pantheon-opencode 2.0.3 → 2.1.0
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/AGENTS.global.md +1 -1
- package/AGENTS.md +1 -1
- package/CHANGELOG.md +68 -0
- package/README.md +13 -6
- package/README.pt-BR.md +12 -5
- package/commands/pantheon-deepwork.md +11 -2
- package/package.json +2 -1
- package/plugin.json +1 -1
- package/prompts/orchestrate-with-zeus.prompt.md +2 -0
- package/pyproject.toml +1 -1
- package/scripts/cost.mjs +4 -5
- package/scripts/decision-router-budget.mjs +27 -6
- package/scripts/decision-router.mjs +76 -34
- package/scripts/doctor.mjs +74 -0
- package/scripts/install/decision-router-config.mjs +32 -7
- package/src/agents/apollo.md +6 -53
- package/src/agents/athena.md +1 -0
- package/src/agents/talos.md +10 -40
- package/src/agents/themis.md +9 -0
- package/src/agents/zeus.md +29 -2
- package/src/mcp/pantheon_vision.py +33 -2
- package/src/pantheon/background-job-board.ts +33 -0
- package/src/pantheon/bus.ts +190 -0
- package/src/pantheon/cost-command.ts +4 -1
- package/src/pantheon/deepwork/deepwork-tool.ts +370 -0
- package/src/pantheon/deepwork/plan-projection.ts +274 -0
- package/src/pantheon/deepwork/plan-schema.ts +93 -0
- package/src/pantheon/deepwork/plan-tracker.ts +336 -0
- package/src/pantheon/execution-receipt.ts +113 -0
- package/src/pantheon/jev-decision-tool.ts +8 -2
- package/src/pantheon/local-decision-budget.ts +36 -6
- package/src/pantheon/model-command.ts +2 -4
- package/src/pantheon/native-vision.ts +3 -1
- package/src/pantheon/prompt-cache-order.ts +89 -0
- package/src/pantheon/skill-search-tool.ts +2 -2
- package/src/pantheon/themis-reflex.ts +164 -0
- package/src/pantheon/tools.ts +2 -0
- package/src/plugin/index.js +979 -60
- package/src/plugins/tui/dist/tui.js +461 -289
- package/src/plugins/tui/dist/tui.js.map +1 -1
- package/src/plugins/tui/package-lock.json +2 -2
- package/src/plugins/tui/package.json +1 -1
- package/src/plugins/tui/src/autonomy-state.ts +14 -16
- package/src/plugins/tui/src/decision-router-state.ts +22 -10
- package/src/plugins/tui/src/pantheon-settings-state.ts +119 -35
- package/src/plugins/tui/src/pantheon-settings.tsx +142 -48
- package/src/plugins/tui/src/preset-state.ts +96 -6
- package/src/plugins/tui/src/update-status.ts +62 -22
|
@@ -0,0 +1,89 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Prompt Prefix Cache Hierarchy Optimizer — Late-2026 KV-Cache Alignment.
|
|
3
|
+
*
|
|
4
|
+
* Enforces deterministic hierarchical ordering for agent prompts to maximize
|
|
5
|
+
* prefix cache hit rates across LLM providers (Anthropic Claude 3.5/3.7,
|
|
6
|
+
* Gemini 2.0/Flash, OpenAI GPT-5/o3/Luna):
|
|
7
|
+
*
|
|
8
|
+
* [Tier 1: Static Identity & Invariants] ->
|
|
9
|
+
* [Tier 2: Tool Schemas & Governance] ->
|
|
10
|
+
* [Tier 3: Repository Standards & Gates] ->
|
|
11
|
+
* [Tier 4: Ephemeral Session History] ->
|
|
12
|
+
* [Tier 5: Current Turn & Scratchpad]
|
|
13
|
+
*
|
|
14
|
+
* Pure TypeScript — zero external dependencies.
|
|
15
|
+
*
|
|
16
|
+
* @module prompt-cache-order
|
|
17
|
+
*/
|
|
18
|
+
|
|
19
|
+
export interface PromptLayers {
|
|
20
|
+
readonly identity: string
|
|
21
|
+
readonly tools?: readonly string[] | undefined
|
|
22
|
+
readonly conventions?: readonly string[] | undefined
|
|
23
|
+
readonly history?: readonly string[] | undefined
|
|
24
|
+
readonly currentTurn: string
|
|
25
|
+
}
|
|
26
|
+
|
|
27
|
+
export interface CachedPromptResult {
|
|
28
|
+
readonly prompt: string
|
|
29
|
+
/** The byte/character offset separating invariant prefix from volatile turns */
|
|
30
|
+
readonly prefixBoundaryIndex: number
|
|
31
|
+
/** Estimated token length (heuristic ~4 chars/token) */
|
|
32
|
+
readonly estimatedTokens: number
|
|
33
|
+
/** Whether the prefix meets typical provider minimum threshold (e.g. 1024 tokens) */
|
|
34
|
+
readonly cacheEligible: boolean
|
|
35
|
+
}
|
|
36
|
+
|
|
37
|
+
/** Standard token threshold for provider-level prompt caching (Anthropic=1024, OpenAI=1024) */
|
|
38
|
+
export const DEFAULT_CACHE_MIN_TOKENS = 1024
|
|
39
|
+
|
|
40
|
+
/**
|
|
41
|
+
* Assemble agent prompt with strict invariant-prefix ordering to maximize KV cache hit rate.
|
|
42
|
+
*/
|
|
43
|
+
export function buildPrefixCachedPrompt(
|
|
44
|
+
layers: PromptLayers,
|
|
45
|
+
minCacheTokens = DEFAULT_CACHE_MIN_TOKENS,
|
|
46
|
+
): CachedPromptResult {
|
|
47
|
+
const invariantParts: string[] = []
|
|
48
|
+
|
|
49
|
+
// 1. Identity & Core System Instructions
|
|
50
|
+
invariantParts.push(layers.identity.trim())
|
|
51
|
+
|
|
52
|
+
// 2. Tools & Permissions Schemas
|
|
53
|
+
if (layers.tools && layers.tools.length > 0) {
|
|
54
|
+
invariantParts.push(`## Tool Specifications\n${layers.tools.map((t) => t.trim()).join('\n\n')}`)
|
|
55
|
+
}
|
|
56
|
+
|
|
57
|
+
// 3. Project Conventions & Architecture Gates
|
|
58
|
+
if (layers.conventions && layers.conventions.length > 0) {
|
|
59
|
+
invariantParts.push(
|
|
60
|
+
`## Project Standards & Gates\n${layers.conventions.map((c) => c.trim()).join('\n\n')}`,
|
|
61
|
+
)
|
|
62
|
+
}
|
|
63
|
+
|
|
64
|
+
const prefix = invariantParts.join('\n\n')
|
|
65
|
+
const prefixBoundaryIndex = prefix.length
|
|
66
|
+
|
|
67
|
+
const volatileParts: string[] = []
|
|
68
|
+
|
|
69
|
+
// 4. Conversation History (if any)
|
|
70
|
+
if (layers.history && layers.history.length > 0) {
|
|
71
|
+
volatileParts.push(
|
|
72
|
+
`## Conversation History\n${layers.history.map((h) => h.trim()).join('\n\n')}`,
|
|
73
|
+
)
|
|
74
|
+
}
|
|
75
|
+
|
|
76
|
+
// 5. Current Turn Payload
|
|
77
|
+
volatileParts.push(layers.currentTurn.trim())
|
|
78
|
+
|
|
79
|
+
const fullPrompt = [prefix, volatileParts.join('\n\n')].join('\n\n')
|
|
80
|
+
const estimatedTokens = Math.ceil(fullPrompt.length / 4)
|
|
81
|
+
const estimatedPrefixTokens = Math.ceil(prefix.length / 4)
|
|
82
|
+
|
|
83
|
+
return {
|
|
84
|
+
prompt: fullPrompt,
|
|
85
|
+
prefixBoundaryIndex,
|
|
86
|
+
estimatedTokens,
|
|
87
|
+
cacheEligible: estimatedPrefixTokens >= minCacheTokens,
|
|
88
|
+
}
|
|
89
|
+
}
|
|
@@ -41,7 +41,7 @@ const TOOL_DESCRIPTION =
|
|
|
41
41
|
'Advisory discovery tool for skills. Queries installed Pantheon skills and external ' +
|
|
42
42
|
'registries (Anthropic catalog, skills.sh, GitHub). Returns matched skill IDs, ' +
|
|
43
43
|
'descriptions, install counts, and provenance tiers. Zeus uses this to recommend ' +
|
|
44
|
-
'skills before prompting the user
|
|
44
|
+
'project-scoped skills (.agents/skills/) before proactively prompting the user via the question tool.'
|
|
45
45
|
|
|
46
46
|
const SKILL_SEARCH_INPUT_SCHEMA: Record<string, unknown> = {
|
|
47
47
|
type: 'object',
|
|
@@ -159,7 +159,7 @@ export function executeSkillSearch(input: Record<string, unknown>): SkillSearchO
|
|
|
159
159
|
installed: false,
|
|
160
160
|
installs: entry.installs ?? null,
|
|
161
161
|
url: entry.url,
|
|
162
|
-
installCommand: `npx skills add ${entry.skill} -a opencode`,
|
|
162
|
+
installCommand: `npx skills add ${entry.skill} -a opencode -y`,
|
|
163
163
|
})
|
|
164
164
|
}
|
|
165
165
|
}
|
|
@@ -0,0 +1,164 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Themis Local Reflex Loop — Self-healing inline verification for implementer agents.
|
|
3
|
+
*
|
|
4
|
+
* Catches syntax errors, lint failures, and sliced test failures directly in the
|
|
5
|
+
* implementer's execution cycle, granting up to 2 bounded retry attempts before
|
|
6
|
+
* escalating to the Zeus coordinator.
|
|
7
|
+
*
|
|
8
|
+
* Emits reactive events on `getSharedBus()` for auditability and observability.
|
|
9
|
+
*
|
|
10
|
+
* Pure TypeScript — zero external dependencies.
|
|
11
|
+
*
|
|
12
|
+
* @module themis-reflex
|
|
13
|
+
*/
|
|
14
|
+
|
|
15
|
+
import { getSharedBus } from './bus.ts'
|
|
16
|
+
|
|
17
|
+
export interface ThemisDiscrepancy {
|
|
18
|
+
readonly file: string
|
|
19
|
+
readonly error: string
|
|
20
|
+
readonly expected?: string
|
|
21
|
+
readonly actual?: string
|
|
22
|
+
readonly suggestedFix?: string
|
|
23
|
+
}
|
|
24
|
+
|
|
25
|
+
export interface ReflexAttemptState {
|
|
26
|
+
taskID: string
|
|
27
|
+
agent: string
|
|
28
|
+
attempt: number
|
|
29
|
+
maxAttempts: number
|
|
30
|
+
discrepancies: ThemisDiscrepancy[]
|
|
31
|
+
timestamp: number
|
|
32
|
+
}
|
|
33
|
+
|
|
34
|
+
export type ThemisReflexDecision =
|
|
35
|
+
| {
|
|
36
|
+
readonly action: 'retry'
|
|
37
|
+
readonly prompt: string
|
|
38
|
+
readonly attempt: number
|
|
39
|
+
readonly maxAttempts: number
|
|
40
|
+
}
|
|
41
|
+
| {
|
|
42
|
+
readonly action: 'escalate'
|
|
43
|
+
readonly verdict: 'BLOCK' | 'FAIL'
|
|
44
|
+
readonly reason: string
|
|
45
|
+
readonly attempts: number
|
|
46
|
+
}
|
|
47
|
+
| {
|
|
48
|
+
readonly action: 'pass'
|
|
49
|
+
readonly verdict: 'PASS'
|
|
50
|
+
}
|
|
51
|
+
|
|
52
|
+
export class ThemisReflexController {
|
|
53
|
+
private readonly states: Map<string, ReflexAttemptState> = new Map()
|
|
54
|
+
private readonly maxRetries: number
|
|
55
|
+
|
|
56
|
+
constructor(options: { maxRetries?: number } = {}) {
|
|
57
|
+
this.maxRetries = options.maxRetries ?? 2
|
|
58
|
+
}
|
|
59
|
+
|
|
60
|
+
/**
|
|
61
|
+
* Evaluate a task execution result against Themis verification findings.
|
|
62
|
+
* If discrepancies are found and attempts remain, returns an actionable retry directive.
|
|
63
|
+
*/
|
|
64
|
+
evaluate(
|
|
65
|
+
taskID: string,
|
|
66
|
+
agent: string,
|
|
67
|
+
discrepancies: readonly ThemisDiscrepancy[],
|
|
68
|
+
): ThemisReflexDecision {
|
|
69
|
+
if (discrepancies.length === 0) {
|
|
70
|
+
const existing = this.states.get(taskID)
|
|
71
|
+
if (existing) {
|
|
72
|
+
getSharedBus().emit('themis:reflex:resolved', {
|
|
73
|
+
taskID,
|
|
74
|
+
agent,
|
|
75
|
+
file: existing.discrepancies[0]?.file ?? '',
|
|
76
|
+
error: '',
|
|
77
|
+
attempt: existing.attempt,
|
|
78
|
+
maxAttempts: this.maxRetries,
|
|
79
|
+
timestamp: Date.now(),
|
|
80
|
+
resolved: true,
|
|
81
|
+
})
|
|
82
|
+
this.states.delete(taskID)
|
|
83
|
+
}
|
|
84
|
+
return { action: 'pass', verdict: 'PASS' }
|
|
85
|
+
}
|
|
86
|
+
|
|
87
|
+
let state = this.states.get(taskID)
|
|
88
|
+
if (!state) {
|
|
89
|
+
state = {
|
|
90
|
+
taskID,
|
|
91
|
+
agent,
|
|
92
|
+
attempt: 0,
|
|
93
|
+
maxAttempts: this.maxRetries,
|
|
94
|
+
discrepancies: [...discrepancies],
|
|
95
|
+
timestamp: Date.now(),
|
|
96
|
+
}
|
|
97
|
+
this.states.set(taskID, state)
|
|
98
|
+
}
|
|
99
|
+
|
|
100
|
+
state.attempt += 1
|
|
101
|
+
state.discrepancies = [...discrepancies]
|
|
102
|
+
state.timestamp = Date.now()
|
|
103
|
+
|
|
104
|
+
const firstDisc = discrepancies[0]
|
|
105
|
+
getSharedBus().emit('themis:reflex:discrepancy', {
|
|
106
|
+
taskID,
|
|
107
|
+
agent,
|
|
108
|
+
file: firstDisc?.file ?? 'unknown',
|
|
109
|
+
error: firstDisc?.error ?? 'unspecified',
|
|
110
|
+
attempt: state.attempt,
|
|
111
|
+
maxAttempts: this.maxRetries,
|
|
112
|
+
timestamp: state.timestamp,
|
|
113
|
+
})
|
|
114
|
+
|
|
115
|
+
if (state.attempt <= this.maxRetries) {
|
|
116
|
+
const prompt = this.formatReflexPrompt(state.attempt, this.maxRetries, discrepancies)
|
|
117
|
+
return {
|
|
118
|
+
action: 'retry',
|
|
119
|
+
prompt,
|
|
120
|
+
attempt: state.attempt,
|
|
121
|
+
maxAttempts: this.maxRetries,
|
|
122
|
+
}
|
|
123
|
+
}
|
|
124
|
+
|
|
125
|
+
// Retries exhausted — escalate to coordinator
|
|
126
|
+
this.states.delete(taskID)
|
|
127
|
+
return {
|
|
128
|
+
action: 'escalate',
|
|
129
|
+
verdict: 'FAIL',
|
|
130
|
+
reason: `themis.reflex.exhausted:${state.attempt - 1}_retries_failed`,
|
|
131
|
+
attempts: state.attempt,
|
|
132
|
+
}
|
|
133
|
+
}
|
|
134
|
+
|
|
135
|
+
/**
|
|
136
|
+
* Format an inline, high-density reflex prompt directing the implementer to self-heal.
|
|
137
|
+
*/
|
|
138
|
+
private formatReflexPrompt(
|
|
139
|
+
attempt: number,
|
|
140
|
+
maxAttempts: number,
|
|
141
|
+
discrepancies: readonly ThemisDiscrepancy[],
|
|
142
|
+
): string {
|
|
143
|
+
const lines = [
|
|
144
|
+
`[THEMIS INLINE REFLEX — Attempt ${attempt}/${maxAttempts}]`,
|
|
145
|
+
'Themis automated audit detected discrepancies in your edits. Fix them directly before returning:',
|
|
146
|
+
]
|
|
147
|
+
|
|
148
|
+
for (const d of discrepancies) {
|
|
149
|
+
lines.push(`- File: ${d.file}`)
|
|
150
|
+
lines.push(` Error: ${d.error}`)
|
|
151
|
+
if (d.expected) lines.push(` Expected: ${d.expected}`)
|
|
152
|
+
if (d.actual) lines.push(` Actual: ${d.actual}`)
|
|
153
|
+
if (d.suggestedFix) lines.push(` Fix hint: ${d.suggestedFix}`)
|
|
154
|
+
}
|
|
155
|
+
|
|
156
|
+
lines.push('Do NOT escalate to Zeus yet. Apply the targeted fix and re-verify your slice.')
|
|
157
|
+
return lines.join('\n')
|
|
158
|
+
}
|
|
159
|
+
|
|
160
|
+
/** Reset state for a taskID */
|
|
161
|
+
reset(taskID: string): void {
|
|
162
|
+
this.states.delete(taskID)
|
|
163
|
+
}
|
|
164
|
+
}
|
package/src/pantheon/tools.ts
CHANGED
|
@@ -48,6 +48,7 @@
|
|
|
48
48
|
|
|
49
49
|
import { createCostCommand } from './cost-command.ts'
|
|
50
50
|
import type { ShadowSink } from './decision-shadow.ts'
|
|
51
|
+
import { createDeepworkTool } from './deepwork/deepwork-tool.ts'
|
|
51
52
|
import { createHashlineEditTool, type HashlineToolResult } from './hashline/tool.ts'
|
|
52
53
|
import { createV2JevDecisionDefinition, type V2JevDecisionRuntime } from './jev-decision-tool.ts'
|
|
53
54
|
import { createV2LocalDecisionDefinition } from './local-decision-tool.ts'
|
|
@@ -545,6 +546,7 @@ export function createV2ToolDefinitions(
|
|
|
545
546
|
)
|
|
546
547
|
definitions.push(todowriteTool)
|
|
547
548
|
definitions.push(createV2SkillSearchDefinition())
|
|
549
|
+
definitions.push(createDeepworkTool())
|
|
548
550
|
if (visionRuntime) definitions.push(createV2VisionTool(visionRuntime))
|
|
549
551
|
if (jevRuntime) definitions.push(createV2JevDecisionDefinition(jevRuntime))
|
|
550
552
|
return definitions
|