@lazyingart/agintiflow 0.20.287 → 0.20.289
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.
- package/docs/state-of-the-art-agent-runtime.md +17 -0
- package/package.json +1 -1
- package/scripts/smoke-dynamic-step-budget.js +1252 -16
- package/scripts/smoke-local-failure-recovery.js +171 -3
- package/scripts/smoke-progressive-tool-selection.js +581 -30
- package/src/agent-runner.js +1430 -104
- package/src/local-failure-recovery.js +162 -9
- package/src/model-client.js +3 -1
- package/src/progressive-tool-selection.js +201 -24
- package/src/provider-handoff.js +86 -0
- package/src/research-routing.js +1 -1
- package/src/scs-evidence.js +79 -11
|
@@ -130,6 +130,13 @@ only after a categorized provider failure and must not replay verified side
|
|
|
130
130
|
effects. Codex and Claude remain explicit opt-in backends rather than hidden
|
|
131
131
|
fallbacks.
|
|
132
132
|
|
|
133
|
+
Timeout compaction also preserves goal-scoped deep-research completion. A
|
|
134
|
+
same-task repair of the immediately preceding report keeps ordinary file,
|
|
135
|
+
shell, and validation tools after a model switch instead of forcing another
|
|
136
|
+
research pass. A genuinely new research request does not inherit that marker
|
|
137
|
+
and still begins with the bounded `deep_research` workflow. Native function
|
|
138
|
+
calling and the text-tool fallback apply the same rule.
|
|
139
|
+
|
|
133
140
|
## Machine Host Protocol
|
|
134
141
|
|
|
135
142
|
The supported subprocess boundary is:
|
|
@@ -161,6 +168,16 @@ existing routine owned by the relevant project:
|
|
|
161
168
|
The agent selects, invokes, monitors, and verifies those routines. It does not
|
|
162
169
|
replace them with prompt-specific shell fragments.
|
|
163
170
|
|
|
171
|
+
## Verification Recovery
|
|
172
|
+
|
|
173
|
+
Failed project tests are authoritative evidence unless the current request
|
|
174
|
+
explicitly asks to change them. AgInTiFlow keeps the failing command and source
|
|
175
|
+
lineage, constrains repair writes to evidence-linked production files, and
|
|
176
|
+
rejects repeated attempts to weaken the test specification. If bounded
|
|
177
|
+
LocalLLM recovery exhausts its authenticated strong routes without producing a
|
|
178
|
+
passing verification, the same session can rebound once to its original
|
|
179
|
+
provider while preserving the goal, source context, and verifier evidence.
|
|
180
|
+
|
|
164
181
|
## Acceptance Gates
|
|
165
182
|
|
|
166
183
|
A primary-backend release is accepted only after all of these pass:
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "@lazyingart/agintiflow",
|
|
3
|
-
"version": "0.20.
|
|
3
|
+
"version": "0.20.289",
|
|
4
4
|
"type": "module",
|
|
5
5
|
"description": "AgInTiFlow is a project-aware agent workspace for hybrid wet-dry R&D, hardware-aware intelligence, software automation, and industrial workflows.",
|
|
6
6
|
"license": "Apache-2.0",
|