Pine/deepseek-harness
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases 1 Wiki Activity

fix(mode): stabilize plan-mode model experience

Browse Source
This commit is contained in:
Tianyi Cui
2026-07-20 22:13:59 +08:00
parent 28a74d6d7e
commit 897dc82d9b
40 changed files with 620 additions and 298 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/plan-acp-agent/README.md
+5 -3
View File
@@ -4,7 +4,9 @@ The coding agent as an ACP server with **session modes** composed — the live c
## What it demonstrates
`session/new` advertises the mode picker (`default` / `plan`) plus the sandbox-mode and approval config options — two independent axes on one session, the composition this example exists to demonstrate. Plan mode adds the plan guidance section and the `exit_plan_mode` tool and touches nothing else: the sandbox keeps whatever mode its own knob says (workspace-write here by default), escalation prompts work in plan exactly as in default, and switching either axis never disturbs the other, in any order. A user who wants a hard read-only floor while planning flips the sandbox-mode option to read-only alongside the mode picker. There is deliberately no per-mode tool list either: `write`/`edit`/`bash` stay present in plan and the section's guidance is what defers changes to after the review (the effects-based generalization is the RFC's deferred item). A blocking decision goes to the user through `ask_user_question`. The model leaves by presenting its plan through `exit_plan_mode`: the plan markdown renders as the tool's call card, the review question arrives as an elicitation form (approve / keep planning, free text welcome), and a keep-planning answer returns the feedback to the model verbatim.
`session/new` advertises the mode picker (`default` / `plan`) plus the sandbox-mode and approval config options — independent axes on one session. The composition owns its full plan instructions in [`cordis.yml`](cordis.yml): persist in the selected mode, inspect before asking, avoid mutations, resolve discoverable facts from the repository, and produce a decision-complete plan through `exit_plan_mode`. These are the most instrumental behaviors shared by the local Codex and Claude Code plan-mode references without importing their product-specific plan files, phase machinery, or protocol tags.
Plan mode adds only the configured guidance section. Every other tool, including `exit_plan_mode`, has the same schema in `default` and `plan`; the exit tool describes itself as plan-only and rejects if called outside plan mode. Keeping both native schemas and Code Mode's SDK stable avoids tool-catalog churn at the transition. The sandbox retains its own mode (workspace-write here by default), escalation prompts work identically, and a user who wants a hard read-only floor selects read-only separately. A blocking user-owned choice goes through `ask_user_question`. In plan mode, `exit_plan_mode` renders the submitted markdown as a call card and asks for approval or corrective feedback through ACP elicitation.
## Run
@@ -12,8 +14,8 @@ The coding agent as an ACP server with **session modes** composed — the live c
pnpm run demo:plan-acp # needs DEEPSEEK_API_KEY (repo-root .env works)
```
Drive it from Zed or any ACP client; the mode picker appears on the session beside the sandbox/approval selects. Switching back to `default` (or an approved `exit_plan_mode`) drops the plan section and the exit tool on the next step; the sandbox and approval knobs stay exactly where the user left them.
Drive it from Zed or any ACP client; the mode picker appears beside the sandbox and approval selects. Switching back to `default`, directly or through an approved `exit_plan_mode`, drops only the plan section on the next step. The tool catalog and the independent knobs stay unchanged.
## Tests
`pnpm run test:snapshot` replays three scenarios keyless (the recorded bash re-executes for real under the host's sandbox runner — Seatbelt on macOS, bwrap on Linux CI). `modes-advertise` (authored): the `modes` advertisement and both config options on `session/new`, both `session/set_mode` round-trips with their optimistic `current_mode_update`, and the loud rejection of an unknown mode id, as committed wire bytes. `plan-mode` (recorded, the header pin): the full arc — setMode(plan), the plan-shaped initial header (full toolset + exit tool + section), a real `cat` run inside plan (under the sandbox's own workspace-write default — the mode does not change it), the plan presented via `exit_plan_mode`, a scripted elicitation approve, the boundary-flushed `mode/set` back with a complete changed `request/header`, then a real edit mid-turn. `plan-mode-reject` (recorded): the keep-planning branch, whose corrective `isError` carries the reviewer's free-text feedback verbatim and leaves the session in plan mode. The sandbox-denial marker stays pinned at the unit tier (`packages/bash/tool-bash/tests` — a recorded denial's stderr would be the backend's dialect and replay only where it was recorded).
`pnpm run test:snapshot` replays the ACP mode and plan-review surfaces keyless, including stable schemas across approval and real filesystem calls under Seatbelt on macOS or bwrap on Linux. `pnpm run test:e2e` adds a self-skipping live-model smoke that verifies the file is unchanged when review appears and changed only after approval. Sandbox denial remains covered at the `dsh-tool-bash` unit tier because recorded backend stderr is platform-specific.