@miller-tech/uap 1.76.6 → 1.77.0
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/dist/.tsbuildinfo +1 -1
- package/dist/coordination/reactor.d.ts.map +1 -1
- package/dist/coordination/reactor.js +24 -0
- package/dist/coordination/reactor.js.map +1 -1
- package/package.json +1 -1
- package/src/policies/enforcers/__pycache__/_common.cpython-312.pyc +0 -0
- package/src/policies/enforcers/delivery_enforcement.py +10 -4
- package/tools/agents/scripts/__pycache__/toolcall_path_normalizer.cpython-312.pyc +0 -0
- package/tools/agents/scripts/anthropic_proxy.py +31 -3
- package/tools/agents/tests/test_coordination_ban.py +68 -0
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"reactor.d.ts","sourceRoot":"","sources":["../../src/coordination/reactor.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,gBAAgB,EAAuB,MAAM,wBAAwB,CAAC;AAC/E,OAAO,EAAE,aAAa,EAAoB,MAAM,qBAAqB,CAAC;AAKtE,MAAM,MAAM,YAAY,GAAG,eAAe,GAAG,aAAa,GAAG,UAAU,GAAG,WAAW,GAAG,MAAM,GAAG,aAAa,CAAC;AAE/G,MAAM,WAAW,cAAc;IAC7B,KAAK,EAAE,YAAY,CAAC;IACpB,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,YAAY,CAAC,EAAE,MAAM,EAAE,CAAC;IACxB,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,GAAG,CAAC,EAAE,MAAM,CAAC;IACb,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,QAAQ,CAAC,EAAE,MAAM,EAAE,CAAC;CACrB;AAED,MAAM,WAAW,aAAa;IAC5B,IAAI,EAAE,cAAc,GAAG,eAAe,GAAG,SAAS,GAAG,WAAW,CAAC;IACjE,MAAM,EAAE,MAAM,CAAC;IACf,UAAU,EAAE,MAAM,CAAC;IACnB,MAAM,EAAE,MAAM,CAAC;CAChB;AAED,MAAM,WAAW,aAAa;IAC5B,MAAM,EAAE,MAAM,CAAC;IACf,KAAK,EAAE,OAAO,CAAC;IACf,MAAM,EAAE,MAAM,CAAC;IACf,OAAO,EAAE,aAAa,EAAE,CAAC;IACzB,YAAY,EAAE,MAAM,EAAE,CAAC;IACvB,UAAU,EAAE,MAAM,CAAC;CACpB;AAED,MAAM,WAAW,cAAc;IAC7B,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,kBAAkB,CAAC,EAAE,MAAM,EAAE,CAAC;IAC9B,cAAc,CAAC,EAAE,MAAM,CAAC;CACzB;AAED,MAAM,WAAW,WAAW;IAC1B,gBAAgB,CAAC,EAAE,gBAAgB,CAAC;IACpC,aAAa,CAAC,EAAE,aAAa,CAAC;CAC/B;
|
|
1
|
+
{"version":3,"file":"reactor.d.ts","sourceRoot":"","sources":["../../src/coordination/reactor.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,gBAAgB,EAAuB,MAAM,wBAAwB,CAAC;AAC/E,OAAO,EAAE,aAAa,EAAoB,MAAM,qBAAqB,CAAC;AAKtE,MAAM,MAAM,YAAY,GAAG,eAAe,GAAG,aAAa,GAAG,UAAU,GAAG,WAAW,GAAG,MAAM,GAAG,aAAa,CAAC;AAE/G,MAAM,WAAW,cAAc;IAC7B,KAAK,EAAE,YAAY,CAAC;IACpB,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,YAAY,CAAC,EAAE,MAAM,EAAE,CAAC;IACxB,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,GAAG,CAAC,EAAE,MAAM,CAAC;IACb,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,QAAQ,CAAC,EAAE,MAAM,EAAE,CAAC;CACrB;AAED,MAAM,WAAW,aAAa;IAC5B,IAAI,EAAE,cAAc,GAAG,eAAe,GAAG,SAAS,GAAG,WAAW,CAAC;IACjE,MAAM,EAAE,MAAM,CAAC;IACf,UAAU,EAAE,MAAM,CAAC;IACnB,MAAM,EAAE,MAAM,CAAC;CAChB;AAED,MAAM,WAAW,aAAa;IAC5B,MAAM,EAAE,MAAM,CAAC;IACf,KAAK,EAAE,OAAO,CAAC;IACf,MAAM,EAAE,MAAM,CAAC;IACf,OAAO,EAAE,aAAa,EAAE,CAAC;IACzB,YAAY,EAAE,MAAM,EAAE,CAAC;IACvB,UAAU,EAAE,MAAM,CAAC;CACpB;AAED,MAAM,WAAW,cAAc;IAC7B,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,kBAAkB,CAAC,EAAE,MAAM,EAAE,CAAC;IAC9B,cAAc,CAAC,EAAE,MAAM,CAAC;CACzB;AAED,MAAM,WAAW,WAAW;IAC1B,gBAAgB,CAAC,EAAE,gBAAgB,CAAC;IACpC,aAAa,CAAC,EAAE,aAAa,CAAC;CAC/B;AAuBD,wBAAgB,OAAO,CACrB,GAAG,EAAE,cAAc,EACnB,IAAI,CAAC,EAAE,cAAc,EACrB,IAAI,CAAC,EAAE,WAAW,GACjB,aAAa,CAqOf"}
|
|
@@ -6,6 +6,14 @@ import { maybeCollaborationInjection } from './collaboration-inject.js';
|
|
|
6
6
|
const DEFAULT_INJECT_THRESHOLD = 0.30;
|
|
7
7
|
const DEFAULT_AUTO_SPAWN_THRESHOLD = 0.80;
|
|
8
8
|
const DEFAULT_MAX_INJECT_CHARS = 1200;
|
|
9
|
+
// #3-B: capabilities that imply writing/modifying source. When a task routes to
|
|
10
|
+
// one of these, the reactor proactively tells the agent to route file writes
|
|
11
|
+
// through `uap deliver` BEFORE it hits the (blocking) delivery-enforcement gate
|
|
12
|
+
// — weak local models otherwise retry the blocked edit or hallucinate completion.
|
|
13
|
+
const CODE_CAPABILITIES = new Set([
|
|
14
|
+
'typescript', 'javascript', 'python', 'rust', 'go', 'cpp', 'java',
|
|
15
|
+
'frontend', 'backend', 'cli', 'api-design',
|
|
16
|
+
]);
|
|
9
17
|
export function resolve(ctx, opts, deps) {
|
|
10
18
|
const injectThreshold = opts?.injectThreshold ?? DEFAULT_INJECT_THRESHOLD;
|
|
11
19
|
const autoSpawnThreshold = opts?.autoSpawnThreshold ?? DEFAULT_AUTO_SPAWN_THRESHOLD;
|
|
@@ -59,10 +67,26 @@ export function resolve(ctx, opts, deps) {
|
|
|
59
67
|
const collabInject = ctx.cwd && !(ctx.surfaced ?? []).includes(collabKey)
|
|
60
68
|
? maybeCollaborationInjection(ctx.cwd, ctx.promptText)
|
|
61
69
|
: null;
|
|
70
|
+
// #3-B: delivery routing. Surface up-front (deduped per session via
|
|
71
|
+
// `deliver:routing`) whenever the task routes to a code capability, so the
|
|
72
|
+
// agent calls deliver instead of editing source directly and tripping the gate.
|
|
73
|
+
const deliverKey = 'deliver:routing';
|
|
74
|
+
const isCodeTask = (routeResult.matchedCapabilities ?? []).some((c) => CODE_CAPABILITIES.has(c));
|
|
75
|
+
const deliverInject = isCodeTask && !(ctx.surfaced ?? []).includes(deliverKey)
|
|
76
|
+
? 'This task creates or modifies source code. Use the `deliver` tool ' +
|
|
77
|
+
'(or run `uap deliver "<one-line description>"`) to write code — direct ' +
|
|
78
|
+
'Edit/Write on source files is gated and will be blocked. Deliver writes ' +
|
|
79
|
+
'the files and verifies them against the gates; do not report completion ' +
|
|
80
|
+
'until deliver reports success.'
|
|
81
|
+
: null;
|
|
62
82
|
// Standalone context blocks (design + board) ride even on low-confidence turns.
|
|
63
83
|
const buildContextBlocks = () => {
|
|
64
84
|
const blocks = [];
|
|
65
85
|
const keys = [];
|
|
86
|
+
if (deliverInject) {
|
|
87
|
+
blocks.push(`## Writing code — route through deliver\n${deliverInject}`);
|
|
88
|
+
keys.push(deliverKey);
|
|
89
|
+
}
|
|
66
90
|
if (collabInject) {
|
|
67
91
|
blocks.push(`## Agent collaboration\n${collabInject}`);
|
|
68
92
|
keys.push(collabKey);
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"reactor.js","sourceRoot":"","sources":["../../src/coordination/reactor.ts"],"names":[],"mappings":"AAAA,OAAO,EAAoB,mBAAmB,EAAE,MAAM,wBAAwB,CAAC;AAC/E,OAAO,EAAiB,gBAAgB,EAAE,MAAM,qBAAqB,CAAC;AACtE,OAAO,EAAE,oBAAoB,EAAE,MAAM,6BAA6B,CAAC;AACnE,OAAO,EAAE,mBAAmB,EAAE,MAAM,mBAAmB,CAAC;AACxD,OAAO,EAAE,2BAA2B,EAAE,MAAM,2BAA2B,CAAC;AA0CxE,MAAM,wBAAwB,GAAG,IAAI,CAAC;AACtC,MAAM,4BAA4B,GAAG,IAAI,CAAC;AAC1C,MAAM,wBAAwB,GAAG,IAAI,CAAC;
|
|
1
|
+
{"version":3,"file":"reactor.js","sourceRoot":"","sources":["../../src/coordination/reactor.ts"],"names":[],"mappings":"AAAA,OAAO,EAAoB,mBAAmB,EAAE,MAAM,wBAAwB,CAAC;AAC/E,OAAO,EAAiB,gBAAgB,EAAE,MAAM,qBAAqB,CAAC;AACtE,OAAO,EAAE,oBAAoB,EAAE,MAAM,6BAA6B,CAAC;AACnE,OAAO,EAAE,mBAAmB,EAAE,MAAM,mBAAmB,CAAC;AACxD,OAAO,EAAE,2BAA2B,EAAE,MAAM,2BAA2B,CAAC;AA0CxE,MAAM,wBAAwB,GAAG,IAAI,CAAC;AACtC,MAAM,4BAA4B,GAAG,IAAI,CAAC;AAC1C,MAAM,wBAAwB,GAAG,IAAI,CAAC;AAEtC,gFAAgF;AAChF,6EAA6E;AAC7E,gFAAgF;AAChF,kFAAkF;AAClF,MAAM,iBAAiB,GAAG,IAAI,GAAG,CAAS;IACxC,YAAY,EAAE,YAAY,EAAE,QAAQ,EAAE,MAAM,EAAE,IAAI,EAAE,KAAK,EAAE,MAAM;IACjE,UAAU,EAAE,SAAS,EAAE,KAAK,EAAE,YAAY;CAC3C,CAAC,CAAC;AAUH,MAAM,UAAU,OAAO,CACrB,GAAmB,EACnB,IAAqB,EACrB,IAAkB;IAElB,MAAM,eAAe,GAAG,IAAI,EAAE,eAAe,IAAI,wBAAwB,CAAC;IAC1E,MAAM,kBAAkB,GAAG,IAAI,EAAE,kBAAkB,IAAI,4BAA4B,CAAC;IACpF,MAAM,cAAc,GAAG,IAAI,EAAE,cAAc,IAAI,wBAAwB,CAAC;IACxE,MAAM,kBAAkB,GAAG,IAAI,EAAE,kBAAkB,IAAI,EAAE,CAAC;IAE1D,MAAM,gBAAgB,GAAG,IAAI,EAAE,gBAAgB,IAAI,mBAAmB,EAAE,CAAC;IACzE,MAAM,aAAa,GAAG,IAAI,EAAE,aAAa,IAAI,gBAAgB,EAAE,CAAC;IAEhE,MAAM,MAAM,GAAkB;QAC5B,MAAM,EAAE,EAAE;QACV,KAAK,EAAE,KAAK;QACZ,MAAM,EAAE,EAAE;QACV,OAAO,EAAE,EAAE;QACX,YAAY,EAAE,EAAE;QAChB,UAAU,EAAE,CAAC;KACd,CAAC;IAEF,IAAI,CAAC,GAAG,CAAC,UAAU,EAAE,CAAC;QACpB,OAAO,MAAM,CAAC;IAChB,CAAC;IAED,MAAM,WAAW,GAAG,gBAAgB,CAAC,SAAS,CAC5C;QACE,EAAE,EAAE,mBAAmB;QACvB,KAAK,EAAE,GAAG,CAAC,UAAU;QACrB,IAAI,EAAE,MAAM;QACZ,MAAM,EAAE,MAAM;QACd,QAAQ,EAAE,CAAC;QACX,MAAM,EAAE,EAAE;QACV,SAAS,EAAE,0BAA0B;QACrC,SAAS,EAAE,0BAA0B;KACtC,EACD,GAAG,CAAC,YAAY,IAAI,EAAE,CACvB,CAAC;IAEF,MAAM,UAAU,GAAG,WAAW,CAAC,UAAU,CAAC;IAE1C,MAAM,eAAe,GAAG,aAAa,CAAC,aAAa,CAAC,GAAG,CAAC,UAAU,CAAC,CAAC;IAEpE,uEAAuE;IACvE,8EAA8E;IAC9E,2EAA2E;IAC3E,6EAA6E;IAC7E,MAAM,SAAS,GAAG,eAAe,CAAC;IAClC,MAAM,YAAY,GAChB,GAAG,CAAC,GAAG,IAAI,CAAC,CAAC,GAAG,CAAC,QAAQ,IAAI,EAAE,CAAC,CAAC,QAAQ,CAAC,SAAS,CAAC;QAClD,CAAC,CAAC,oBAAoB,CAAC,GAAG,CAAC,GAAG,EAAE,GAAG,CAAC,UAAU,EAAE,GAAG,CAAC,YAAY,CAAC;QACjE,CAAC,CAAC,IAAI,CAAC;IAEX,4EAA4E;IAC5E,uEAAuE;IACvE,uEAAuE;IACvE,MAAM,QAAQ,GAAG,cAAc,CAAC;IAChC,MAAM,WAAW,GACf,GAAG,CAAC,GAAG,IAAI,CAAC,CAAC,GAAG,CAAC,QAAQ,IAAI,EAAE,CAAC,CAAC,QAAQ,CAAC,QAAQ,CAAC;QACjD,CAAC,CAAC,mBAAmB,CAAC,GAAG,CAAC,GAAG,CAAC;QAC9B,CAAC,CAAC,IAAI,CAAC;IAEX,2EAA2E;IAC3E,kEAAkE;IAClE,+EAA+E;IAC/E,2CAA2C;IAC3C,MAAM,SAAS,GAAG,eAAe,CAAC;IAClC,MAAM,YAAY,GAChB,GAAG,CAAC,GAAG,IAAI,CAAC,CAAC,GAAG,CAAC,QAAQ,IAAI,EAAE,CAAC,CAAC,QAAQ,CAAC,SAAS,CAAC;QAClD,CAAC,CAAC,2BAA2B,CAAC,GAAG,CAAC,GAAG,EAAE,GAAG,CAAC,UAAU,CAAC;QACtD,CAAC,CAAC,IAAI,CAAC;IAEX,oEAAoE;IACpE,2EAA2E;IAC3E,gFAAgF;IAChF,MAAM,UAAU,GAAG,iBAAiB,CAAC;IACrC,MAAM,UAAU,GAAG,CAAC,WAAW,CAAC,mBAAmB,IAAI,EAAE,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,iBAAiB,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC;IACjG,MAAM,aAAa,GACjB,UAAU,IAAI,CAAC,CAAC,GAAG,CAAC,QAAQ,IAAI,EAAE,CAAC,CAAC,QAAQ,CAAC,UAAU,CAAC;QACtD,CAAC,CAAC,oEAAoE;YACpE,yEAAyE;YACzE,0EAA0E;YAC1E,0EAA0E;YAC1E,gCAAgC;QAClC,CAAC,CAAC,IAAI,CAAC;IAEX,gFAAgF;IAChF,MAAM,kBAAkB,GAAG,GAAuC,EAAE;QAClE,MAAM,MAAM,GAAa,EAAE,CAAC;QAC5B,MAAM,IAAI,GAAa,EAAE,CAAC;QAC1B,IAAI,aAAa,EAAE,CAAC;YAClB,MAAM,CAAC,IAAI,CAAC,4CAA4C,aAAa,EAAE,CAAC,CAAC;YACzE,IAAI,CAAC,IAAI,CAAC,UAAU,CAAC,CAAC;QACxB,CAAC;QACD,IAAI,YAAY,EAAE,CAAC;YACjB,MAAM,CAAC,IAAI,CAAC,2BAA2B,YAAY,EAAE,CAAC,CAAC;YACvD,IAAI,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC;QACvB,CAAC;QACD,IAAI,WAAW,EAAE,CAAC;YAChB,MAAM,CAAC,IAAI,CAAC,2BAA2B,WAAW,EAAE,CAAC,CAAC;YACtD,IAAI,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC;QACtB,CAAC;QACD,IAAI,YAAY,EAAE,CAAC;YACjB,MAAM,CAAC,IAAI,CAAC,iCAAiC,YAAY,EAAE,CAAC,CAAC;YAC7D,IAAI,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC;QACvB,CAAC;QACD,OAAO,EAAE,MAAM,EAAE,MAAM,CAAC,IAAI,CAAC,MAAM,CAAC,EAAE,IAAI,EAAE,CAAC;IAC/C,CAAC,CAAC;IAEF,IAAI,UAAU,GAAG,eAAe,IAAI,CAAC,CAAC,eAAe,IAAI,eAAe,CAAC,MAAM,KAAK,CAAC,CAAC,EAAE,CAAC;QACvF,6EAA6E;QAC7E,MAAM,SAAS,GAAG,kBAAkB,EAAE,CAAC;QACvC,IAAI,SAAS,CAAC,MAAM,EAAE,CAAC;YACrB,MAAM,CAAC,MAAM,GAAG,SAAS,CAAC,MAAM,CAAC;YACjC,MAAM,CAAC,MAAM,GAAG,yCAAyC,CAAC;YAC1D,MAAM,CAAC,YAAY,GAAG,SAAS,CAAC,IAAI,CAAC;YACrC,MAAM,CAAC,UAAU,GAAG,UAAU,CAAC;QACjC,CAAC;QACD,OAAO,MAAM,CAAC;IAChB,CAAC;IAED,MAAM,iBAAiB,GAAG,WAAW,CAAC,iBAAiB,IAAI,EAAE,CAAC;IAC9D,MAAM,iBAAiB,GAAG,WAAW,CAAC,iBAAiB,IAAI,EAAE,CAAC;IAE9D,MAAM,WAAW,GAAG,IAAI,GAAG,CAAC,GAAG,CAAC,QAAQ,IAAI,EAAE,CAAC,CAAC;IAEhD,MAAM,KAAK,GAAiB,EAAE,CAAC;IAE/B,wEAAwE;IACxE,yEAAyE;IACzE,kDAAkD;IAClD,IAAI,UAAU,IAAI,eAAe,EAAE,CAAC;QAClC,KAAK,MAAM,CAAC,IAAI,iBAAiB,EAAE,CAAC;YAClC,MAAM,GAAG,GAAG,SAAS,CAAC,EAAE,CAAC;YACzB,IAAI,CAAC,WAAW,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,CAAC;gBAC1B,KAAK,CAAC,IAAI,CAAC,EAAE,GAAG,EAAE,IAAI,EAAE,CAAC,EAAE,UAAU,EAAE,IAAI,EAAE,OAAO,EAAE,CAAC,CAAC;YAC1D,CAAC;QACH,CAAC;QAED,KAAK,MAAM,EAAE,IAAI,iBAAiB,EAAE,CAAC;YACnC,MAAM,GAAG,GAAG,SAAS,EAAE,EAAE,CAAC;YAC1B,IAAI,CAAC,WAAW,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,CAAC;gBAC1B,KAAK,CAAC,IAAI,CAAC,EAAE,GAAG,EAAE,IAAI,EAAE,EAAE,EAAE,UAAU,EAAE,IAAI,EAAE,OAAO,EAAE,CAAC,CAAC;YAC3D,CAAC;QACH,CAAC;IACH,CAAC;IAED,IAAI,eAAe,EAAE,CAAC;QACpB,KAAK,MAAM,CAAC,IAAI,eAAe,EAAE,CAAC;YAChC,MAAM,GAAG,GAAG,WAAW,CAAC,CAAC,YAAY,EAAE,CAAC;YACxC,IAAI,CAAC,WAAW,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,CAAC;gBAC1B,KAAK,CAAC,IAAI,CAAC,EAAE,GAAG,EAAE,IAAI,EAAE,CAAC,CAAC,YAAY,EAAE,UAAU,EAAE,IAAI,EAAE,SAAS,EAAE,KAAK,EAAE,CAAC,CAAC,KAAK,EAAE,CAAC,CAAC;YACzF,CAAC;QACH,CAAC;IACH,CAAC;IAED,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,UAAU,GAAG,CAAC,CAAC,UAAU,CAAC,CAAC;IAElD,IAAI,aAAa,GAAG,EAAE,CAAC;IACvB,MAAM,UAAU,GAAiB,EAAE,CAAC;IACpC,KAAK,MAAM,IAAI,IAAI,KAAK,EAAE,CAAC;QACzB,MAAM,IAAI,GAAG,IAAI,CAAC,IAAI,KAAK,SAAS;YAClC,CAAC,CAAC,SAAS,IAAI,CAAC,IAAI,OAAO,IAAI,CAAC,KAAK,IAAI,IAAI,CAAC,IAAI,iBAAiB,IAAI,CAAC,UAAU,CAAC,OAAO,CAAC,CAAC,CAAC,GAAG;YAChG,CAAC,CAAC,SAAS,IAAI,CAAC,IAAI,mBAAmB,IAAI,CAAC,UAAU,CAAC,OAAO,CAAC,CAAC,CAAC,GAAG,CAAC;QACvE,IAAI,aAAa,CAAC,MAAM,GAAG,IAAI,CAAC,MAAM,IAAI,cAAc,EAAE,CAAC;YACzD,aAAa,IAAI,IAAI,GAAG,IAAI,CAAC;YAC7B,UAAU,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;QACxB,CAAC;IACH,CAAC;IAED,MAAM,MAAM,GAAG,aAAa,CAAC,IAAI,EAAE,CAAC;IAEpC,MAAM,OAAO,GAAoB,EAAE,CAAC;IACpC,MAAM,YAAY,GAAa,EAAE,CAAC;IAElC,KAAK,MAAM,IAAI,IAAI,UAAU,EAAE,CAAC;QAC9B,YAAY,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;QAC5B,IAAI,IAAI,CAAC,IAAI,KAAK,OAAO,EAAE,CAAC;YAC1B,MAAM,WAAW,GAAG,kBAAkB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,IAAI,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC;YAC1E,IAAI,UAAU,IAAI,kBAAkB,IAAI,WAAW,EAAE,CAAC;gBACpD,OAAO,CAAC,IAAI,CAAC;oBACX,IAAI,EAAE,cAAc;oBACpB,MAAM,EAAE,IAAI,CAAC,IAAI;oBACjB,UAAU,EAAE,IAAI,CAAC,UAAU;oBAC3B,MAAM,EAAE,8BAA8B,IAAI,CAAC,IAAI,wBAAwB,UAAU,OAAO,kBAAkB,EAAE;iBAC7G,CAAC,CAAC;YACL,CAAC;iBAAM,CAAC;gBACN,OAAO,CAAC,IAAI,CAAC;oBACX,IAAI,EAAE,eAAe;oBACrB,MAAM,EAAE,IAAI,CAAC,IAAI;oBACjB,UAAU,EAAE,IAAI,CAAC,UAAU;oBAC3B,MAAM,EAAE,qBAAqB,IAAI,CAAC,IAAI,WAAW;iBAClD,CAAC,CAAC;YACL,CAAC;QACH,CAAC;aAAM,IAAI,IAAI,CAAC,IAAI,KAAK,OAAO,EAAE,CAAC;YACjC,OAAO,CAAC,IAAI,CAAC;gBACX,IAAI,EAAE,eAAe;gBACrB,MAAM,EAAE,IAAI,CAAC,IAAI;gBACjB,UAAU,EAAE,IAAI,CAAC,UAAU;gBAC3B,MAAM,EAAE,qBAAqB,IAAI,CAAC,IAAI,WAAW;aAClD,CAAC,CAAC;QACL,CAAC;aAAM,IAAI,IAAI,CAAC,IAAI,KAAK,SAAS,EAAE,CAAC;YACnC,OAAO,CAAC,IAAI,CAAC;gBACX,IAAI,EAAE,SAAS;gBACf,MAAM,EAAE,IAAI,CAAC,IAAI;gBACjB,UAAU,EAAE,IAAI,CAAC,UAAU;gBAC3B,MAAM,EAAE,WAAW,IAAI,CAAC,IAAI,UAAU;aACvC,CAAC,CAAC;QACL,CAAC;IACH,CAAC;IAED,4EAA4E;IAC5E,4BAA4B;IAC5B,MAAM,SAAS,GAAG,kBAAkB,EAAE,CAAC;IACvC,IAAI,SAAS,CAAC,MAAM,EAAE,CAAC;QACrB,MAAM,CAAC,MAAM,GAAG,GAAG,SAAS,CAAC,MAAM,GAAG,MAAM,CAAC,CAAC,CAAC,MAAM,GAAG,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC;QACtE,YAAY,CAAC,IAAI,CAAC,GAAG,SAAS,CAAC,IAAI,CAAC,CAAC;IACvC,CAAC;SAAM,CAAC;QACN,MAAM,CAAC,MAAM,GAAG,MAAM,CAAC;IACzB,CAAC;IACD,MAAM,CAAC,KAAK,GAAG,KAAK,CAAC;IACrB,MAAM,CAAC,MAAM;QACX,sBAAsB,UAAU,CAAC,OAAO,CAAC,CAAC,CAAC,EAAE;YAC7C,GAAG,eAAe,IAAI,eAAe,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,oBAAoB,CAAC,CAAC,CAAC,EAAE,EAAE;YAC9E,GAAG,YAAY,CAAC,CAAC,CAAC,2BAA2B,CAAC,CAAC,CAAC,EAAE,EAAE;YACpD,GAAG,WAAW,CAAC,CAAC,CAAC,kBAAkB,CAAC,CAAC,CAAC,EAAE,EAAE;YAC1C,GAAG,YAAY,CAAC,CAAC,CAAC,0BAA0B,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC;IACtD,MAAM,CAAC,OAAO,GAAG,OAAO,CAAC;IACzB,MAAM,CAAC,YAAY,GAAG,YAAY,CAAC;IACnC,MAAM,CAAC,UAAU,GAAG,UAAU,CAAC;IAE/B,OAAO,MAAM,CAAC;AAChB,CAAC"}
|
package/package.json
CHANGED
|
Binary file
|
|
@@ -85,11 +85,17 @@ def main() -> None:
|
|
|
85
85
|
if os.environ.get("UAP_DELIVER_BYPASS") == "1":
|
|
86
86
|
emit(True, "UAP_DELIVER_BYPASS override set")
|
|
87
87
|
|
|
88
|
+
# #3-F: terse, imperative, model-parseable. Weak local models otherwise
|
|
89
|
+
# retry the blocked edit or hallucinate completion ("the files exist") when
|
|
90
|
+
# the message is a passive explanation. State the exact next action.
|
|
88
91
|
msg = (
|
|
89
|
-
f"
|
|
90
|
-
"
|
|
91
|
-
"
|
|
92
|
-
"
|
|
92
|
+
f"BLOCKED: do not edit '{rel_posix}' directly. "
|
|
93
|
+
"To create or change code, call the `deliver` tool "
|
|
94
|
+
"(or run: uap deliver \"<one-line description of the change>\"). "
|
|
95
|
+
"Deliver writes the files and verifies them against the gates. "
|
|
96
|
+
"Do NOT retry this edit. Do NOT say the file is written until deliver "
|
|
97
|
+
"reports success. "
|
|
98
|
+
"(Sanctioned manual edit only: set UAP_DELIVER_BYPASS=1.)"
|
|
93
99
|
)
|
|
94
100
|
|
|
95
101
|
mode = os.environ.get("UAP_ENFORCE_DELIVERY", "block").lower()
|
|
Binary file
|
|
@@ -189,6 +189,23 @@ PROXY_TOOL_STATE_FINALIZE_THRESHOLD = int(
|
|
|
189
189
|
PROXY_TOOL_STATE_REVIEW_CYCLE_LIMIT = int(
|
|
190
190
|
os.environ.get("PROXY_TOOL_STATE_REVIEW_CYCLE_LIMIT", "3")
|
|
191
191
|
)
|
|
192
|
+
# #4 Coordination no-op suppression. Pure-bookkeeping tools (task/board status
|
|
193
|
+
# updates) are never productive to repeat, so they are banned after fewer cycle
|
|
194
|
+
# detections than the generic per-tool threshold. This targets the 35B-A3B
|
|
195
|
+
# "update the task instead of doing the actual work" loop (observed live:
|
|
196
|
+
# cycling_tools=['TaskUpdate']). Tunable via env; set the threshold to 0 to
|
|
197
|
+
# disable the faster ban and fall back to the generic threshold.
|
|
198
|
+
PROXY_COORDINATION_TOOLS = {
|
|
199
|
+
t.strip()
|
|
200
|
+
for t in os.environ.get(
|
|
201
|
+
"PROXY_COORDINATION_TOOLS",
|
|
202
|
+
"TaskUpdate,TaskCreate,TaskList,TaskGet,TaskOutput",
|
|
203
|
+
).split(",")
|
|
204
|
+
if t.strip()
|
|
205
|
+
}
|
|
206
|
+
PROXY_COORDINATION_BAN_THRESHOLD = int(
|
|
207
|
+
os.environ.get("PROXY_COORDINATION_BAN_THRESHOLD", "2")
|
|
208
|
+
)
|
|
192
209
|
# Force finalize after N consecutive forced_budget_exhausted events where
|
|
193
210
|
# neither cycling nor stagnation was detected — catches "distinct but
|
|
194
211
|
# unproductive" tool spam that defeats per-tool cycle detection.
|
|
@@ -3662,15 +3679,26 @@ def _resolve_state_machine_tool_choice(
|
|
|
3662
3679
|
for part in fp.split("|"):
|
|
3663
3680
|
raw_names.append(part.split(":")[0])
|
|
3664
3681
|
monitor.cycling_tool_names = list(dict.fromkeys(raw_names))
|
|
3665
|
-
# Cycle 18 Option 2: track per-tool cycle counts and ban after
|
|
3682
|
+
# Cycle 18 Option 2: track per-tool cycle counts and ban after N cycles.
|
|
3683
|
+
# #4: coordination/bookkeeping tools (TaskUpdate etc.) are banned
|
|
3684
|
+
# faster (PROXY_COORDINATION_BAN_THRESHOLD) since repeating them is
|
|
3685
|
+
# never productive; other tools keep the generic threshold of 3.
|
|
3666
3686
|
for name in monitor.cycling_tool_names:
|
|
3667
3687
|
monitor.tool_cycle_counts[name] = monitor.tool_cycle_counts.get(name, 0) + 1
|
|
3668
|
-
|
|
3688
|
+
is_coord = (
|
|
3689
|
+
name in PROXY_COORDINATION_TOOLS
|
|
3690
|
+
and PROXY_COORDINATION_BAN_THRESHOLD > 0
|
|
3691
|
+
)
|
|
3692
|
+
ban_at = PROXY_COORDINATION_BAN_THRESHOLD if is_coord else 3
|
|
3693
|
+
if monitor.tool_cycle_counts[name] >= ban_at and name not in monitor.session_banned_tools:
|
|
3669
3694
|
monitor.session_banned_tools.add(name)
|
|
3670
3695
|
logger.warning(
|
|
3671
|
-
"TOOL BAN: '%s' banned for session after %d cycle detections"
|
|
3696
|
+
"TOOL BAN: '%s' banned for session after %d cycle detections "
|
|
3697
|
+
"(threshold=%d%s)",
|
|
3672
3698
|
name,
|
|
3673
3699
|
monitor.tool_cycle_counts[name],
|
|
3700
|
+
ban_at,
|
|
3701
|
+
", coordination" if is_coord else "",
|
|
3674
3702
|
)
|
|
3675
3703
|
logger.warning(
|
|
3676
3704
|
"TOOL STATE MACHINE: entering review (cycle=%s repeat=%d stagnation=%d cycles=%d cycling_tools=%s)",
|
|
@@ -0,0 +1,68 @@
|
|
|
1
|
+
"""Tests for #4: coordination-tool no-op loop suppression (faster ban).
|
|
2
|
+
|
|
3
|
+
A pure-bookkeeping tool that cycles (e.g. TaskUpdate) is banned after
|
|
4
|
+
PROXY_COORDINATION_BAN_THRESHOLD (2) cycle detections, while a generic tool
|
|
5
|
+
keeps the original threshold of 3. Banning removes the tool from the offered
|
|
6
|
+
set, physically breaking the "update the task instead of doing the work" loop.
|
|
7
|
+
"""
|
|
8
|
+
import importlib.util
|
|
9
|
+
import unittest
|
|
10
|
+
from pathlib import Path
|
|
11
|
+
|
|
12
|
+
proxy_path = Path(__file__).resolve().parents[1] / "scripts" / "anthropic_proxy.py"
|
|
13
|
+
spec = importlib.util.spec_from_file_location("anthropic_proxy", proxy_path)
|
|
14
|
+
ap = importlib.util.module_from_spec(spec)
|
|
15
|
+
spec.loader.exec_module(ap)
|
|
16
|
+
|
|
17
|
+
|
|
18
|
+
def _cycling_body(tool="TaskUpdate"):
|
|
19
|
+
# >= PROXY_TOOL_STATE_MIN_MESSAGES (6) messages, last user message carries a
|
|
20
|
+
# tool_result -> active agentic loop.
|
|
21
|
+
msgs = [{"role": "user", "content": "do the task"}]
|
|
22
|
+
for i in range(3):
|
|
23
|
+
msgs.append({"role": "assistant", "content": [
|
|
24
|
+
{"type": "tool_use", "id": f"t{i}", "name": tool, "input": {}}]})
|
|
25
|
+
msgs.append({"role": "user", "content": [
|
|
26
|
+
{"type": "tool_result", "tool_use_id": f"t{i}", "content": "ok"}]})
|
|
27
|
+
return {"messages": msgs}
|
|
28
|
+
|
|
29
|
+
|
|
30
|
+
def _drive_one_cycle(monitor, tool):
|
|
31
|
+
"""Put the monitor in act-phase with a detected cycle of `tool` and run the
|
|
32
|
+
state-machine resolver once (which runs the ban loop)."""
|
|
33
|
+
fp = f"{tool}:deadbeef"
|
|
34
|
+
monitor.tool_turn_phase = "act"
|
|
35
|
+
monitor.tool_state_forced_budget_remaining = 5
|
|
36
|
+
monitor.tool_state_review_cycles = 0
|
|
37
|
+
win = max(2, ap.PROXY_TOOL_STATE_CYCLE_WINDOW)
|
|
38
|
+
monitor.tool_call_history = [fp] * (win + 2)
|
|
39
|
+
monitor.last_tool_fingerprint = fp
|
|
40
|
+
monitor.tool_state_stagnation_streak = max(1, ap.PROXY_TOOL_STATE_STAGNATION_THRESHOLD)
|
|
41
|
+
ap._resolve_state_machine_tool_choice(
|
|
42
|
+
_cycling_body(tool), monitor, has_tool_results=True, last_user_has_tool_result=True
|
|
43
|
+
)
|
|
44
|
+
|
|
45
|
+
|
|
46
|
+
class CoordinationBanTest(unittest.TestCase):
|
|
47
|
+
def test_config_defaults(self):
|
|
48
|
+
self.assertIn("TaskUpdate", ap.PROXY_COORDINATION_TOOLS)
|
|
49
|
+
self.assertEqual(ap.PROXY_COORDINATION_BAN_THRESHOLD, 2)
|
|
50
|
+
|
|
51
|
+
def test_coordination_tool_banned_after_two_cycles(self):
|
|
52
|
+
m = ap.SessionMonitor(context_window=132096)
|
|
53
|
+
_drive_one_cycle(m, "TaskUpdate")
|
|
54
|
+
self.assertNotIn("TaskUpdate", m.session_banned_tools, "should not ban after 1 cycle")
|
|
55
|
+
_drive_one_cycle(m, "TaskUpdate")
|
|
56
|
+
self.assertIn("TaskUpdate", m.session_banned_tools, "should ban after 2 cycles (coordination)")
|
|
57
|
+
|
|
58
|
+
def test_generic_tool_needs_three_cycles(self):
|
|
59
|
+
m = ap.SessionMonitor(context_window=132096)
|
|
60
|
+
for _ in range(2):
|
|
61
|
+
_drive_one_cycle(m, "glob")
|
|
62
|
+
self.assertNotIn("glob", m.session_banned_tools, "generic tool must NOT ban at 2")
|
|
63
|
+
_drive_one_cycle(m, "glob")
|
|
64
|
+
self.assertIn("glob", m.session_banned_tools, "generic tool bans at 3")
|
|
65
|
+
|
|
66
|
+
|
|
67
|
+
if __name__ == "__main__":
|
|
68
|
+
unittest.main()
|