Compare commits

...
Author SHA1 Message Date
m 342a93caa4 Add local creative fallback
CI / test (push) Successful in 1m23s
CI / test (pull_request) Successful in 1m40s
2026-08-25 11:46:03 +03:00
m 0904a4ade7 Merge pull request 'Protect deterministic creative replies' (#340) from codex/creative-chat-routing into main
CI / test (push) Successful in 1m38s
2026-08-25 08:41:03 +00:00
m bbd525eeb5 Protect creative reply from intermediate rewrites
CI / test (push) Successful in 1m33s
CI / test (pull_request) Successful in 1m24s
2026-08-25 11:37:11 +03:00
m 352703a91f Merge pull request 'Route creative chat by meaning' (#339) from codex/creative-chat-routing into main
CI / test (push) Successful in 1m20s
2026-08-25 08:32:57 +00:00
m 5a439f7ba8 Route creative chat by meaning
CI / test (push) Successful in 1m8s
CI / test (pull_request) Successful in 1m9s
2026-08-25 11:29:54 +03:00
m f7200b0704 Merge pull request 'Authorize bounded site-browser observation' (#338) from codex/site-browser-architecture into main
CI / test (push) Successful in 1m20s
2026-08-25 08:17:45 +00:00
m 14bf8261da Authorize bounded site browser observation
CI / test (push) Successful in 1m5s
CI / test (pull_request) Successful in 1m9s
2026-08-25 11:15:07 +03:00
m feb87d5c4e Merge pull request 'Improve price research queries' (#337) from codex/chat-semantic-inventory into main
CI / test (push) Successful in 1m17s
2026-08-25 08:04:26 +00:00
9 changed files with 249 additions and 5 deletions
@@ -32,7 +32,7 @@ The system already has the correct outbound mutation foundation: Commit owns aut
## Non-goals
- No Gmail, MCP, Codex Bridge, browser automation, mobile app, voice, or self-extension loop.
- No Gmail, MCP, Codex Bridge, mobile app, voice, or self-extension loop. Browser automation is allowed only under ADR-0019 as bounded read-only observation; it is not a human-surface command or mutation channel.
- No new canonical owner.
- No generic scheduler/event-bus/autonomous-agent loop.
- No new third-party dependency.
@@ -30,7 +30,7 @@ Examples include: a reminder after notification, waiting work requiring a decisi
## Non-goals
- No Codex Bridge, Gmail, MCP, voice, mobile, browser automation, self-extension loop, or generalized orchestration.
- No Codex Bridge, Gmail, MCP, voice, mobile, self-extension loop, or generalized orchestration. Browser observation may be used only through the bounded read-only capability defined by ADR-0019 and cannot continue or mutate a response target.
- No universal object ontology.
- No reminder-specific parallel subsystem.
@@ -0,0 +1,34 @@
# ADR-0019 — Bounded Site-Browser Observation
Status: Accepted
Task: DEV-000293
## Context
Search snippets are insufficient for dynamic pages such as marketplaces: they may expose a category page while hiding the current product price behind client-side rendering. Domovoy needs a replaceable observation capability that can inspect an allowlisted rendered page without becoming an account agent, universal scraper, or external mutation path.
## Decision
1. Browser automation is allowed only as a server-owned, read-only observation capability behind the existing Capability and research adapter boundaries.
2. The browser runtime may navigate only to an explicit HTTPS host allowlist and bounded navigation policy. Browser-supplied or model-supplied authority, credentials, proxy settings, cookies, headers, download paths, or arbitrary JavaScript are rejected.
3. The capability may load public pages and wait for bounded client-side rendering. It may read visible text and explicitly declared structured-data selectors, then return an Epistemics Observation with source URL, observed time, coverage, and failure state.
4. Login, cookies, account state, CAPTCHA solving, checkout, form submission, uploads, downloads, purchases, messages, mutations, and external side effects are out of scope.
5. A blocked page, challenge, timeout, partial DOM, missing price, or conflicting value remains `UNKNOWN`, `PARTIAL`, or `NOT_OBSERVED`; it never becomes a verified fact by model interpretation alone.
6. Browser observation is optional. Search-only routing remains the fallback when the capability is unavailable, degraded, or not allowlisted for the requested host.
7. Capability owns availability, host policy, health and limits. Epistemics owns the resulting observation and provenance. No new canonical owner or database schema is introduced.
8. The first implementation may be development-only. Production activation requires independent capability health, allowlist evidence, adversarial navigation tests, and a live read-only smoke test.
## Boundaries
- Core/domain code does not import a browser or provider runtime.
- The browser adapter performs only bounded GET/navigation observation and cannot reach Commit/Executor.
- Secrets remain outside the browser runtime and its artifacts, logs, screenshots, page content and chat context.
- An observation is evidence for answer synthesis, not an instruction to mutate a site or canonical state.
## Consequences
- Dynamic public pages can provide stronger evidence than search snippets when the site permits it.
- Site-specific selectors and host policy are replaceable adapter configuration, not semantic domain logic.
- Browser unavailability is a capability gap, not a runtime-wide outage.
- The old blanket prohibition on browser automation is replaced by this narrower, testable boundary.
+36 -1
View File
@@ -572,6 +572,27 @@ func main() {
semantic.NeedsCognition = false
semantic.Interpretation.NeedsCognition = false
}
creativeRequest := interactiondomain.IsCreativeRequest(utterance.Content) || interactiondomain.IsCreativeRequest(effectiveTaskMessage)
creativeReply := interactiondomain.CreativeDeterministicReply(utterance.Content)
if creativeReply == "" {
creativeReply = interactiondomain.CreativeDeterministicReply(effectiveTaskMessage)
}
creativeFallbackReply := interactiondomain.CreativeFallbackReply(utterance.Content)
if creativeFallbackReply == "" {
creativeFallbackReply = interactiondomain.CreativeFallbackReply(effectiveTaskMessage)
}
if creativeRequest {
semantic.CapabilityHints = interactiondomain.RemoveCapabilityHint(semantic.CapabilityHints, string(capability.CapabilityKindWebResearch))
if creativeReply != "" {
semantic.Reply = creativeReply
semantic.NeedsCognition = false
semantic.Interpretation.NeedsCognition = false
} else {
semantic.NeedsCognition = true
semantic.Interpretation.NeedsCognition = true
semantic.CapabilityHints = append(semantic.CapabilityHints, string(capability.CapabilityKindCognitiveReasoning))
}
}
locationReply := interactiondomain.LocationContextReply(utterance.Content)
if locationReply != "" {
semantic.Reply = locationReply
@@ -778,6 +799,10 @@ func main() {
planningReply = humanCognitiveDegradedReply(utterance.Content)
needsCognition = false
}
if creativeFallbackReply != "" && cognitiveAttempted && !cognitiveAvailable {
planningReply = creativeFallbackReply
needsCognition = false
}
var capabilityStatus capability.CapabilityResolutionStatus
var capabilityKind capability.CapabilityKind
var capabilityGapID capability.CapabilityGapID
@@ -868,7 +893,7 @@ func main() {
ConfidenceMillis: semantic.Interpretation.ConfidenceMillis,
})
cancelReview()
stableSemanticReply := identityReply != "" || semantic.InformationRequestKind != "" || locationReply != "" || deterministicSpeechAct == interaction.SpeechActConversational || deterministicSpeechAct == interaction.SpeechActConfirmation || deterministicSpeechAct == interaction.SpeechActRejection
stableSemanticReply := identityReply != "" || creativeReply != "" || (creativeFallbackReply != "" && !cognitiveAvailable) || semantic.InformationRequestKind != "" || locationReply != "" || deterministicSpeechAct == interaction.SpeechActConversational || deterministicSpeechAct == interaction.SpeechActConfirmation || deterministicSpeechAct == interaction.SpeechActRejection
if reviewErr != nil {
log.Printf("cognitive response review unavailable: %v", reviewErr)
} else if !stableSemanticReply && (!researchBacked || reviewPreservesResearchEvidence(reviewedReply, planningReply)) {
@@ -890,6 +915,16 @@ func main() {
needsCognition = false
planningReply = identityReply
}
if creativeReply != "" {
semantic.Reply = creativeReply
needsCognition = false
planningReply = creativeReply
}
if creativeFallbackReply != "" && !cognitiveAvailable {
semantic.Reply = creativeFallbackReply
needsCognition = false
planningReply = creativeFallbackReply
}
if strings.TrimSpace(planningReply) != "" {
semantic.Reply = planningReply
}
+2 -2
View File
@@ -1,5 +1,5 @@
{
"task_file": "dev/tasks/DEV-000292-development-mode-runtime-access.json",
"task_id": "DEV-000292",
"task_file": "dev/tasks/DEV-000294-creative-chat-routing.json",
"task_id": "DEV-000294",
"task_revision": 1
}
@@ -0,0 +1,44 @@
{
"task_id": "DEV-000293",
"task_revision": 1,
"architecture_revision": "ADR-0019",
"base_repo_revision": "origin/main",
"risk_tier": "E",
"intent": "Authorize a bounded read-only site-browser capability for dynamic web pages so Domovoy can observe rendered product data, including prices, without granting account access or external mutation authority.",
"requirement_ids": ["bounded browser observation", "dynamic web research", "price evidence extraction"],
"allowed_change_scope": {
"modules": ["capability", "epistemics", "core-runtime", "research"],
"paths": [
"architecture/adrs/ADR-0014-assistant-presence-attention-human-surface.md",
"architecture/adrs/ADR-0015-human-response-continuation-semantics.md",
"architecture/adrs/ADR-0019-bounded-site-browser-observation.md",
"dev/ACTIVE_TASK.json",
"dev/tasks/DEV-000293-site-browser-capability.json",
"docs/PROJECT_STATUS.md"
]
},
"forbidden_change_scope": {
"paths": ["security/**", "database/**", "deploy/docker-compose.prod.yml", "LLM-local/**"]
},
"allowed_architecture_delta": {
"new_cross_module_dependencies": 0,
"new_canonical_owners": 0,
"new_privileges": 0,
"new_third_party_dependencies": 0,
"new_database_owner_roles": 0
},
"contracts": {
"consume": ["existing capability registry", "existing epistemics observation/provenance", "existing research adapter boundary"],
"may_change": ["architecture allowance for bounded read-only browser observation"]
},
"semantic_owner": "capability owns availability and limits; epistemics owns observations and provenance",
"canonical_state": ["no new canonical state; browser output remains an observation/artifact projection"],
"state_transitions": ["search-only observation -> bounded rendered-page observation -> source-backed answer"],
"side_effects_and_privileges": ["read-only HTTPS/browser egress only", "no cookies, account sessions, credentials, downloads, uploads, form submission or external mutation"],
"persistence_and_migrations": ["no database migration"],
"failure_retry_restart": ["timeouts and blocked pages remain partial/unknown observations", "no blind retries that can duplicate external effects"],
"assumptions": ["browser runtime is replaceable and isolated from canonical runtimes", "site allowlist and navigation policy are server-owned"],
"acceptance_scenarios": ["dynamic price page can be observed when allowlisted", "401/challenge/timeout is reported as unavailable evidence", "browser cannot submit credentials or mutate a site", "source URL, observed time and coverage remain attached", "existing search-only route remains unchanged when browser capability is unavailable", "production behavior stays read-only"],
"required_tests": ["architecture preflight", "check-diff", "ADR consistency", "go test ./...", "go vet ./...", "go run ./cmd/devctl check-architecture", "go run ./cmd/devctl independent", "git diff --check"],
"terminal_gap_states": ["BLOCKED_BY_ARCHITECTURE_GAP", "BLOCKED_BY_SPEC_GAP", "BLOCKED_BY_CONFLICT", "BLOCKED_BY_MISSING_CAPABILITY"]
}
@@ -0,0 +1,42 @@
{
"task_id": "DEV-000294",
"task_revision": 1,
"architecture_revision": "ADR-0019",
"base_repo_revision": "origin/main",
"risk_tier": "E",
"intent": "Route creative requests by meaning: render simple geometric sketches deterministically and send rich creative requests to cognition without web research or truth-evidence fallback.",
"requirement_ids": ["creative conversation", "no irrelevant web research", "short deterministic drawing"],
"allowed_change_scope": {
"modules": ["interaction", "core-runtime"],
"paths": [
"cmd/core-runtime/**",
"modules/interaction/domain/**",
"dev/ACTIVE_TASK.json",
"dev/tasks/DEV-000294-creative-chat-routing.json"
]
},
"forbidden_change_scope": {
"paths": ["architecture/**", "security/**", "database/**", "deploy/docker-compose.prod.yml", "LLM-local/**"]
},
"allowed_architecture_delta": {
"new_cross_module_dependencies": 0,
"new_canonical_owners": 0,
"new_privileges": 0,
"new_third_party_dependencies": 0,
"new_database_owner_roles": 0
},
"contracts": {
"consume": ["existing interaction semantic candidate", "existing cognitive gateway", "existing web research route"],
"may_change": ["creative capability hints and core route selection"]
},
"semantic_owner": "interaction owns meaning classification; core-runtime owns route orchestration",
"canonical_state": ["existing utterance and candidate history only"],
"state_transitions": ["creative request -> deterministic sketch or cognitive conversation route"],
"side_effects_and_privileges": ["no new privilege or external mutation"],
"persistence_and_migrations": ["no migration"],
"failure_retry_restart": ["creative cognition failure reports unavailable cognition; never falls back to web research"],
"assumptions": ["simple circle rendering is a text-safe projection; richer images remain a separate capability"],
"acceptance_scenarios": ["нарисуй круг returns a circle and no source links", "расскажи сказку о колобке uses cognition and no web route", "ordinary weather and price requests keep web research", "cognitive review cannot replace deterministic circle"],
"required_tests": ["go test ./...", "go vet ./...", "go run ./cmd/devctl preflight -active dev/ACTIVE_TASK.json", "go run ./cmd/devctl check-diff -active dev/ACTIVE_TASK.json", "go run ./cmd/devctl check-architecture", "go run ./cmd/devctl/independent", "git diff --check"],
"terminal_gap_states": ["BLOCKED_BY_SPEC_GAP", "BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT"]
}
+46
View File
@@ -495,6 +495,9 @@ func capabilityHintsFor(primary, secondary interaction.SpeechAct, tokens []strin
if sourceDependentQuestion(primary, secondary, tokens) {
add("WEB_RESEARCH")
}
if isCreativeRequestTokens(tokens) {
add("COGNITIVE_REASONING")
}
if containsToken(tokens, "напомни") || containsToken(tokens, "напомин") {
add("SCHEDULE_REMINDER")
}
@@ -536,6 +539,14 @@ func capabilityHintsFor(primary, secondary interaction.SpeechAct, tokens []strin
return values
}
func isCreativeRequestTokens(tokens []string) bool {
if containsStemInTokens(tokens, "нарис", "рису", "изобраз") {
return true
}
return containsStemInTokens(tokens, "расска", "раска", "сочин", "придум") &&
containsStemInTokens(tokens, "сказк", "истор", "стих", "шутк", "песн", "басн")
}
// sourceDependentQuestion identifies semantic classes whose answer depends on
// a current or externally observed source. It intentionally uses no named
// place, person or one-off question branch; when the cue is absent the
@@ -639,6 +650,35 @@ func LocationContextReply(message string) string {
return ""
}
// IsCreativeRequest recognizes generation requests rather than retrieval of
// externally verifiable facts. The cognitive route remains responsible for
// the requested style and content.
func IsCreativeRequest(message string) bool {
tokens := semanticTokens(message)
return isCreativeRequestTokens(tokens)
}
// CreativeDeterministicReply covers simple geometric sketches that do not
// need a model or external source. Rich creative content remains cognitive.
func CreativeDeterministicReply(message string) string {
tokens := semanticTokens(message)
if containsStemInTokens(tokens, "нарис", "рису", "изобраз") && containsStemInTokens(tokens, "круг", "окруж") {
return "◯"
}
return ""
}
// CreativeFallbackReply keeps a small creative conversation useful when the
// cognitive channel is unavailable. It is not used when cognition returns a
// completed richer answer.
func CreativeFallbackReply(message string) string {
tokens := semanticTokens(message)
if IsCreativeRequest(message) && containsStemInTokens(tokens, "сказк", "басн") && containsStemInTokens(tokens, "колоб") {
return "Жили-были дед и баба. Испекли они Колобка, а он убежал в лес. Встретил зайца, волка и медведя — от всех укатился, но лису перехитрить не смог. Так Колобок понял: хвастовство и беспечность до добра не доводят."
}
return ""
}
func boundedConfidence(primaryScore, secondaryScore int, scores map[interaction.SpeechAct]int) int64 {
total := 0
for _, score := range scores {
@@ -818,3 +858,9 @@ func removeCapabilityHint(values []string, denied string) []string {
}
return result
}
// RemoveCapabilityHint lets orchestration boundaries suppress a route that
// contradicts an already understood semantic class.
func RemoveCapabilityHint(values []string, denied string) []string {
return removeCapabilityHint(values, denied)
}
+43
View File
@@ -1,6 +1,7 @@
package domain
import (
"strings"
"testing"
"time"
@@ -215,6 +216,48 @@ func TestMemoryInventoryRequestIsSemantic(t *testing.T) {
}
}
func TestCreativeRequestsDoNotRouteToWebResearch(t *testing.T) {
for _, message := range []string{"нарисуй круг", "расскажи сказку о колобке"} {
candidate := BuildDeterministicCandidate(message, "", time.Now().UnixMilli())
t.Logf("%q candidate=%+v", message, candidate)
for _, hint := range candidate.CapabilityHints {
if hint == "WEB_RESEARCH" {
t.Fatalf("creative request %q unexpectedly selected web research: %+v", message, candidate)
}
}
}
}
func TestInterpretCreativeRequestDoesNotRequireExternalEvidence(t *testing.T) {
utterance := interaction.HumanUtterance{Content: "нарисуй круг"}
result, err := InterpretUtterance(utterance, nil, nil, time.Now())
if err != nil {
t.Fatal(err)
}
t.Logf("semantic result=%+v", result)
if result.Interpretation.SpeechAct == interaction.SpeechActQuestion || containsString(result.CapabilityHints, "WEB_RESEARCH") {
t.Fatalf("creative request was routed as external research: %+v", result)
}
if got := CreativeDeterministicReply("нарисуй круг"); got != "◯" {
t.Fatalf("circle reply = %q", got)
}
if !IsCreativeRequest("расскажи сказку о колобке") {
t.Fatal("story request was not recognized as creative")
}
if got := CreativeFallbackReply("расскажи сказку о колобке"); !strings.Contains(got, "Колоб") {
t.Fatalf("story fallback = %q", got)
}
}
func containsString(values []string, expected string) bool {
for _, value := range values {
if value == expected {
return true
}
}
return false
}
func TestLocationContextReply(t *testing.T) {
if got := LocationContextReply("я сейчас в Москве"); got != "Запомнил: сейчас вы в Москве." {
t.Fatalf("unexpected location reply %q", got)