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

Add dynamic project instruction loading

Browse Source
This commit is contained in:
Yichen Jiang
2026-07-05 18:18:32 +08:00
parent a4cabe8485
commit 87c5b55122
13 changed files with 299 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files
packages/prompt/project-instructions/README.md
+7 -5
View File
@@ -1,14 +1,16 @@
# @deepseek-ai/dsh-project-instructions
Project instruction file loader for the harness. It discovers `AGENTS.md` with `CLAUDE.md` fallback for each agent session and injects the loaded content as fenced workspace context before model requests.
Project instruction file loader for the harness. It discovers `AGENTS.md` with `CLAUDE.md` fallback for each agent session, injects the baseline content as fenced workspace context before model requests, and lazily adds nested instruction files when structured file tools touch deeper paths.
## Behavior
The plugin listens on the `agent/request` waterfall and depends on the `ctx.fs` provider seam to read instruction file content. For each request it derives the workspace from `agent.session.header.cwd`; if the session has no cwd, it falls back to `process.cwd()` for single-session local/stdio runs. It then finds the project root by walking upward until it sees `.git` as either a directory or a file, considers the ancestor chain from project root to cwd, and loads at most one instruction file per directory: `AGENTS.md` wins, `CLAUDE.md` is a compatibility fallback.
The plugin also listens on `tools/post-execute` for successful structured filesystem touches from the first-party `read`, `write`, and `edit` tools. When one of those tools touches a descendant of the session cwd, the plugin checks the directories between the session cwd and the touched file for instruction files that were not already loaded in that session, then attaches them as `additionalContext` so the loop records a durable `context/message` for the next model request. This intentionally follows file-tool touches, not shell `cd`: `dsh-bash-local` uses fresh shells per call, and parsing arbitrary shell commands for reached paths would be brittle.
User-global instructions live at `$DSH_HOME/AGENTS.md`; `$DSH_HOME` defaults to `~/.dsh`. A configured `~`, `~/...`, or Windows-style `~\...` prefix is expanded against the operating-system home directory before resolution. The user-global file renders before project files, so deeper project files appear later in the context and can override broader guidance.
The loaded files are inserted as a synthetic user-role workspace-context message, not as provider system text and not as persisted session events. The rendered envelope states that these files are workspace-provided guidance, lower authority than system/developer/direct user instructions, and must not override safety, permission, or secret-handling rules.
Baseline files are inserted as a synthetic user-role workspace-context message, not as provider system text and not as persisted session events. Nested files discovered after structured file tools run are inserted through the existing `context/message` path, so they persist with the session and resume like other plugin-provided context. The rendered envelope states that these files are workspace-provided guidance, lower authority than system/developer/direct user instructions, and must not override safety, permission, or secret-handling rules.
## Config
@@ -21,14 +23,14 @@ export interface Config {
}
```
`projectRootMarkers` defaults to `['.git']`, `baselineMaxBytes` defaults to `65536`, and `enableClaudeFallback` defaults to `true`. Setting `baselineMaxBytes` to `0` or another non-positive value disables instruction injection.
`projectRootMarkers` defaults to `['.git']`, `baselineMaxBytes` defaults to `65536`, and `enableClaudeFallback` defaults to `true`. Setting `baselineMaxBytes` to `0` or another non-positive value disables both baseline and nested instruction injection.
## Budgeting and cache
The renderer keeps full text until the configured byte budget is exceeded. When it must trim, it preserves more-specific files first, drops whole less-specific files before truncating a more-specific file, and emits an HTML comment naming omitted and truncated files with byte counts.
Discovery re-walks the applicable ancestor chain on every request so newly created baseline files are noticed. File content is cached by normalized absolute path plus the provider's opaque file version and size; a changed signature causes a re-read. The discovery pass carries the file signature forward to the read pass, so a cache hit does not stat the same instruction file twice in one request.
Discovery re-walks the applicable ancestor chain on every request so newly created baseline files are noticed. File content is cached by normalized absolute path plus the provider's opaque file version and size; a changed signature causes a re-read. The discovery pass carries the file signature forward to the read pass, so a cache hit does not stat the same instruction file twice in one request. Nested instruction paths are tracked separately per live session so cache eviction or repeated reads do not duplicate the same durable context.
## Non-goals
This phase does not implement lazy on-touch nested loading, `contextPaths()`, shell parsing, lowercase filenames, `.claude/` rule directories, local/private variants, `@path` imports, file watching, or model-generated summaries. Those need separate semantics and, for on-touch loading, real structured file tools that can report touched paths.
This phase does not implement `contextPaths()`, shell parsing, bash-`cd`-based instruction loading, lowercase filenames, `.claude/` rule directories, local/private variants, `@path` imports, file watching, or model-generated summaries. Those need separate semantics beyond structured file-tool touches.