pantheon-opencode 1.6.2-beta.4 → 1.6.2-beta.6
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/CHANGELOG.md +27 -0
- package/README.md +202 -29
- package/README.pt-BR.md +9 -0
- package/bin/pantheon-init.mjs +6 -1
- package/package.json +2 -1
- package/plugin.json +1 -1
- package/pyproject.toml +1 -1
- package/scripts/decision-router-budget.mjs +7 -8
- package/scripts/decision-router-v2-budget.mjs +218 -0
- package/scripts/decision-router-v2-egress.mjs +272 -0
- package/scripts/decision-router.mjs +268 -46
- package/scripts/docs-check-changelog.mjs +5 -3
- package/scripts/install/decision-layer-preview.mjs +376 -0
- package/scripts/install/decision-router-config.mjs +162 -34
- package/scripts/mcp_resources.py +11 -9
- package/scripts/release-notes.mjs +30 -7
- package/scripts/versioning.mjs +41 -5
- package/src/mcp/eval_store.py +11 -4
- package/src/mcp/mcp_resources.py +11 -9
- package/src/mcp/memory_mcp.py +293 -150
- package/src/pantheon/jev-decision-tool.ts +336 -0
- package/src/pantheon/jev-decisions.ts +625 -0
- package/src/pantheon/local-decision-budget.ts +4 -1
- package/src/pantheon/local-decision-egress-gate.ts +16 -136
- package/src/pantheon/local-decision-prompt.ts +184 -45
- package/src/pantheon/local-decision-provider-matrix.ts +14 -62
- package/src/pantheon/local-decision-skills.ts +115 -5
- package/src/pantheon/local-decision-tool.ts +35 -1
- package/src/pantheon/local-decision-turn.ts +15 -252
- package/src/pantheon/local-decision-usage-evidence.internal.ts +80 -119
- package/src/pantheon/v2-decision-budget.ts +456 -0
- package/src/pantheon/v2-decision-contracts.ts +300 -0
- package/src/pantheon/v2-decision-egress.ts +203 -0
- package/src/pantheon/v2-decision-events.ts +180 -0
- package/src/pantheon/v2-decision-heuristic.ts +121 -0
- package/src/pantheon/v2-decision-points/p1-agent.ts +91 -0
- package/src/pantheon/v2-decision-points/p10-claims.ts +243 -0
- package/src/pantheon/v2-decision-points/p11-catalog.ts +220 -0
- package/src/pantheon/v2-decision-points/p2-skill.ts +83 -0
- package/src/pantheon/v2-decision-points/p3-complexity.ts +108 -0
- package/src/pantheon/v2-decision-points/p4-memory.ts +137 -0
- package/src/pantheon/v2-decision-points/p5-themis-verify.ts +190 -0
- package/src/pantheon/v2-decision-points/p6-council.ts +161 -0
- package/src/pantheon/v2-decision-points/p7-loop.ts +263 -0
- package/src/pantheon/v2-decision-points/p8-guardrail.ts +367 -0
- package/src/pantheon/v2-decision-points/p9-memory-compaction.ts +304 -0
- package/src/pantheon/v2-decision-points.ts +242 -0
- package/src/pantheon/v2-decision-policy.ts +166 -0
- package/src/pantheon/v2-decision-premises.ts +146 -0
- package/src/pantheon/v2-decision-prompt-surface.ts +115 -0
- package/src/pantheon/v2-decision-rank.ts +91 -0
- package/src/pantheon/v2-decision-registry.ts +391 -0
- package/src/pantheon/v2-decision-resource-ids.ts +127 -0
- package/src/pantheon/v2-decision-shadow.ts +123 -0
- package/src/pantheon/v2-decision-telemetry.ts +214 -0
- package/src/pantheon/v2-decision-verify.ts +51 -0
- package/src/pantheon/v2-native-decisions.ts +365 -0
- package/src/pantheon/v2-tools.ts +24 -2
- package/src/plugin-v2.ts +360 -14
- package/src/plugins/tui/dist/tui-v2.js +3 -3
- package/src/plugins/tui/dist/tui.js +1 -1
- package/src/plugins/tui/package-lock.json +2 -2
- package/src/plugins/tui/package.json +1 -1
- package/src/pantheon/local-decision-provider-protocol.ts +0 -335
- package/src/pantheon/local-decision-provider.internal.ts +0 -260
|
@@ -0,0 +1,214 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* V2 decision telemetry — allowlist sanitisation, closing defect D10.
|
|
3
|
+
*
|
|
4
|
+
* Telemetry must never carry prompt, response, user content, credentials or
|
|
5
|
+
* free-form text. The strongest way to guarantee that is not to scrub known
|
|
6
|
+
* keys but to accept only a small allowlist whose values are constrained to
|
|
7
|
+
* short tokens and bounded numbers:
|
|
8
|
+
*
|
|
9
|
+
* schemaVersion, point, decision, effect, reason, mode, durationMs, candidateCount
|
|
10
|
+
*
|
|
11
|
+
* Any other key is dropped and only a count is reported. Neither the key nor
|
|
12
|
+
* its value is ever copied into the result, so no attacker-controlled string
|
|
13
|
+
* reaches telemetry through a rejection path.
|
|
14
|
+
*
|
|
15
|
+
* `effect` is fixed to `'none'`. An event that claims anything else is rejected
|
|
16
|
+
* outright: this layer has no measured effect to report, and silently rewriting
|
|
17
|
+
* the claim would hide the caller's error.
|
|
18
|
+
*
|
|
19
|
+
* The sink is in-memory and bounded. It performs no I/O, so the offline harness
|
|
20
|
+
* can assert exactly what would be emitted.
|
|
21
|
+
*/
|
|
22
|
+
|
|
23
|
+
import type { DecisionPointId } from './v2-decision-registry.ts'
|
|
24
|
+
|
|
25
|
+
export const TELEMETRY_SCHEMA_VERSION = 1
|
|
26
|
+
export const TELEMETRY_FIELDS = [
|
|
27
|
+
'schemaVersion',
|
|
28
|
+
'point',
|
|
29
|
+
'decision',
|
|
30
|
+
'effect',
|
|
31
|
+
'reason',
|
|
32
|
+
'mode',
|
|
33
|
+
'durationMs',
|
|
34
|
+
'candidateCount',
|
|
35
|
+
] as const
|
|
36
|
+
|
|
37
|
+
export type TelemetryField = (typeof TELEMETRY_FIELDS)[number]
|
|
38
|
+
export type TelemetryMode = 'shadow' | 'active' | 'off' | 'migration'
|
|
39
|
+
|
|
40
|
+
export interface DecisionTelemetryEvent {
|
|
41
|
+
readonly schemaVersion: typeof TELEMETRY_SCHEMA_VERSION
|
|
42
|
+
readonly point: DecisionPointId
|
|
43
|
+
readonly decision: string
|
|
44
|
+
readonly effect: 'none'
|
|
45
|
+
readonly reason: string
|
|
46
|
+
readonly mode?: TelemetryMode
|
|
47
|
+
readonly durationMs?: number
|
|
48
|
+
readonly candidateCount?: number
|
|
49
|
+
}
|
|
50
|
+
|
|
51
|
+
export type TelemetryRejection = 'INVALID_INPUT' | 'MISSING_REQUIRED_FIELD' | 'EFFECT_NOT_MEASURED'
|
|
52
|
+
|
|
53
|
+
export type TelemetryResult =
|
|
54
|
+
| {
|
|
55
|
+
readonly ok: true
|
|
56
|
+
readonly event: DecisionTelemetryEvent
|
|
57
|
+
readonly droppedFieldCount: number
|
|
58
|
+
}
|
|
59
|
+
| {
|
|
60
|
+
readonly ok: false
|
|
61
|
+
readonly reasonCode: TelemetryRejection
|
|
62
|
+
readonly droppedFieldCount: number
|
|
63
|
+
}
|
|
64
|
+
|
|
65
|
+
const TOKEN = /^[A-Za-z0-9][A-Za-z0-9._:-]{0,63}$/
|
|
66
|
+
const POINT_ID = /^P(?:[1-9]|1[01])$/
|
|
67
|
+
const TELEMETRY_KEYS: ReadonlySet<string> = new Set(TELEMETRY_FIELDS)
|
|
68
|
+
const MODES: ReadonlySet<string> = new Set(['shadow', 'active', 'off', 'migration'])
|
|
69
|
+
const DURATION_MAX_MS = 60_000
|
|
70
|
+
const CANDIDATE_MAX = 10_000
|
|
71
|
+
function isRecord(value: unknown): value is Record<string, unknown> {
|
|
72
|
+
return typeof value === 'object' && value !== null && !Array.isArray(value)
|
|
73
|
+
}
|
|
74
|
+
|
|
75
|
+
function isToken(value: unknown): value is string {
|
|
76
|
+
return typeof value === 'string' && TOKEN.test(value)
|
|
77
|
+
}
|
|
78
|
+
|
|
79
|
+
function isPointId(value: unknown): value is DecisionPointId {
|
|
80
|
+
return typeof value === 'string' && POINT_ID.test(value)
|
|
81
|
+
}
|
|
82
|
+
|
|
83
|
+
function rejected(reasonCode: TelemetryRejection, droppedFieldCount: number): TelemetryResult {
|
|
84
|
+
return { ok: false, reasonCode, droppedFieldCount }
|
|
85
|
+
}
|
|
86
|
+
|
|
87
|
+
/**
|
|
88
|
+
* Reduce an arbitrary record to the telemetry allowlist.
|
|
89
|
+
*
|
|
90
|
+
* Required survivors: `point`, `decision` and `reason`. `effect`, when present,
|
|
91
|
+
* must already be `'none'`; the output always carries `'none'`.
|
|
92
|
+
*/
|
|
93
|
+
export function sanitizeTelemetry(input: unknown): TelemetryResult {
|
|
94
|
+
if (!isRecord(input)) return rejected('INVALID_INPUT', 0)
|
|
95
|
+
|
|
96
|
+
let droppedFieldCount = 0
|
|
97
|
+
const drop = (): void => {
|
|
98
|
+
droppedFieldCount += 1
|
|
99
|
+
}
|
|
100
|
+
|
|
101
|
+
let point: DecisionPointId | null = null
|
|
102
|
+
let decision: string | null = null
|
|
103
|
+
let reason: string | null = null
|
|
104
|
+
let mode: TelemetryMode | null = null
|
|
105
|
+
let durationMs: number | null = null
|
|
106
|
+
let candidateCount: number | null = null
|
|
107
|
+
let effectNotMeasured = false
|
|
108
|
+
|
|
109
|
+
for (const key of Object.keys(input).sort()) {
|
|
110
|
+
const value = input[key]
|
|
111
|
+
if (!TELEMETRY_KEYS.has(key)) {
|
|
112
|
+
drop()
|
|
113
|
+
continue
|
|
114
|
+
}
|
|
115
|
+
switch (key) {
|
|
116
|
+
case 'schemaVersion':
|
|
117
|
+
if (value !== TELEMETRY_SCHEMA_VERSION) drop()
|
|
118
|
+
break
|
|
119
|
+
case 'point':
|
|
120
|
+
if (isPointId(value)) point = value
|
|
121
|
+
else drop()
|
|
122
|
+
break
|
|
123
|
+
case 'decision':
|
|
124
|
+
if (isToken(value)) decision = value
|
|
125
|
+
else drop()
|
|
126
|
+
break
|
|
127
|
+
case 'reason':
|
|
128
|
+
if (isToken(value)) reason = value
|
|
129
|
+
else drop()
|
|
130
|
+
break
|
|
131
|
+
case 'mode':
|
|
132
|
+
if (typeof value === 'string' && MODES.has(value)) {
|
|
133
|
+
mode = value as TelemetryMode
|
|
134
|
+
} else {
|
|
135
|
+
drop()
|
|
136
|
+
}
|
|
137
|
+
break
|
|
138
|
+
case 'durationMs':
|
|
139
|
+
if (
|
|
140
|
+
typeof value === 'number' &&
|
|
141
|
+
Number.isFinite(value) &&
|
|
142
|
+
value >= 0 &&
|
|
143
|
+
value <= DURATION_MAX_MS
|
|
144
|
+
) {
|
|
145
|
+
durationMs = value
|
|
146
|
+
} else {
|
|
147
|
+
drop()
|
|
148
|
+
}
|
|
149
|
+
break
|
|
150
|
+
case 'candidateCount':
|
|
151
|
+
if (
|
|
152
|
+
Number.isInteger(value) &&
|
|
153
|
+
(value as number) >= 0 &&
|
|
154
|
+
(value as number) <= CANDIDATE_MAX
|
|
155
|
+
) {
|
|
156
|
+
candidateCount = value as number
|
|
157
|
+
} else {
|
|
158
|
+
drop()
|
|
159
|
+
}
|
|
160
|
+
break
|
|
161
|
+
case 'effect':
|
|
162
|
+
if (value === 'none') break
|
|
163
|
+
effectNotMeasured = true
|
|
164
|
+
drop()
|
|
165
|
+
break
|
|
166
|
+
default:
|
|
167
|
+
drop()
|
|
168
|
+
}
|
|
169
|
+
}
|
|
170
|
+
|
|
171
|
+
if (effectNotMeasured) return rejected('EFFECT_NOT_MEASURED', droppedFieldCount)
|
|
172
|
+
if (point === null || decision === null || reason === null) {
|
|
173
|
+
return rejected('MISSING_REQUIRED_FIELD', droppedFieldCount)
|
|
174
|
+
}
|
|
175
|
+
|
|
176
|
+
const event: DecisionTelemetryEvent = {
|
|
177
|
+
schemaVersion: TELEMETRY_SCHEMA_VERSION,
|
|
178
|
+
point,
|
|
179
|
+
decision,
|
|
180
|
+
effect: 'none',
|
|
181
|
+
reason,
|
|
182
|
+
...(mode !== null ? { mode } : {}),
|
|
183
|
+
...(durationMs !== null ? { durationMs } : {}),
|
|
184
|
+
...(candidateCount !== null ? { candidateCount } : {}),
|
|
185
|
+
}
|
|
186
|
+
return { ok: true, event: Object.freeze(event), droppedFieldCount }
|
|
187
|
+
}
|
|
188
|
+
|
|
189
|
+
/** In-memory, bounded telemetry collector. No I/O, no network. */
|
|
190
|
+
export interface TelemetrySink {
|
|
191
|
+
emit(input: unknown): TelemetryResult
|
|
192
|
+
events(): readonly DecisionTelemetryEvent[]
|
|
193
|
+
size(): number
|
|
194
|
+
}
|
|
195
|
+
|
|
196
|
+
export function createTelemetrySink(maxEvents = 1_000): TelemetrySink {
|
|
197
|
+
if (!Number.isInteger(maxEvents) || maxEvents < 0 || maxEvents > 100_000) {
|
|
198
|
+
throw new RangeError('maxEvents must be an integer in 0..100000')
|
|
199
|
+
}
|
|
200
|
+
const collected: DecisionTelemetryEvent[] = []
|
|
201
|
+
return {
|
|
202
|
+
emit(input: unknown) {
|
|
203
|
+
const result = sanitizeTelemetry(input)
|
|
204
|
+
if (result.ok && collected.length < maxEvents) collected.push(result.event)
|
|
205
|
+
return result
|
|
206
|
+
},
|
|
207
|
+
events() {
|
|
208
|
+
return Object.freeze([...collected])
|
|
209
|
+
},
|
|
210
|
+
size() {
|
|
211
|
+
return collected.length
|
|
212
|
+
},
|
|
213
|
+
}
|
|
214
|
+
}
|
|
@@ -0,0 +1,51 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* V2 verification contract — explicit verdicts, never inferred.
|
|
3
|
+
*
|
|
4
|
+
* `failed_to_respond` during mandatory verification is `BLOCK`; during advisory
|
|
5
|
+
* verification it is `FAIL`. Neither is ever `PASS`, and no scalar can stand in
|
|
6
|
+
* for a verdict.
|
|
7
|
+
*/
|
|
8
|
+
|
|
9
|
+
export const VERIFY_VERDICTS = ['PASS', 'FAIL', 'BLOCK'] as const
|
|
10
|
+
export type VerifyVerdict = (typeof VERIFY_VERDICTS)[number]
|
|
11
|
+
|
|
12
|
+
export interface VerifyRequirement {
|
|
13
|
+
readonly id: string
|
|
14
|
+
readonly mandatory: boolean
|
|
15
|
+
}
|
|
16
|
+
|
|
17
|
+
export type VerifyReasonCode = 'VERDICT' | 'FAILED_TO_RESPOND' | 'UNRECOGNIZED_VERDICT'
|
|
18
|
+
|
|
19
|
+
export interface VerifyOutcome {
|
|
20
|
+
readonly requirementId: string
|
|
21
|
+
readonly verdict: VerifyVerdict
|
|
22
|
+
readonly reasonCode: VerifyReasonCode
|
|
23
|
+
readonly mandatory: boolean
|
|
24
|
+
}
|
|
25
|
+
|
|
26
|
+
/**
|
|
27
|
+
* Resolve a verification verdict.
|
|
28
|
+
*
|
|
29
|
+
* The verdict must be stated explicitly. `failed_to_respond` during a
|
|
30
|
+
* mandatory verification is `BLOCK`; during an advisory verification it is
|
|
31
|
+
* `FAIL`. Neither case is ever `PASS`, and no certainty scalar can stand in
|
|
32
|
+
* for a verdict.
|
|
33
|
+
*/
|
|
34
|
+
export function resolveVerify(requirement: VerifyRequirement, response: unknown): VerifyOutcome {
|
|
35
|
+
const base = { requirementId: requirement.id, mandatory: requirement.mandatory }
|
|
36
|
+
if (typeof response === 'string') {
|
|
37
|
+
if (response === 'PASS' || response === 'FAIL' || response === 'BLOCK') {
|
|
38
|
+
return { ...base, verdict: response, reasonCode: 'VERDICT' }
|
|
39
|
+
}
|
|
40
|
+
return {
|
|
41
|
+
...base,
|
|
42
|
+
verdict: requirement.mandatory ? 'BLOCK' : 'FAIL',
|
|
43
|
+
reasonCode: 'UNRECOGNIZED_VERDICT',
|
|
44
|
+
}
|
|
45
|
+
}
|
|
46
|
+
return {
|
|
47
|
+
...base,
|
|
48
|
+
verdict: requirement.mandatory ? 'BLOCK' : 'FAIL',
|
|
49
|
+
reasonCode: 'FAILED_TO_RESPOND',
|
|
50
|
+
}
|
|
51
|
+
}
|
|
@@ -0,0 +1,365 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Offline scaffolding for OpenCode V2.2.5's native Decisions API.
|
|
3
|
+
*
|
|
4
|
+
* This module intentionally has no host call site. It resolves only model refs
|
|
5
|
+
* already present in the host model catalog, prepares a minimal current-turn
|
|
6
|
+
* request, and consumes a mocked/host response as non-authoritative advice.
|
|
7
|
+
* Nothing here calls a provider, reads credentials, or changes host behavior.
|
|
8
|
+
*/
|
|
9
|
+
import type { Model } from '@opencode/plugin'
|
|
10
|
+
import type { Context } from '@opencode/plugin/promise/plugin'
|
|
11
|
+
import type { SessionPrompt } from '@opencode/plugin/promise/session'
|
|
12
|
+
import { PANTHEON_AGENT_CATALOG } from './pantheon-decision-catalog.ts'
|
|
13
|
+
|
|
14
|
+
export const NATIVE_DECISION_ANSWER_NAME = 'specialist'
|
|
15
|
+
export const NATIVE_DECISION_MAX_PROMPT_LENGTH = 12_000
|
|
16
|
+
|
|
17
|
+
type NativeDecisionGenerator = Context['experimental']['generate']['decisions']
|
|
18
|
+
export type NativeDecisionRequest = Parameters<NativeDecisionGenerator>[0]
|
|
19
|
+
export type NativeDecisionResponse = Awaited<ReturnType<NativeDecisionGenerator>>
|
|
20
|
+
export type NativeDecisionModelRef = NativeDecisionRequest['model']
|
|
21
|
+
|
|
22
|
+
export const NATIVE_STOCK_DECISION_ROUTES = Object.freeze([
|
|
23
|
+
Object.freeze({
|
|
24
|
+
providerID: 'opencode',
|
|
25
|
+
id: 'jev-1.13',
|
|
26
|
+
kind: 'opencode-console-jev',
|
|
27
|
+
billing: 'OpenCode Console',
|
|
28
|
+
}),
|
|
29
|
+
Object.freeze({
|
|
30
|
+
providerID: 'openrouter',
|
|
31
|
+
id: 'typesafe/jev-1.13',
|
|
32
|
+
kind: 'openrouter-jev',
|
|
33
|
+
billing: 'OpenRouter',
|
|
34
|
+
}),
|
|
35
|
+
] as const)
|
|
36
|
+
|
|
37
|
+
type NativeDecisionRouteKind =
|
|
38
|
+
| (typeof NATIVE_STOCK_DECISION_ROUTES)[number]['kind']
|
|
39
|
+
| 'custom-decision'
|
|
40
|
+
|
|
41
|
+
export interface ResolvedNativeDecisionRoute {
|
|
42
|
+
readonly model: NativeDecisionModelRef
|
|
43
|
+
readonly kind: NativeDecisionRouteKind
|
|
44
|
+
readonly billing: string
|
|
45
|
+
}
|
|
46
|
+
|
|
47
|
+
export type NativeDecisionRouteFailure =
|
|
48
|
+
| 'AUTO_MODEL_UNAVAILABLE'
|
|
49
|
+
| 'INVALID_MODEL_REF'
|
|
50
|
+
| 'DIRECT_OPENAI_UNSUPPORTED'
|
|
51
|
+
| 'MODEL_NOT_IN_CATALOG'
|
|
52
|
+
| 'AMBIGUOUS_MODEL'
|
|
53
|
+
| 'MODEL_NOT_DECISION'
|
|
54
|
+
|
|
55
|
+
export type NativeDecisionRouteResolution =
|
|
56
|
+
| { readonly ok: true; readonly route: ResolvedNativeDecisionRoute }
|
|
57
|
+
| { readonly ok: false; readonly reasonCode: NativeDecisionRouteFailure }
|
|
58
|
+
|
|
59
|
+
export type NativeDecisionSelection = 'auto' | NativeDecisionModelRef
|
|
60
|
+
|
|
61
|
+
type DecisionModelCatalogEntry = Pick<Model.Info, 'providerID' | 'id' | 'type'>
|
|
62
|
+
|
|
63
|
+
function readModelRef(value: unknown): NativeDecisionModelRef | null {
|
|
64
|
+
if (value === null || typeof value !== 'object' || Array.isArray(value)) return null
|
|
65
|
+
const record = value as Record<string, unknown>
|
|
66
|
+
if (
|
|
67
|
+
typeof record.providerID !== 'string' ||
|
|
68
|
+
record.providerID.length === 0 ||
|
|
69
|
+
record.providerID.trim() !== record.providerID ||
|
|
70
|
+
typeof record.id !== 'string' ||
|
|
71
|
+
record.id.length === 0 ||
|
|
72
|
+
record.id.trim() !== record.id ||
|
|
73
|
+
(record.variant !== undefined && typeof record.variant !== 'string')
|
|
74
|
+
) {
|
|
75
|
+
return null
|
|
76
|
+
}
|
|
77
|
+
return {
|
|
78
|
+
providerID: record.providerID,
|
|
79
|
+
id: record.id,
|
|
80
|
+
...(record.variant === undefined ? {} : { variant: record.variant }),
|
|
81
|
+
}
|
|
82
|
+
}
|
|
83
|
+
|
|
84
|
+
function findRouteKind(
|
|
85
|
+
model: NativeDecisionModelRef,
|
|
86
|
+
): Pick<ResolvedNativeDecisionRoute, 'kind' | 'billing'> {
|
|
87
|
+
const stock = NATIVE_STOCK_DECISION_ROUTES.find(
|
|
88
|
+
(candidate) => candidate.providerID === model.providerID && candidate.id === model.id,
|
|
89
|
+
)
|
|
90
|
+
if (stock !== undefined) return { kind: stock.kind, billing: stock.billing }
|
|
91
|
+
if (model.providerID === 'openrouter') {
|
|
92
|
+
return { kind: 'custom-decision', billing: 'OpenRouter' }
|
|
93
|
+
}
|
|
94
|
+
if (model.providerID === 'opencode') {
|
|
95
|
+
return { kind: 'custom-decision', billing: 'OpenCode Console' }
|
|
96
|
+
}
|
|
97
|
+
return { kind: 'custom-decision', billing: 'Provider configured in OpenCode' }
|
|
98
|
+
}
|
|
99
|
+
|
|
100
|
+
/**
|
|
101
|
+
* Resolve Auto strictly against the exact effective session model. It never
|
|
102
|
+
* searches for a substitute, changes providers, or infers decision capability
|
|
103
|
+
* from a chat model's provider or id.
|
|
104
|
+
*/
|
|
105
|
+
export function resolveNativeDecisionRoute(
|
|
106
|
+
selection: NativeDecisionSelection,
|
|
107
|
+
effectiveSessionModel: unknown,
|
|
108
|
+
modelCatalog: readonly DecisionModelCatalogEntry[],
|
|
109
|
+
): NativeDecisionRouteResolution {
|
|
110
|
+
const requested =
|
|
111
|
+
selection === 'auto' ? readModelRef(effectiveSessionModel) : readModelRef(selection)
|
|
112
|
+
if (selection === 'auto' && requested === null) {
|
|
113
|
+
return { ok: false, reasonCode: 'AUTO_MODEL_UNAVAILABLE' }
|
|
114
|
+
}
|
|
115
|
+
if (requested === null) return { ok: false, reasonCode: 'INVALID_MODEL_REF' }
|
|
116
|
+
if (requested.providerID === 'openai') {
|
|
117
|
+
return { ok: false, reasonCode: 'DIRECT_OPENAI_UNSUPPORTED' }
|
|
118
|
+
}
|
|
119
|
+
|
|
120
|
+
const matches = modelCatalog.filter(
|
|
121
|
+
(model) => model.providerID === requested.providerID && model.id === requested.id,
|
|
122
|
+
)
|
|
123
|
+
if (matches.length === 0) return { ok: false, reasonCode: 'MODEL_NOT_IN_CATALOG' }
|
|
124
|
+
if (matches.length !== 1) return { ok: false, reasonCode: 'AMBIGUOUS_MODEL' }
|
|
125
|
+
if (matches[0]?.type !== 'decision') {
|
|
126
|
+
return { ok: false, reasonCode: 'MODEL_NOT_DECISION' }
|
|
127
|
+
}
|
|
128
|
+
|
|
129
|
+
return {
|
|
130
|
+
ok: true,
|
|
131
|
+
route: Object.freeze({ model: requested, ...findRouteKind(requested) }),
|
|
132
|
+
}
|
|
133
|
+
}
|
|
134
|
+
|
|
135
|
+
export interface NativeDecisionConsentEntry {
|
|
136
|
+
readonly providerID: string
|
|
137
|
+
readonly modelID: string
|
|
138
|
+
readonly enabled: boolean
|
|
139
|
+
}
|
|
140
|
+
|
|
141
|
+
export type NativeDecisionConsentResult =
|
|
142
|
+
| { readonly allowed: true; readonly reasonCode: 'EXACT_ROUTE_CONSENT' }
|
|
143
|
+
| {
|
|
144
|
+
readonly allowed: false
|
|
145
|
+
readonly reasonCode: 'CONSENT_MISSING' | 'CONSENT_AMBIGUOUS' | 'CONSENT_DISABLED'
|
|
146
|
+
}
|
|
147
|
+
|
|
148
|
+
/** Pure provider+model-specific consent check. Missing consent is always off. */
|
|
149
|
+
export function evaluateNativeDecisionConsent(
|
|
150
|
+
route: ResolvedNativeDecisionRoute,
|
|
151
|
+
entries: readonly NativeDecisionConsentEntry[] = [],
|
|
152
|
+
): NativeDecisionConsentResult {
|
|
153
|
+
const exact = entries.filter(
|
|
154
|
+
(entry) => entry.providerID === route.model.providerID && entry.modelID === route.model.id,
|
|
155
|
+
)
|
|
156
|
+
if (exact.length === 0) return { allowed: false, reasonCode: 'CONSENT_MISSING' }
|
|
157
|
+
if (exact.length !== 1) return { allowed: false, reasonCode: 'CONSENT_AMBIGUOUS' }
|
|
158
|
+
if (exact[0]?.enabled !== true) return { allowed: false, reasonCode: 'CONSENT_DISABLED' }
|
|
159
|
+
return { allowed: true, reasonCode: 'EXACT_ROUTE_CONSENT' }
|
|
160
|
+
}
|
|
161
|
+
|
|
162
|
+
/**
|
|
163
|
+
* Evidence required to bind a prompt callback to exactly one current turn.
|
|
164
|
+
* OpenCode V2.2.5's SessionPrompt event does not supply these queue/activity
|
|
165
|
+
* facts, so this contract is currently test-only and cannot authorize a call.
|
|
166
|
+
*/
|
|
167
|
+
export interface NativeDecisionCorrelationEvidence {
|
|
168
|
+
readonly sessionID: string
|
|
169
|
+
readonly messageID: string
|
|
170
|
+
readonly currentMessageID: string
|
|
171
|
+
readonly delivery: 'steer' | 'queue'
|
|
172
|
+
readonly pendingMessageIDs: readonly string[]
|
|
173
|
+
readonly activeMessageID: string | null
|
|
174
|
+
readonly continuation: boolean
|
|
175
|
+
}
|
|
176
|
+
|
|
177
|
+
export type NativeDecisionCorrelationFailure =
|
|
178
|
+
| 'CORRELATION_UNAVAILABLE'
|
|
179
|
+
| 'CORRELATION_MISMATCH'
|
|
180
|
+
| 'QUEUED_TURN'
|
|
181
|
+
| 'CONCURRENT_TURN'
|
|
182
|
+
| 'CONTINUATION_TURN'
|
|
183
|
+
|
|
184
|
+
function isNonBlankID(value: unknown): value is string {
|
|
185
|
+
return typeof value === 'string' && value.trim().length > 0
|
|
186
|
+
}
|
|
187
|
+
|
|
188
|
+
export function validateNativeDecisionCorrelation(
|
|
189
|
+
event: SessionPrompt,
|
|
190
|
+
evidence: NativeDecisionCorrelationEvidence | undefined,
|
|
191
|
+
): NativeDecisionCorrelationFailure | null {
|
|
192
|
+
if (event === null || typeof event !== 'object' || Array.isArray(event)) {
|
|
193
|
+
return 'CORRELATION_MISMATCH'
|
|
194
|
+
}
|
|
195
|
+
if (!isNonBlankID(event.sessionID) || !isNonBlankID(event.messageID)) {
|
|
196
|
+
return 'CORRELATION_MISMATCH'
|
|
197
|
+
}
|
|
198
|
+
if (evidence === undefined) return 'CORRELATION_UNAVAILABLE'
|
|
199
|
+
if (evidence === null || typeof evidence !== 'object' || Array.isArray(evidence)) {
|
|
200
|
+
return 'CORRELATION_UNAVAILABLE'
|
|
201
|
+
}
|
|
202
|
+
if (
|
|
203
|
+
!isNonBlankID(evidence.sessionID) ||
|
|
204
|
+
!isNonBlankID(evidence.messageID) ||
|
|
205
|
+
!isNonBlankID(evidence.currentMessageID) ||
|
|
206
|
+
(evidence.delivery !== 'steer' && evidence.delivery !== 'queue') ||
|
|
207
|
+
!Array.isArray(evidence.pendingMessageIDs) ||
|
|
208
|
+
evidence.pendingMessageIDs.some((id) => !isNonBlankID(id)) ||
|
|
209
|
+
(evidence.activeMessageID !== null && !isNonBlankID(evidence.activeMessageID)) ||
|
|
210
|
+
typeof evidence.continuation !== 'boolean'
|
|
211
|
+
) {
|
|
212
|
+
return 'CORRELATION_UNAVAILABLE'
|
|
213
|
+
}
|
|
214
|
+
if (
|
|
215
|
+
event.sessionID !== evidence.sessionID ||
|
|
216
|
+
event.messageID !== evidence.messageID ||
|
|
217
|
+
event.messageID !== evidence.currentMessageID
|
|
218
|
+
) {
|
|
219
|
+
return 'CORRELATION_MISMATCH'
|
|
220
|
+
}
|
|
221
|
+
if (event.delivery !== 'steer' || evidence.delivery !== 'steer') return 'QUEUED_TURN'
|
|
222
|
+
if (
|
|
223
|
+
evidence.activeMessageID !== null ||
|
|
224
|
+
evidence.pendingMessageIDs.length !== 1 ||
|
|
225
|
+
evidence.pendingMessageIDs[0] !== event.messageID
|
|
226
|
+
) {
|
|
227
|
+
return 'CONCURRENT_TURN'
|
|
228
|
+
}
|
|
229
|
+
if (evidence.continuation) return 'CONTINUATION_TURN'
|
|
230
|
+
return null
|
|
231
|
+
}
|
|
232
|
+
|
|
233
|
+
const DECISION_AGENT_IDS = new Set(
|
|
234
|
+
PANTHEON_AGENT_CATALOG.filter((agent) => agent.id !== 'zeus').map((agent) => agent.id),
|
|
235
|
+
)
|
|
236
|
+
|
|
237
|
+
export type NativeDecisionPreparationFailure =
|
|
238
|
+
| NativeDecisionRouteFailure
|
|
239
|
+
| NativeDecisionCorrelationFailure
|
|
240
|
+
| 'CONSENT_MISSING'
|
|
241
|
+
| 'CONSENT_AMBIGUOUS'
|
|
242
|
+
| 'CONSENT_DISABLED'
|
|
243
|
+
| 'PROMPT_EMPTY'
|
|
244
|
+
| 'PROMPT_TOO_LONG'
|
|
245
|
+
| 'FILES_NOT_ALLOWED'
|
|
246
|
+
| 'INVALID_CANDIDATES'
|
|
247
|
+
|
|
248
|
+
export type NativeDecisionPreparation =
|
|
249
|
+
| { readonly ok: true; readonly request: NativeDecisionRequest }
|
|
250
|
+
| { readonly ok: false; readonly reasonCode: NativeDecisionPreparationFailure }
|
|
251
|
+
|
|
252
|
+
/**
|
|
253
|
+
* Prepare only the current user text and a fixed question. This function does
|
|
254
|
+
* not invoke the native API and does not accept history, tools, paths, files,
|
|
255
|
+
* system instructions, memory, or catalog descriptions as inputs.
|
|
256
|
+
*/
|
|
257
|
+
export function prepareNativeDecisionRequest(
|
|
258
|
+
route: ResolvedNativeDecisionRoute,
|
|
259
|
+
event: SessionPrompt,
|
|
260
|
+
eligibleAgentIDs: readonly string[],
|
|
261
|
+
consentEntries: readonly NativeDecisionConsentEntry[] = [],
|
|
262
|
+
correlationEvidence?: NativeDecisionCorrelationEvidence,
|
|
263
|
+
): NativeDecisionPreparation {
|
|
264
|
+
if (
|
|
265
|
+
event === null ||
|
|
266
|
+
typeof event !== 'object' ||
|
|
267
|
+
event.prompt === null ||
|
|
268
|
+
typeof event.prompt !== 'object'
|
|
269
|
+
) {
|
|
270
|
+
return { ok: false, reasonCode: 'CORRELATION_MISMATCH' }
|
|
271
|
+
}
|
|
272
|
+
const consent = evaluateNativeDecisionConsent(route, consentEntries)
|
|
273
|
+
if (!consent.allowed) return { ok: false, reasonCode: consent.reasonCode }
|
|
274
|
+
|
|
275
|
+
const correlationFailure = validateNativeDecisionCorrelation(event, correlationEvidence)
|
|
276
|
+
if (correlationFailure !== null) return { ok: false, reasonCode: correlationFailure }
|
|
277
|
+
|
|
278
|
+
const promptFiles = (event.prompt as { files?: unknown }).files
|
|
279
|
+
if (promptFiles !== undefined && (!Array.isArray(promptFiles) || promptFiles.length > 0)) {
|
|
280
|
+
return { ok: false, reasonCode: 'FILES_NOT_ALLOWED' }
|
|
281
|
+
}
|
|
282
|
+
const promptText = event.prompt.text
|
|
283
|
+
if (typeof promptText !== 'string' || promptText.trim().length === 0) {
|
|
284
|
+
return { ok: false, reasonCode: 'PROMPT_EMPTY' }
|
|
285
|
+
}
|
|
286
|
+
if (promptText.length > NATIVE_DECISION_MAX_PROMPT_LENGTH) {
|
|
287
|
+
return { ok: false, reasonCode: 'PROMPT_TOO_LONG' }
|
|
288
|
+
}
|
|
289
|
+
|
|
290
|
+
const candidates = new Set<string>()
|
|
291
|
+
for (const id of eligibleAgentIDs) {
|
|
292
|
+
if (typeof id !== 'string' || !DECISION_AGENT_IDS.has(id) || candidates.has(id)) {
|
|
293
|
+
return { ok: false, reasonCode: 'INVALID_CANDIDATES' }
|
|
294
|
+
}
|
|
295
|
+
candidates.add(id)
|
|
296
|
+
}
|
|
297
|
+
if (candidates.size > DECISION_AGENT_IDS.size) {
|
|
298
|
+
return { ok: false, reasonCode: 'INVALID_CANDIDATES' }
|
|
299
|
+
}
|
|
300
|
+
|
|
301
|
+
const request = {
|
|
302
|
+
model: route.model,
|
|
303
|
+
input: [{ type: 'text', text: promptText }],
|
|
304
|
+
questions: [
|
|
305
|
+
{
|
|
306
|
+
type: 'choice',
|
|
307
|
+
name: NATIVE_DECISION_ANSWER_NAME,
|
|
308
|
+
instructions:
|
|
309
|
+
'Choose one listed specialist only when the current request clearly needs it; otherwise choose none. This is advisory only.',
|
|
310
|
+
choices: [...[...candidates].map((value) => ({ value })), { value: 'none' }],
|
|
311
|
+
},
|
|
312
|
+
],
|
|
313
|
+
} satisfies NativeDecisionRequest
|
|
314
|
+
|
|
315
|
+
return { ok: true, request: Object.freeze(request) }
|
|
316
|
+
}
|
|
317
|
+
|
|
318
|
+
export interface NativeDecisionAdvice {
|
|
319
|
+
readonly agent: string | null
|
|
320
|
+
readonly reasonCode: 'RECOMMENDED' | 'ABSTAIN' | 'INVALID_RESPONSE' | 'LOW_CONFIDENCE'
|
|
321
|
+
readonly effect: 'none'
|
|
322
|
+
}
|
|
323
|
+
|
|
324
|
+
/** Bounded consumer: validates one candidate and returns advice only. */
|
|
325
|
+
export function consumeNativeDecisionAdvice(
|
|
326
|
+
response: unknown,
|
|
327
|
+
eligibleAgentIDs: readonly string[],
|
|
328
|
+
): NativeDecisionAdvice {
|
|
329
|
+
const abstain = (reasonCode: NativeDecisionAdvice['reasonCode']): NativeDecisionAdvice =>
|
|
330
|
+
Object.freeze({ agent: null, reasonCode, effect: 'none' })
|
|
331
|
+
if (response === null || typeof response !== 'object' || Array.isArray(response)) {
|
|
332
|
+
return abstain('INVALID_RESPONSE')
|
|
333
|
+
}
|
|
334
|
+
const answers = (response as { answers?: unknown }).answers
|
|
335
|
+
if (!Array.isArray(answers)) return abstain('INVALID_RESPONSE')
|
|
336
|
+
const matches = answers.filter(
|
|
337
|
+
(answer): answer is Record<string, unknown> =>
|
|
338
|
+
answer !== null &&
|
|
339
|
+
typeof answer === 'object' &&
|
|
340
|
+
!Array.isArray(answer) &&
|
|
341
|
+
answer.name === NATIVE_DECISION_ANSWER_NAME,
|
|
342
|
+
)
|
|
343
|
+
if (matches.length !== 1) return abstain(matches.length === 0 ? 'ABSTAIN' : 'INVALID_RESPONSE')
|
|
344
|
+
|
|
345
|
+
const answer = matches[0]
|
|
346
|
+
if (answer?.type !== 'choice') return abstain('ABSTAIN')
|
|
347
|
+
if (answer.choice === 'none') return abstain('ABSTAIN')
|
|
348
|
+
if (typeof answer.choice !== 'string' || !eligibleAgentIDs.includes(answer.choice)) {
|
|
349
|
+
return abstain('INVALID_RESPONSE')
|
|
350
|
+
}
|
|
351
|
+
if (
|
|
352
|
+
typeof answer.confidence !== 'number' ||
|
|
353
|
+
!Number.isFinite(answer.confidence) ||
|
|
354
|
+
answer.confidence < 0 ||
|
|
355
|
+
answer.confidence > 1
|
|
356
|
+
) {
|
|
357
|
+
return abstain('INVALID_RESPONSE')
|
|
358
|
+
}
|
|
359
|
+
if (!DECISION_AGENT_IDS.has(answer.choice)) return abstain('INVALID_RESPONSE')
|
|
360
|
+
return Object.freeze({ agent: answer.choice, reasonCode: 'RECOMMENDED', effect: 'none' })
|
|
361
|
+
}
|
|
362
|
+
|
|
363
|
+
// This type assertion keeps the source contract tied to @opencode/plugin@2.2.5.
|
|
364
|
+
// Any removal or schema change to the native method fails `npm run typecheck`.
|
|
365
|
+
export type OpenCodeV225DecisionApi = Pick<Context['experimental']['generate'], 'decisions'>
|
package/src/pantheon/v2-tools.ts
CHANGED
|
@@ -48,9 +48,11 @@
|
|
|
48
48
|
|
|
49
49
|
import { createCostCommand } from './cost-command.ts'
|
|
50
50
|
import { createHashlineEditTool, type HashlineToolResult } from './hashline/tool.ts'
|
|
51
|
+
import { createV2JevDecisionDefinition, type V2JevDecisionRuntime } from './jev-decision-tool.ts'
|
|
51
52
|
import { createV2LocalDecisionDefinition } from './local-decision-tool.ts'
|
|
52
53
|
import { createModelCommand } from './model-command.ts'
|
|
53
54
|
import type { ToolContextLike } from './tool-context.ts'
|
|
55
|
+
import type { ShadowSink } from './v2-decision-shadow.ts'
|
|
54
56
|
import { createV2VisionTool, type V2VisionRuntime } from './v2-vision.ts'
|
|
55
57
|
|
|
56
58
|
// ─── Contract types ─────────────────────────────────────────────────────
|
|
@@ -420,6 +422,17 @@ async function executeModel(
|
|
|
420
422
|
|
|
421
423
|
// ─── Tool set ───────────────────────────────────────────────────────────
|
|
422
424
|
|
|
425
|
+
/**
|
|
426
|
+
* Optional shadow sinks wired at tool registration.
|
|
427
|
+
*
|
|
428
|
+
* `complexity` belongs to P3, whose declared host surface is `tool.invoke`: the
|
|
429
|
+
* local selector's own input carries the complexity vector and gate status, so
|
|
430
|
+
* the observation is constructed where the payload can actually feed it.
|
|
431
|
+
*/
|
|
432
|
+
export interface V2ToolShadowOptions {
|
|
433
|
+
readonly complexitySink?: ShadowSink
|
|
434
|
+
}
|
|
435
|
+
|
|
423
436
|
/**
|
|
424
437
|
* Build the V2 tool definitions.
|
|
425
438
|
*
|
|
@@ -434,7 +447,11 @@ async function executeModel(
|
|
|
434
447
|
* per-call definition with no construction-time I/O. A fresh descriptor array
|
|
435
448
|
* prevents callers from mutating another registration's definitions.
|
|
436
449
|
*/
|
|
437
|
-
export function createV2ToolDefinitions(
|
|
450
|
+
export function createV2ToolDefinitions(
|
|
451
|
+
visionRuntime?: V2VisionRuntime,
|
|
452
|
+
shadow: V2ToolShadowOptions = {},
|
|
453
|
+
jevRuntime?: V2JevDecisionRuntime,
|
|
454
|
+
): V2ToolDef[] {
|
|
438
455
|
const definitions: V2ToolDef[] = [
|
|
439
456
|
{
|
|
440
457
|
name: 'hashline_edit',
|
|
@@ -475,7 +492,12 @@ export function createV2ToolDefinitions(visionRuntime?: V2VisionRuntime): V2Tool
|
|
|
475
492
|
execute: executeModel,
|
|
476
493
|
},
|
|
477
494
|
]
|
|
478
|
-
definitions.push(
|
|
495
|
+
definitions.push(
|
|
496
|
+
createV2LocalDecisionDefinition({
|
|
497
|
+
...(shadow.complexitySink === undefined ? {} : { sink: shadow.complexitySink }),
|
|
498
|
+
}),
|
|
499
|
+
)
|
|
479
500
|
if (visionRuntime) definitions.push(createV2VisionTool(visionRuntime))
|
|
501
|
+
if (jevRuntime) definitions.push(createV2JevDecisionDefinition(jevRuntime))
|
|
480
502
|
return definitions
|
|
481
503
|
}
|