start-command 0.42.0 → 0.43.0

package/CHANGELOG.md

@@ -1,5 +1,18 @@
 # start-command
 
+## 0.43.0
+
+### Minor Changes
+
+- 8567804: `--use-formal-ai`: bounded typed source edits (#261).
+
+  - **Bound preimage:** an `edit` or `write` of an existing file is bound to the complete bytes the `read` tool returned for that exact path in the same task. `cat`, a directory listing or a stale read is not enough, and a change of the target while start writes is refused. `read` returns the text as the numbered frame Formal AI binds as a whole-file Read and decodes back exactly.
+  - **Owned change only:** `old_string` must occur exactly once (or `replace_all` is set), and only those byte regions change. Unicode, CRLF, a byte order mark and a missing final newline are preserved, and symlinks are followed only inside the working directory.
+  - **Receipt and verification:** each change keeps a mutation receipt (`tool`, `target`, `preimage`, `postimage`, byte `matches`, `readback`) in the evidence and returns it to Formal AI. The readback is verified byte for byte against the preimage and the matches, independently of fingerprints; any other byte change ends the run with 70.
+  - **Unresolved failures:** a failed `read`, `edit`, `write` or `bash` change stays open until a verified change of the same target (or the same command) resolves it. A later read, a `cp`/`mv` over the file or Formal AI's answer does not resolve it, and the run ends `incomplete` with the original failure and its exit status.
+  - **Gap records:** `SourcePreimageUnavailable` is a known `workspace-change-verification-gap/v1` reason. Other JSON gap records and Formal AI's Links Notation `obligation_gap` answers end `incomplete` with exit 65 instead of success, unless they only repeat tool output (including a read frame's text). There is still no shell or vendor fallback.
+  - **Dependency:** log sanitization now depends on `@link-foundation/sensitive-data-sanitizer` ^0.14.0, its current release.
+
 ## 0.42.0
 
 ### Minor Changes

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "start-command",
-  "version": "0.42.0",
+  "version": "0.43.0",
   "description": "Gamification of coding, execute any command with ability to auto-report issues on GitHub",
   "main": "src/bin/cli.js",
   "exports": {
@@ -68,7 +68,7 @@
     ]
   },
   "dependencies": {
-    "@link-foundation/sensitive-data-sanitizer": "^0.12.2",
+    "@link-foundation/sensitive-data-sanitizer": "^0.14.0",
     "command-stream": "^2.6.0",
     "lino-objects-codec": "^0.9.5"
   },

package/src/lib/formal-ai-client.js

@@ -15,6 +15,7 @@ const { API_PREFIX } = require('./formal-ai-runtime');
 const {
   FILE_TOOLS,
   describeFileCall,
+  echoedByTool,
   normalizeFileArgs,
 } = require('./formal-ai-files');
 
@@ -144,7 +145,12 @@ const GAP_SCHEMAS = Object.freeze({
     'EditToolUnavailable',
     'ReadToolUnavailable',
   ],
-  'workspace-change-verification-gap/v1': ['VerificationRunToolUnavailable'],
+  // SourcePreimageUnavailable: Formal AI could not bind the complete Read
+  // before its edit to the change (issue #261).
+  'workspace-change-verification-gap/v1': [
+    'VerificationRunToolUnavailable',
+    'SourcePreimageUnavailable',
+  ],
 });
 const CAPABILITIES = new Set([
   'read',
@@ -196,10 +202,7 @@ function recognizeCapabilityGap(content, messages = []) {
   ) {
     return null;
   }
-  const echoed = messages.some(
-    (m) => m && m.role === 'tool' && String(m.content || '').includes(text)
-  );
-  if (echoed) {
+  if (echoedByTool(text, messages)) {
     return null;
   }
   return {