Add token economy composer mode / 增加省 token 协作模式

[SivanCola]

Summary

#4071 has already landed in main-v2. This PR is now rebased directly on the latest main-v2 and should be reviewed as the follow-up delta that adds a desktop/composer token economy mode. The goal is to keep the first provider-visible request lean for economy sessions while preserving the existing full/default request shape.

What changed

• Added a TokenMode boot profile:
  • full / empty keeps the existing default capability surface.
  • economy keeps only core coding tools visible at boot and moves optional sources behind connect_tool_source.
• Added connect_tool_source for economy mode to enable optional capabilities on demand:
  • skills
  • MCP servers, including enabled built-in MCP servers
  • CodeGraph
  • LSP
  • web_fetch
  • install_source
  • task
• Kept cache-sensitive behavior explicit:
  • full mode and default mode now have a request snapshot regression test that compares the system prompt, tool schema order, and complete provider-visible schemas.
  • economy mode has a fixed first-request tool list/order regression test.
  • skills index, MCP tools, CodeGraph steering, LSP tools, task, web_fetch, and install_source are hidden from the economy first request until explicitly enabled.
• Added desktop tab support for token mode:
  • persisted per tab only when set to economy
  • inherited by new blank sessions
  • included in Meta and TabMeta
  • rebuilds the controller when changed so the provider sees the new request profile
  • rejects switching while a turn is running
• Reworked composer state into a single ComposerProfile model:
  • keeps collaboration mode, tool approval mode, token mode, goal, and goal draft state independent
  • prevents stale backend tab/meta hydration from erasing a pending local plan/token/goal selection
  • replaces the narrower goalDraftMode helper and test
• Added desktop UI affordances:
  • token saver item in the collaboration menu
  • active token saver chip in the composer control row
  • localized English and Chinese copy
  • CSS sizing so plan/goal/token chips do not truncate or shift the toolbar
• Added docs:
  • docs/COLLABORATION_MODES.zh-CN.md
  • docs/TOOL_APPROVAL_MODES.zh-CN.md
  • linked both from docs/GUIDE.zh-CN.md

Cache behavior

This intentionally creates two stable request profiles rather than mutating the existing full profile:

• Existing default/full mode should keep the same provider-visible system prompt and tool schemas.
• Economy mode uses a smaller first-request prefix by omitting optional skills/MCP/CodeGraph/LSP/tool schemas up front.
• Optional capabilities are enabled only after connect_tool_source, so later turns pay for them only when the task actually needs them.

Rebase update

After #4071 was merged, this branch was rebased with only the token economy commit on top of the latest main-v2. The PR diff now excludes the history/memory retrieval implementation from #4071 and contains only the token economy follow-up changes.

Validation

Passed:

• git diff --check
• go test ./internal/boot -count=1
• go test ./... from desktop/
• pnpm --dir desktop/frontend test:typecheck
• pnpm --dir desktop/frontend test
• pnpm --dir desktop/frontend check:css

Local environment notes:

• go test ./... with the normal user environment failed only because machine-local MCP configuration polluted internal/control TestRemoveMCPServerRemovesUnconnectedLazyPlaceholder.
• Re-running go test ./... with an isolated home removed that MCP-config failure; the remaining local failures were macOS sandbox confinement tests (internal/sandbox TestSandboxEnforcesWrites and internal/tool/builtin TestBashSandboxConfinement) where this host did not deny writes outside the sandbox root. Those failures are unrelated to this token economy change.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: af98812842

ℹ️ About Codex in GitHub

Your team has set up Codex to 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 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[esengine]

Merged — another excellent one, and thanks for the clean rebase onto main-v2 after #4071 landed.

I measured the effect end-to-end: economy cuts the first-request prefix ~56% on a lean project (37 tools / ~11.8k tok → 20 / ~5.2k) and more as skills/MCP/CodeGraph/LSP stack up — the economy prefix stays flat while full grows. That's exactly the lever this project is built around.

What I appreciated most: the default/full profile is provably byte-identical (full == default on the complete provider-visible schemas, not just names), so existing users' cache is untouched — and you fenced both profiles with regression tests. The connect-on-demand round-trip is the right tradeoff for the lean-coding common case.

One note for the record (not a change request): in economy the agent spends a turn on connect_tool_source before using an optional source, so very tool-heavy sessions recover less of the saving — might be worth a sentence in the docs if you revisit. Thanks again.