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

Fix workspace context review findings

Browse Source
This commit is contained in:
Yichen Jiang
2026-07-13 16:31:03 +08:00
parent adf6b8a1ab
commit aa62b5109a
40 changed files with 708 additions and 252 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
packages/prompt/workspace-context/README.md
View File

@@ -8,7 +8,7 @@ The baseline is composed once per agent-loop instance on `agent/session-prefix`.
The plugin also listens on `tools/post-execute` for successful first-party `read`, `write`, and `edit` calls. Each touch checks newly reached descendant scopes and every previously loaded scope. A new file is attached through the result's `additionalContexts`; a changed file or candidate switch appends a replacement; a missing final candidate appends a removal notice. Native calls and Code Mode sub-dispatches share this path: `run_code` defers each nested context until its outer result, so the loop still appends updates after tool-call/result adjacency is complete. This follows structured filesystem activity rather than shell `cd`, because each local bash call starts a fresh shell and parsing arbitrary shell syntax would be unreliable.
Instruction reads use the optional `ctx.fs` provider. The plugin does not statically inject `fs`, so providerless product trees still boot and instruction loading becomes a no-op until a provider is present. It calls `ctx.fs.lstat` before resolving a candidate, rejecting a final-component symlink instead of following repository-owned links across the trust boundary. A provider failure after a file was loaded is treated as temporarily unavailable, not as proof that the file was deleted.
Instruction reads use the optional `ctx.fs` provider. The plugin does not statically inject `fs`, so providerless product trees still boot and instruction loading becomes a no-op until a provider is present. It calls `ctx.fs.lstat` before resolving a candidate, rejecting a final-component symlink instead of following repository-owned links across the trust boundary. Prefix cancellation and dynamic tool cancellation propagate through resolution, metadata probes, and streaming reads. A provider failure after a file was loaded is treated as temporarily unavailable, not as proof that the file was deleted.
## Prompt Shape
@@ -46,7 +46,7 @@ The core `context/message` envelope is disabled for these messages because the p
## State And Refresh
Model-visible text contains no hidden state markers. Each dynamic context event instead carries JSON metadata with a versioned list of `{ action, scope, path, previousPath?, digest? }` changes. On every relevant tool touch, the plugin reconstructs loaded state from its visible session events and overlays a short in-memory pending window for context returned by `tools/post-execute` but not yet appended by the loop.
Model-visible text contains no hidden state markers. Each dynamic context event instead carries JSON metadata with a versioned list of `{ action, scope, path, previousPath?, digest? }` changes. On every relevant tool touch, the plugin reconstructs loaded state from its visible session events and overlays a short in-memory pending window for context present on the immutable top-level `tools/result` but not yet appended by the loop. Nested Code Mode results stage pending changes under the outer execution token for same-run duplicate suppression; the outer result rolls that state back and recommits only contexts that survived outer policy.
An unchanged path and SHA-1 content digest is not injected again. Resume works because visible metadata is persisted in the session log. Compaction re-arms a scope after its context event leaves the visible surface. A removal is a tombstone, so a later candidate reappearance is loaded again. Only changes actually rendered within the byte budget enter metadata and pending state; an omitted change remains eligible for a later touch.
@@ -59,19 +59,20 @@ export interface Config {
dshHome?: string
projectRootMarkers?: string[]
maxBytes: number
maxSourceBytes?: number
instructionFileCandidates?: string[]
}
```
`maxBytes` is required so each deployment makes its prompt-budget choice explicitly. `projectRootMarkers` defaults to `['.git']`, and `instructionFileCandidates` defaults to `['AGENTS.md', 'CLAUDE.md']`. In each project directory, the first existing candidate wins; with defaults, `AGENTS.md` is native and `CLAUDE.md` is the compatibility fallback. Candidate entries must be same-directory file names, so empty entries, `.`/`..`, and entries containing `/` or `\` are ignored.
`maxBytes` is required so each deployment makes its prompt-budget choice explicitly. `maxSourceBytes` limits each source instruction file before rendering and defaults to 1 MiB. `projectRootMarkers` defaults to `['.git']`, and `instructionFileCandidates` defaults to `['AGENTS.md', 'CLAUDE.md']`. In each project directory, the first existing candidate wins; with defaults, `AGENTS.md` is native and `CLAUDE.md` is the compatibility fallback. Candidate entries must be same-directory file names, so empty entries, `.`/`..`, and entries containing `/` or `\` are ignored.
The user-global file is always `$DSH_HOME/AGENTS.md`; the candidate list only controls project scopes. `$DSH_HOME` defaults to `~/.dsh`, and configured `~`, `~/...`, and Windows-style `~\...` prefixes are expanded against the operating-system home directory. A non-positive or non-finite byte budget disables both baseline and dynamic loading.
The user-global file is always `$DSH_HOME/AGENTS.md`; the candidate list only controls project scopes. `$DSH_HOME` defaults to `~/.dsh`, and configured `~`, `~/...`, and Windows-style `~\...` prefixes are expanded against the operating-system home directory. A non-positive or non-finite render budget disables both baseline and dynamic loading; configured `maxSourceBytes` must be a positive integer.
## Budgeting And Cache
## Budgeting And Bounded Reads
Rendering preserves the most specific instruction files first. It drops whole broader files before truncating the most-specific file and emits a visible `Workspace instruction budget ...` notice naming omitted and truncated paths. The rendered bytes never exceed `maxBytes`.
Each discovered candidate is read and identified by normalized absolute path, the provider's opaque version, and a SHA-1 content digest. Comparing the digest after reading prevents a same-version, same-size rewrite from returning stale cached text. Discovery carries the provider version into reading so one pass does not stat the same file twice. Cache identity is separate from loaded-state identity: persisted structured metadata, not cached prose, controls duplicate suppression.
Instruction content is read through `streamText()` under `maxSourceBytes`, even when provider metadata omits size or a file grows after its metadata probe. An oversized file is ignored without falling through to a lower-priority same-directory candidate; during dynamic reconciliation it is temporarily unavailable rather than removed. The plugin keeps no process-wide prose cache: every reconciliation observes current content, computes its SHA-1 only after the bounded read, and relies on persisted structured metadata for duplicate suppression.
## Non-goals