test.ts
• IPC integration: TEST_INTEGRATION=1 bun x jest tests/ipc/runtimeExecuteBash.test.ts -t "execute simple bash commands"

---

Generated with mux • Model: openai:gpt-5.2 • Thinking: xhigh

[ThomasK33]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[ThomasK33]

Fixed in db76e83 - coerceTimeoutMs now accepts 0 and only rejects negative values, so timeout_secs: 0 correctly translates to an immediate non-blocking check for both agent and bash tasks.

[ThomasK33]

@codex review

Follow-up: fixed static-check by aligning RefreshController unit tests with the current (post-revert) RefreshController API.

[chatgpt-codex-connector]

💡 Codex Review

mux/src/node/services/tools/task_await.ts

Lines 174 to 176 in f84415d

	try {
	const report = await taskService.waitForAgentReport(taskId, {
	timeoutMs,

Handle timeout_secs=0 for agent tasks

task_await now allows timeout_secs=0 (per schema/docs for non‑blocking bash), but createTaskAwaitTool forwards timeoutMs directly into taskService.waitForAgentReport. When timeout_secs is 0 and any agent task is included (explicitly or via the default “all active tasks”), waitForAgentReport asserts timeoutMs > 0 and throws, so those agent results come back as error instead of queued|running|awaiting_report. Consider treating 0 as “no wait” for agent tasks (return current status) or defaulting to the 10‑minute wait when timeout_secs is 0.

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[ThomasK33]

Fixed in cac38ca.

• task_await now treats timeout_secs: 0 (or values that round down to 0ms) as non-blocking for agent tasks.
• We validate scope first, then return the current task status (queued/running/awaiting_report) instead of calling waitForAgentReport (which asserts timeoutMs > 0).
• Added a regression test covering timeout_secs=0.

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[ThomasK33]

@codex review

Addressed the truncation metadata issue for task(kind="bash") by:

• adding truncated?: { reason; totalLines } to the task completed result schema
• including a truncation warning + metadata in reportMarkdown when present
• making bash tool init lazy so agent-only task tests don't require bash deps

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[ThomasK33]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Another round soon, please!

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[ThomasK33]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[ThomasK33]

@codex review

Fixed pending task finalization to skip task(kind="bash") (and added a unit test).

Also extended workspace activity snapshots to include the last thinking level for the VS Code extension.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[ThomasK33]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[ThomasK33]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[ThomasK33]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".