feat(system-prompt): prompt variables, persona-as-section, tool-guidance ownership

One principle: every fact in the assembled prompt has exactly one owner.

- dsh-system-prompt: merge-extensible AssembleContext on assemble();
  a variable(name, provider) registry; {{name}} interpolation in
  renderPrompt, strict (unknown/valueless/malformed references throw);
  duplicate section and variable names rejected; assembly carries
  resolved section text + variables through the assemble waterfall.
- dsh-agent declares AssembleContext.agent; dsh-agent-loop registers
  the agent:persona section (order 0 - identity renders before tool
  guidance) and the model/cwd variables, and drops its string join:
  renderPrompt(assembly) IS the full prompt.
- Tool guidance moves to its owners: descriptions carry per-tool
  semantics; sections only cross-call habits (tool:bash exit-code
  habit at order 105; read's not-shell nudge). todo/subagent need no
  section - their descriptions already carry the contract.
- SubagentProvider.inheritsParentContext (spawn/acp false, fork true);
  dsh-tool-subagent derives truthful per-provider wording and resolves
  the provider at load (backend must be listed first).
- Example personas shrink to identity + behavior with {{model}} (and
  {{cwd}} in the ACP tree); the welcome banner stops enumerating tools.

RFC: docs/rfc/implemented/architecture/2026-07-05-prompt-variables-and-tool-guidance-ownership.md
This commit is contained in:
Tianyi Cui
2026-07-05 01:54:46 +08:00
parent 1e2efc861e
commit f256f3961d
41 changed files with 746 additions and 177 deletions
+10 -23
View File
@@ -23,9 +23,8 @@
- deepseek-v4-flash
- deepseek-v4-pro
# Local bash executor for agent-core's tool-bash schema.
# FIXME(config-comments): keep this executor note from implying bash is the
# whole tool set; filesystem, subagent, and todo_write are loaded below.
# Local bash executor for agent-core's tool-bash schema (one of several tool
# stacks in this tree: filesystem, subagent, and todo_write load below).
- id: bash
name: '@deepseek-ai/dsh-bash-local'
config:
@@ -39,28 +38,16 @@
config:
model: deepseek-v4-flash
persistenceRoot: !!js process.env.DSH_SNAPSHOT_SESSIONS_ROOT ?? './.sessions'
# The persona: identity + behavior only. Tool guidance lives with each tool
# plugin (descriptions + prompt sections); {{model}} and {{cwd}} are prompt
# variables the agent loop resolves per session (every ACP session carries
# the client's cwd, so the persona can state the workspace).
systemPrompt: |
You are a coding assistant driven over the Agent Client Protocol.
You are a coding assistant powered by the {{model}} model, driven over
the Agent Client Protocol. Your working directory is {{cwd}}.
Your tools are read/write/edit for file operations, bash (plus
bash_output/bash_kill for background tasks), and subagent. Use read to
inspect UTF-8 text files, write to create or replace files, and edit for
targeted literal replacements. Use bash for shell commands, tests,
searches, and operations that are not ordinary file reads or edits. Each
bash call runs in a fresh shell — pass workdir instead of cd. Check the
[exit code: N] marker; verify your work. Keep answers brief and factual.
Use the subagent tool to delegate a focused, self-contained subtask to
a fresh child agent (it works in its own context and returns only its
final result) — give it a complete, standalone instruction. Use
subagent_fork instead when the subtask needs THIS conversation's
context: the child inherits the log so far.
For multi-step work, use the todo_write tool to track a task list:
send the WHOLE list each call (it replaces the previous one), keep at
most one task in_progress (exactly one while work remains), and mark a
task completed as soon as it is done. Skip it for trivial single-step
tasks.
Verify your work by running the code or tests. Keep answers brief and
factual.
# The subagent seam + both in-process backends + two model-facing tools, as leaf
# entries after the app (which provides ctx.agents/ctx.tools). spawn (a fresh