fix(workspace-context): guard baseline restoration

This commit is contained in:
fz
2026-08-03 22:05:17 +08:00
parent eff7b758b1
commit 19be29cd0b
31 changed files with 249 additions and 105 deletions
+3 -3
View File
@@ -6,7 +6,7 @@ Per-session workspace instruction loading for `AGENTS.md`-compatible files. The
## Lifecycle
The baseline is injected at the first `agent/step` of each live session. It reads `$DSH_HOME/AGENTS.md` followed by, in each directory from the project root to `agent.session.header.cwd`, every existing base candidate and then every existing local-overlay candidate. Within one directory, candidates whose content is byte-identical after trimming leading and trailing whitespace collapse to the earliest candidate in configured order, so a `CLAUDE.md` that merely duplicates its sibling `AGENTS.md` is rendered once. The durable sourced `user/message` enters the same request as the claimed prompt. If a later surface replacement such as compaction shadows that baseline, `system-prompt/assemble` recomposes and injects the current chain before the loop snapshots its next request.
The baseline is injected at the first `agent/step` of each live session. It reads `$DSH_HOME/AGENTS.md` followed by, in each directory from the project root to `agent.session.header.cwd`, every existing base candidate and then every existing local-overlay candidate. Within one directory, candidates whose content is byte-identical after trimming leading and trailing whitespace collapse to the earliest candidate in configured order, so a `CLAUDE.md` that merely duplicates its sibling `AGENTS.md` is rendered once. The durable sourced `user/message` enters the same request as the claimed prompt. If a later surface replacement such as compaction shadows that baseline, a model-request `system-prompt/assemble` recomposes and injects the current chain before the loop snapshots that request; inspection-only assemblies such as TUI `/status` do not mutate the session.
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. Each configured candidate name is an independent scope in its directory: a newly present file is attached through the result's `additionalContexts`; a changed file appends a replacement; a file that disappears or becomes a per-directory duplicate of an earlier 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.
@@ -52,7 +52,7 @@ Model-visible text contains no hidden state markers. Each baseline or dynamic co
An unchanged path and SHA-1 content digest is not injected again. A per-session, per-scope provider cache stores only `{ path, version, digest, trimmedDigest }`: when the provider's opaque `FsVersion` and the effective visible state both match, reconciliation skips the content read; a changed version triggers a bounded read and SHA-1 confirmation before any model-visible update. The `trimmedDigest` — SHA-1 over the whitespace-trimmed content — is the per-directory duplicate key, so an unchanged file can still be removed when an earlier candidate converges on its content. Resume works because SHA-1 state is persisted in the typed source, while an empty in-memory version cache merely causes one confirming read. Compaction re-arms a scope after its context event leaves the visible surface even when the cached version is unchanged. A removal is a tombstone, so a later candidate reappearance is loaded again. Only model-visible changes actually rendered within the byte budget enter the source, pending state, and version cache; an omitted change remains eligible for a later touch, while a same-digest version refresh updates only the provider cache.
The initial baseline event itself is not rewritten. Its typed changes remain authoritative only while that event is in the visible session surface. After a surface replacement removes it, prompt assembly recomposes the current baseline for that replacement generation and injects it before the first post-replacement request; a queued baseline that is discarded with its step remains eligible for the next request. A successful filesystem touch can still append later replacements or removals. The in-memory scope marker and provider-version cache only select and accelerate probes. A hot plugin remount retains a baseline only when its typed event remains visible, while rebuilding current scope and version tracking; otherwise it injects a current baseline. A resumed loop always recomposes the current baseline and also reconciles still-visible dynamic scopes before its first request. There is no file watcher, so an on-disk change becomes visible at the next successful `read`, `write`, or `edit` touch, when prompt assembly restores a shadowed baseline, or when a resumed loop prepares its baseline.
The initial baseline event itself is not rewritten. Its typed changes remain authoritative only while that event is in the visible session surface. After a surface replacement removes it, model-request prompt assembly recomposes the current baseline and rechecks cancellation, visibility, and the current replacement generation immediately before injecting it. Concurrent preparations can read in parallel, but only the first commit queues a baseline; inspection-only assemblies never restore one. A successful filesystem touch can still append later replacements or removals. The in-memory scope marker and provider-version cache only select and accelerate probes. A hot plugin remount retains a baseline only when its typed event remains visible, while rebuilding current scope and version tracking; otherwise it injects a current baseline. A resumed loop always recomposes the current baseline and also reconciles still-visible dynamic scopes before its first request. There is no file watcher, so an on-disk change becomes visible at the next successful `read`, `write`, or `edit` touch, when a model request restores a shadowed baseline, or when a resumed loop prepares its baseline.
## Configuration
@@ -162,7 +162,7 @@ Append-only; newly visible content follows the reusable request prefix and does
## Known Limitations and Deferred Work
- **Discovery follows structured fs tools, not shell navigation** — a `bash` command that changes directories does not trigger nested instruction discovery because shell syntax and per-call shell state are not a reliable filesystem seam.
- **Refresh is event-driven** — there is no watcher; external edits become visible on the next successful first-party `read`, `write`, or `edit`, when prompt assembly restores a shadowed baseline, or when a resumed loop prepares its baseline.
- **Refresh is event-driven** — there is no watcher; external edits become visible on the next successful first-party `read`, `write`, or `edit`, when model-request prompt assembly restores a shadowed baseline, or when a resumed loop prepares its baseline.
- **Candidate semantics stay intentionally small** — lowercase names, `.claude/rules/`, and `@path` imports are not interpreted; project scopes load `AGENTS.local.md`/`CLAUDE.local.md` overlays by default, but the user-global `$DSH_HOME` scope has no local overlay and other custom names require explicit candidate configuration.
- **Per-directory dedup is content-based** — sibling candidates collapse only when byte-identical after trimming leading and trailing whitespace; a `CLAUDE.md` that symlinks its sibling `AGENTS.md` resolves to the same content and collapses like any duplicate, while a distinct real copy that has drifted from `AGENTS.md` loads in full alongside it.
- **Symlinked instruction files are followed across the trust boundary** — a candidate whose final component is a symlink is resolved and its target loaded, so a cloned repository can surface off-tree file content as lower-authority workspace guidance (it never overrides system, developer, or direct user instructions). Confine `ctx.fs` with the filesystem policy gate or an OS sandbox when loading untrusted repositories.