hablas-ai 2.2.0 → 2.2.2

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (3) hide show
  1. package/CHANGELOG.md +18 -18
  2. package/dist/index.js +329 -208
  3. package/package.json +1 -1
package/CHANGELOG.md CHANGED
@@ -1,25 +1,25 @@
1
1
  # Changelog
2
2
 
3
- ## [2.2.0] — 2026-06-12 · Stateful Runtime and Thinking Hooks
4
-
5
- ### Added
6
- - project-local `.hablas/session.json` persistence for workspace session state
7
- - larger default history and context budgets for the rebuilt runtime
8
- - engine hooks for model-start / model-stop so the UI can show real thinking states
3
+ ## [2.2.2] — 2026-06-12 · System Prompt Strength Rebuild
9
4
 
10
5
  ### Changed
11
- - direct and non-interactive sessions now use the configured `contextBudget` instead of falling back to the old small default
12
- - project sessions are automatically saved after turns and restored on the same workspace
13
- - web-search and large file context handling were tightened so useful output survives into model context better
14
-
15
- ### Fixed
16
- - `LARGE_WEB_RESULT_TOKENS is not defined` crash
17
- - thinking hooks are now wired at the engine level instead of existing only in the UI layer
18
- - repeated file reads are less likely to happen because full `read_file` context is preserved and cached more consistently
6
+ - Rebuilt the single-agent system prompt at a deeper architectural level.
7
+ - Replaced weaker phrase-driven prompt behavior with a stronger role / mission / behavior-class / tool-contract structure.
8
+ - Strengthened the prompt across all major execution domains:
9
+ - product reasoning
10
+ - architecture reasoning
11
+ - implementation
12
+ - debugging
13
+ - review
14
+ - devops thinking
15
+ - web and asset acquisition
16
+ - Strengthened tool discipline and anti-fake-completion rules.
17
+ - Reduced dependence on brittle literal examples and pushed the prompt toward generalized intent interpretation.
19
18
 
20
- ## [2.1.2] — 2026-06-12 · Full-Read Context Fix
19
+ ## [2.2.1] — 2026-06-12 · Runtime Persistence and Web-Search Crash Fix
21
20
 
22
21
  ### Fixed
23
- - `read_file` results are now preserved in full inside the model context instead of being aggressively truncated.
24
- - This prevents the agent from rereading the same large file just because it only saw the first chunk in context.
25
- - Tool result boxes now summarize the first meaningful line instead of flattening huge file bodies into unreadable one-liners.
22
+ - Restored the missing project-state import in the interactive runtime that caused `ensureProjectStateDir is not defined` at startup.
23
+ - Fixed the web-search follow-up crash path by restoring the missing large-web token constant and the associated output-preservation path.
24
+ - Project session state is now saved after slash commands as well as normal turns.
25
+ - Read-cache preservation is now actually written back after successful `read_file` operations and invalidated when related files change.