fix: make agent recovery and concealment reliable (#2189)

This commit is contained in:
evandance
2026-08-05 19:44:11 +08:00
committed by GitHub
parent 0848a5b4fa
commit 09feefe96b
58 changed files with 3606 additions and 328 deletions
+3 -1
View File
@@ -316,7 +316,9 @@ These codes remain available to in-process hosts through the wrapped
rule and shipped-tree summary. Agents consuming a host that explicitly
enabled concealment should match `error.type == "validation"` and
`error.subtype == "command_unavailable"` instead of branching on a
rule-specific reason.
rule-specific reason. The canonical
[`validation/command_unavailable` contract](../../errs/ERROR_CONTRACT.md#concealed-commands-validationcommand_unavailable)
defines its exit code, wire fields, and consumer behavior.
## Where to go next
+5
View File
@@ -29,6 +29,11 @@ import "io/fs"
// the CLI builds. Later additions or removals of top-level directories do
// not change the manifest; files within an owned skill directory are read
// live. Base and Overlay must contain only valid skill directories.
// A skill may declare hard dependencies under
// metadata.requires.skills in SKILL.md. Every declared dependency must be
// present in the final composed manifest; Allow is never widened and Remove
// is never overridden to satisfy one. A same-named Overlay replacement uses
// the replacement SKILL.md's dependency metadata, not the base copy's.
// Declaring this asset composition is a build-integrity commitment: invalid
// selection, content, ownership, or reference remaps abort the build rather
// than silently falling back to host defaults.