@theoremjs/agents 0.3.4-next.811.gf4e3b3f → 0.3.4-next.815.g7f5fa76
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/esm/src/guardrails/error.js +4 -1
- package/esm/src/guardrails/labels.d.ts +2 -2
- package/esm/src/guardrails/labels.js +14 -2
- package/esm/src/kernel/engine/runner/gates.js +49 -5
- package/esm/src/kernel/engine/runner/state.d.ts +2 -0
- package/esm/src/kernel/engine/runner/steps.js +35 -4
- package/package.json +1 -1
|
@@ -157,7 +157,10 @@ function withPublicLabels(event, labels) {
|
|
|
157
157
|
event.tool.phase === 'error' &&
|
|
158
158
|
event.tool.failure.error === undefined) {
|
|
159
159
|
const { tool } = event;
|
|
160
|
-
|
|
160
|
+
// why: a tool's network failure names the tool; error.network names the model provider.
|
|
161
|
+
const error = tool.failure.kind === 'network'
|
|
162
|
+
? labelText('tool.unreachable', { tool: tool.name }, labels)
|
|
163
|
+
: kindText(tool.failure.kind, labels, { tool: tool.name });
|
|
161
164
|
return { ...event, tool: { ...tool, failure: { ...tool.failure, error } } };
|
|
162
165
|
}
|
|
163
166
|
return event;
|
|
@@ -1,7 +1,7 @@
|
|
|
1
1
|
/** Values that fill a label template's `{param}` placeholders. */
|
|
2
2
|
export type LabelParams = Record<string, string | number>;
|
|
3
3
|
/** Keys are stable API. */
|
|
4
|
-
export declare const LABEL_KEYS: readonly ["provider.decision_failed", "provider.decision_unavailable", "provider.decision_model_missing", "provider.decision_ended", "provider.decision_images_off", "provider.decision_box_model", "provider.budget_exhausted", "provider.adapter_version", "provider.adapter_methods", "provider.adapter_schemas", "provider.fallback_slot", "provider.model_fallback_slot", "provider.host_model", "turn.media_refused_image", "turn.media_refused_video", "turn.media_refused_file", "provider.turn_profile", "provider.model_binding_missing", "provider.session_closed", "provider.settled_call", "provider.session_unavailable", "provider.live_input", "provider.tool_cancel_missing", "provider.kernel_gate", "provider.event_invalid", "provider.response_identity", "provider.credential_resolution", "provider.credential_shape", "provider.stored_continuation", "provider.continuation_mismatch", "provider.continuation_contract", "provider.stateless_history", "provider.events_after_completion", "provider.duplicate_call", "provider.tool_cancel_live", "provider.complete_unavailable", "provider.parallel_tools", "provider.binding_identity", "provider.profile_type", "provider.event_type", "provider.feature", "provider.input_kind", "provider.output_kind", "provider.builtin", "provider.credential_missing", "provider.profile_configuration", "continue.instruction", "canary.bind_note", "user_data.note", "context.note", "taint.blocked", "taint.reason_steered", "taint.reason_tainted", "taint.destination_blocked", "taint.destination_confirm", "advisory.notice_elevated", "advisory.notice_high", "advisory.guidance", "attachments.too_many_files", "attachments.too_many_of_kind", "attachments.file_too_large", "attachments.turn_too_large", "attachments.not_accepted", "attachments.mime_not_allowed", "attachments.limits_unconfigured", "quota.exhausted", "quota.busy", "error.config", "error.request", "error.conflict", "error.not_found", "error.input", "error.action", "error.auth", "error.rate_limit", "error.busy", "error.unsupported", "error.unavailable", "error.bad_response", "error.network", "error.timeout", "error.safety", "error.blocked", "error.declined", "error.failed", "error.cancelled", "error.internal", "repair.default_guidance", "repair.items_guidance", "repair.prompt_header", "repair.prompt_intro", "repair.prompt_instructions", "repair.history_heading", "repair.section_previous_output", "repair.section_validator_rejection", "repair.section_repair_guidance", "repair.section_instructions", "structured.answer_tool", "structured.answer_missing", "structured.cut_off", "structured.not_json", "compaction.request", "compaction.tool_call", "compaction.tool_result", "compaction.tool_files", "decision.box_image", "detect.blocked", "detect.call_blocked", "detect.output_blocked", "detect.shown_blocked", "detect.hint.ids", "detect.hint.financial", "detect.hint.network", "detect.hint.credentials", "detect.hint.injection", "detect.hint.tool_instructions", "detect.hint.canary_leak", "detect.hint.prompt_leak", "detect.hint.marker_leak", "detect.hint.ungiven_images", "detect.hint.ungiven_links", "detect.hint.tool_leak", "detect.hint.own", "egress.default_repair_guidance", "egress.refusal", "egress.refusal_cut", "egress.rejection", "egress.rejection_found", "egress.hint_unscannable", "egress.hint_provider_tool_leak", "thought.omitted_image", "thought.omitted_link", "thought.omitted_instructions", "session.abandon_gated", "session.tool_denied", "session.tool_aborted", "session.sign_in", "sign_in.link", "sign_in.pending", "sign_in.done", "sign_in.declined", "sign_in.expired", "sign_in.out_of_scope", "session.gate_expired", "session.provider_state_unowned", "session.conversation_unknown", "session.full", "session.busy", "session.turn_ended", "session.gate_pending", "session.part_skipped", "live.session_ended", "voice.unsupported", "voice.permission", "voice.unavailable", "voice.failed", "voice.empty", "stop.cancelled", "stop.interrupted", "stop.length", "stop.filtered", "stop.provider_error", "stop.stream_incomplete", "stop.step_limit", "composer.model", "composer.effort", "tool.awaiting_user", "tool.completed_hidden", "tool.t2_loader_needs_snapshot", "tool.t2_loader_shape", "tool.t2_loader_output_invalid", "tool.execution_uncertain", "tool.input_invalid", "tool.input_invalid_after_mutate", "tool.output_invalid_after_mutate", "tool.handler_no_output", "tool.did_not_run", "tool.page_no_answer", "tool.page_timed_out", "tool.output_invalid", "tool.not_loaded_t2", "tool.wait_unknown", "tool.wait_cycle", "tool.after_field", "tool.background_field", "tool.background_started", "tool.late_result", "tool.not_visible", "tool.builtin_not_enabled", "tool.provider_native", "tool.not_registered", "tool.builtin_needs_snapshot", "tool.not_allowed", "tool.not_eligible", "tool.unsupported_type", "tool.agent_call_limit", "tool.agent_failed", "tool.activity_failed", "protocol.not_served", "protocol.profile_type", "protocol.options", "protocol.option_clash", "protocol.base_options", "protocol.id_clash", "protocol.ws_headers", "protocol.builtin_unwritten", "protocol.provider_missing", "protocol.connection", "protocol.part_schema", "protocol.part_invalid", "protocol.key_invalid", "protocol.no_key", "protocol.reply_shape", "protocol.endpoint_url", "protocol.endpoint_host", "protocol.http", "protocol.no_operation", "protocol.definition", "protocol.provider_id_clash", "protocol.history_needed"];
|
|
4
|
+
export declare const LABEL_KEYS: readonly ["provider.decision_failed", "provider.decision_unavailable", "provider.decision_model_missing", "provider.decision_ended", "provider.decision_images_off", "provider.decision_box_model", "provider.budget_exhausted", "provider.adapter_version", "provider.adapter_methods", "provider.adapter_schemas", "provider.fallback_slot", "provider.model_fallback_slot", "provider.host_model", "turn.media_refused_image", "turn.media_refused_video", "turn.media_refused_file", "provider.turn_profile", "provider.model_binding_missing", "provider.session_closed", "provider.settled_call", "provider.session_unavailable", "provider.live_input", "provider.tool_cancel_missing", "provider.kernel_gate", "provider.event_invalid", "provider.response_identity", "provider.credential_resolution", "provider.credential_shape", "provider.stored_continuation", "provider.continuation_mismatch", "provider.continuation_contract", "provider.stateless_history", "provider.events_after_completion", "provider.duplicate_call", "provider.tool_cancel_live", "provider.complete_unavailable", "provider.parallel_tools", "provider.binding_identity", "provider.profile_type", "provider.event_type", "provider.feature", "provider.input_kind", "provider.output_kind", "provider.builtin", "provider.credential_missing", "provider.profile_configuration", "continue.instruction", "canary.bind_note", "user_data.note", "context.note", "taint.blocked", "taint.reason_steered", "taint.reason_tainted", "taint.destination_blocked", "taint.destination_confirm", "advisory.notice_elevated", "advisory.notice_high", "advisory.guidance", "attachments.too_many_files", "attachments.too_many_of_kind", "attachments.file_too_large", "attachments.turn_too_large", "attachments.not_accepted", "attachments.mime_not_allowed", "attachments.limits_unconfigured", "quota.exhausted", "quota.busy", "error.config", "error.request", "error.conflict", "error.not_found", "error.input", "error.action", "error.auth", "error.rate_limit", "error.busy", "error.unsupported", "error.unavailable", "error.bad_response", "error.network", "error.timeout", "error.safety", "error.blocked", "error.declined", "error.failed", "error.cancelled", "error.internal", "repair.default_guidance", "repair.items_guidance", "repair.cut_off_guidance", "repair.prompt_header", "repair.prompt_intro", "repair.prompt_instructions", "repair.history_heading", "repair.section_previous_output", "repair.section_validator_rejection", "repair.section_repair_guidance", "repair.section_instructions", "structured.answer_tool", "structured.answer_missing", "structured.cut_off", "structured.not_json", "compaction.request", "compaction.tool_call", "compaction.tool_result", "compaction.tool_files", "decision.box_image", "detect.blocked", "detect.call_blocked", "detect.output_blocked", "detect.shown_blocked", "detect.hint.ids", "detect.hint.financial", "detect.hint.network", "detect.hint.credentials", "detect.hint.injection", "detect.hint.tool_instructions", "detect.hint.canary_leak", "detect.hint.prompt_leak", "detect.hint.marker_leak", "detect.hint.ungiven_images", "detect.hint.ungiven_links", "detect.hint.tool_leak", "detect.hint.own", "egress.default_repair_guidance", "egress.refusal", "egress.refusal_cut", "egress.rejection", "egress.rejection_found", "egress.hint_unscannable", "egress.hint_provider_tool_leak", "thought.omitted_image", "thought.omitted_link", "thought.omitted_instructions", "session.abandon_gated", "session.tool_denied", "session.tool_aborted", "session.sign_in", "sign_in.link", "sign_in.pending", "sign_in.done", "sign_in.declined", "sign_in.expired", "sign_in.out_of_scope", "session.gate_expired", "session.provider_state_unowned", "session.conversation_unknown", "session.full", "session.busy", "session.turn_ended", "session.gate_pending", "session.part_skipped", "live.session_ended", "voice.unsupported", "voice.permission", "voice.unavailable", "voice.failed", "voice.empty", "stop.cancelled", "stop.interrupted", "stop.length", "stop.filtered", "stop.provider_error", "stop.stream_incomplete", "stop.step_limit", "composer.model", "composer.effort", "tool.awaiting_user", "tool.completed_hidden", "tool.t2_loader_needs_snapshot", "tool.t2_loader_shape", "tool.t2_loader_output_invalid", "tool.execution_uncertain", "tool.input_invalid", "tool.input_invalid_after_mutate", "tool.output_invalid_after_mutate", "tool.handler_no_output", "tool.did_not_run", "tool.unreachable", "tool.page_no_answer", "tool.page_timed_out", "tool.output_invalid", "tool.not_loaded_t2", "tool.wait_unknown", "tool.wait_cycle", "tool.after_field", "tool.background_field", "tool.background_started", "tool.late_result", "tool.not_visible", "tool.builtin_not_enabled", "tool.provider_native", "tool.not_registered", "tool.builtin_needs_snapshot", "tool.not_allowed", "tool.not_eligible", "tool.unsupported_type", "tool.agent_call_limit", "tool.agent_failed", "tool.activity_failed", "protocol.not_served", "protocol.profile_type", "protocol.options", "protocol.option_clash", "protocol.base_options", "protocol.id_clash", "protocol.ws_headers", "protocol.builtin_unwritten", "protocol.provider_missing", "protocol.connection", "protocol.part_schema", "protocol.part_invalid", "protocol.key_invalid", "protocol.no_key", "protocol.reply_shape", "protocol.endpoint_url", "protocol.endpoint_host", "protocol.http", "protocol.no_operation", "protocol.definition", "protocol.provider_id_clash", "protocol.history_needed"];
|
|
5
5
|
/** One of {@linkcode LABEL_KEYS}. */
|
|
6
6
|
export type LabelKey = (typeof LABEL_KEYS)[number];
|
|
7
7
|
/**
|
|
@@ -10,7 +10,7 @@ export type LabelKey = (typeof LABEL_KEYS)[number];
|
|
|
10
10
|
* puts into history. Only these overrides travel to the browser; the rest
|
|
11
11
|
* (repair prompts, canary, taint, egress) stay on the host.
|
|
12
12
|
*/
|
|
13
|
-
export declare const CLIENT_LABEL_KEYS: readonly ["attachments.too_many_files", "attachments.too_many_of_kind", "attachments.file_too_large", "attachments.turn_too_large", "attachments.not_accepted", "attachments.mime_not_allowed", "attachments.limits_unconfigured", "detect.blocked", "error.config", "error.request", "error.conflict", "error.not_found", "error.input", "error.action", "error.auth", "error.rate_limit", "error.busy", "error.unsupported", "error.unavailable", "error.bad_response", "error.network", "error.timeout", "error.safety", "error.blocked", "error.declined", "error.failed", "error.cancelled", "error.internal", "session.abandon_gated", "session.tool_denied", "session.sign_in", "session.gate_expired", "session.turn_ended", "session.gate_pending", "session.part_skipped", "live.session_ended", "tool.awaiting_user", "tool.completed_hidden", "tool.execution_uncertain", "voice.unsupported", "voice.permission", "voice.unavailable", "voice.failed", "voice.empty", "stop.cancelled", "stop.interrupted", "stop.length", "stop.filtered", "stop.provider_error", "stop.stream_incomplete", "stop.step_limit", "composer.model", "composer.effort"];
|
|
13
|
+
export declare const CLIENT_LABEL_KEYS: readonly ["attachments.too_many_files", "attachments.too_many_of_kind", "attachments.file_too_large", "attachments.turn_too_large", "attachments.not_accepted", "attachments.mime_not_allowed", "attachments.limits_unconfigured", "detect.blocked", "error.config", "error.request", "error.conflict", "error.not_found", "error.input", "error.action", "error.auth", "error.rate_limit", "error.busy", "error.unsupported", "error.unavailable", "error.bad_response", "error.network", "error.timeout", "error.safety", "error.blocked", "error.declined", "error.failed", "error.cancelled", "error.internal", "session.abandon_gated", "session.tool_denied", "session.sign_in", "session.gate_expired", "session.turn_ended", "session.gate_pending", "session.part_skipped", "live.session_ended", "tool.awaiting_user", "tool.completed_hidden", "tool.execution_uncertain", "tool.unreachable", "voice.unsupported", "voice.permission", "voice.unavailable", "voice.failed", "voice.empty", "stop.cancelled", "stop.interrupted", "stop.length", "stop.filtered", "stop.provider_error", "stop.stream_incomplete", "stop.step_limit", "composer.model", "composer.effort"];
|
|
14
14
|
/**
|
|
15
15
|
* Lines only {@linkcode overrideLabels} changes: builder and host errors from providers and
|
|
16
16
|
* protocols, and the decision box hint. A profile's `labels` never reach them.
|
|
@@ -92,6 +92,7 @@ export const LABEL_KEYS = [
|
|
|
92
92
|
'error.internal',
|
|
93
93
|
'repair.default_guidance',
|
|
94
94
|
'repair.items_guidance',
|
|
95
|
+
'repair.cut_off_guidance',
|
|
95
96
|
'repair.prompt_header',
|
|
96
97
|
'repair.prompt_intro',
|
|
97
98
|
'repair.prompt_instructions',
|
|
@@ -180,6 +181,7 @@ export const LABEL_KEYS = [
|
|
|
180
181
|
'tool.output_invalid_after_mutate',
|
|
181
182
|
'tool.handler_no_output',
|
|
182
183
|
'tool.did_not_run',
|
|
184
|
+
'tool.unreachable',
|
|
183
185
|
'tool.page_no_answer',
|
|
184
186
|
'tool.page_timed_out',
|
|
185
187
|
'tool.output_invalid',
|
|
@@ -270,6 +272,7 @@ export const CLIENT_LABEL_KEYS = [
|
|
|
270
272
|
'tool.awaiting_user',
|
|
271
273
|
'tool.completed_hidden',
|
|
272
274
|
'tool.execution_uncertain',
|
|
275
|
+
'tool.unreachable',
|
|
273
276
|
'voice.unsupported',
|
|
274
277
|
'voice.permission',
|
|
275
278
|
'voice.unavailable',
|
|
@@ -431,6 +434,7 @@ const DEFAULTS = {
|
|
|
431
434
|
'error.internal': 'Sorry, something went wrong.',
|
|
432
435
|
'repair.default_guidance': 'Revise the previous output so it satisfies the validator rejection. Preserve the intended user-facing substance unless the guidance says otherwise.',
|
|
433
436
|
'repair.items_guidance': 'The previous output holds only the items that were rejected; the rest of your reply is kept. Return only these items, corrected, in the same order.',
|
|
437
|
+
'repair.cut_off_guidance': 'Your answer was cut off before it ended. Give the whole answer again, written shorter so it fits. Keep what matters most.',
|
|
434
438
|
'repair.prompt_header': '## OUTPUT REPAIR REQUEST',
|
|
435
439
|
'repair.prompt_intro': 'The previous assistant output was rejected by a host validator and must be revised.',
|
|
436
440
|
'repair.prompt_instructions': '1. Inspect the previous output and validator rejection.\n2. Apply the repair guidance without inventing unsupported facts.\n3. Return only the corrected assistant output required by the active profile.',
|
|
@@ -517,6 +521,7 @@ const DEFAULTS = {
|
|
|
517
521
|
'tool.execution_uncertain': 'This action has an unresolved execution. Check the destination before trying it again.',
|
|
518
522
|
'tool.input_invalid': 'Tool input validation failed',
|
|
519
523
|
'tool.did_not_run': "Tool '{tool}' did not run.",
|
|
524
|
+
'tool.unreachable': "Sorry, a tool couldn't be reached. Please try again.",
|
|
520
525
|
'tool.input_invalid_after_mutate': 'Tool input validation failed after mutate',
|
|
521
526
|
'tool.output_invalid_after_mutate': 'Tool output validation failed after mutate',
|
|
522
527
|
'tool.handler_no_output': 'Handler returned no output',
|
|
@@ -648,7 +653,7 @@ export const LABEL_NOTES = {
|
|
|
648
653
|
'error.unsupported': 'Shown to the user when the model or its route cannot serve the request. Takes {tool} when one step failed.',
|
|
649
654
|
'error.unavailable': 'Shown to the user when the model provider is down. Takes {tool} when one step failed.',
|
|
650
655
|
'error.bad_response': 'Shown to the user when the model provider answers with something that cannot be used. Takes {tool} when one step failed.',
|
|
651
|
-
'error.network': 'Shown to the user when the request never reached the model provider.
|
|
656
|
+
'error.network': 'Shown to the user when the request never reached the model provider. A tool that could not be reached shows tool.unreachable.',
|
|
652
657
|
'error.timeout': 'Shown to the user when the model takes longer than the host allows. Takes {tool} when one step failed.',
|
|
653
658
|
'error.safety': 'Shown to the user when Theorem or the model provider holds the reply back. Takes {tool} when one step failed.',
|
|
654
659
|
'error.blocked': "Shown to the user when a guardrail or host policy stops one of the agent's steps. Takes {tool} when one step failed.",
|
|
@@ -658,6 +663,7 @@ export const LABEL_NOTES = {
|
|
|
658
663
|
'error.internal': 'Shown to the user when something inside Theorem breaks. Takes {tool} when one step failed.',
|
|
659
664
|
'repair.default_guidance': "Sent to the model when a validator rejects its output and the host gives no guidance of its own; also the Repair guidance setting's default.",
|
|
660
665
|
'repair.items_guidance': 'Sent to the model when an item rule rejects some items of a list in its structured reply and asks for those items alone to be rewritten.',
|
|
666
|
+
'repair.cut_off_guidance': 'Sent to the model when its answer was cut off at the output limit and it is asked to give it again, shorter.',
|
|
661
667
|
'repair.prompt_header': 'The heading that opens the repair request sent to the model after a validator or the egress check rejects its output.',
|
|
662
668
|
'repair.prompt_intro': 'The first line of the repair request, saying why the model is asked to revise.',
|
|
663
669
|
'repair.prompt_instructions': 'The numbered steps at the end of the repair request.',
|
|
@@ -743,6 +749,7 @@ export const LABEL_NOTES = {
|
|
|
743
749
|
'tool.execution_uncertain': 'Shown when a tool operation has no confirmed outcome and cannot execute again.',
|
|
744
750
|
'tool.input_invalid': "Told to the model when a tool call's input fails the tool's input schema.",
|
|
745
751
|
'tool.did_not_run': 'Given by runTool and runToolStep for a call that never settled when the invoke gave no error of its own. Takes {tool}.',
|
|
752
|
+
'tool.unreachable': 'Shown to the user when a tool call never reached its host. Takes {tool}.',
|
|
746
753
|
'tool.input_invalid_after_mutate': "Told to the model when a tool call's input fails its schema after a guardrail changed it.",
|
|
747
754
|
'tool.output_invalid_after_mutate': "Told to the model when a tool's output fails its schema after a guardrail changed it.",
|
|
748
755
|
'tool.handler_no_output': "Told to the model when a tool's handler returns nothing.",
|
|
@@ -845,11 +852,16 @@ const LABEL_PLACEHOLDERS = {
|
|
|
845
852
|
'tool.not_allowed': ['tool', 'profile'],
|
|
846
853
|
'tool.not_eligible': TOOL,
|
|
847
854
|
'tool.did_not_run': TOOL,
|
|
855
|
+
'tool.unreachable': TOOL,
|
|
848
856
|
'tool.unsupported_type': TOOL,
|
|
849
857
|
'tool.agent_call_limit': TOOL,
|
|
850
858
|
'tool.agent_failed': TOOL,
|
|
851
859
|
'tool.activity_failed': TOOL,
|
|
852
|
-
|
|
860
|
+
// why: a tool that could not be reached is worded by tool.unreachable, so error.network never names one.
|
|
861
|
+
...Object.fromEntries(LABEL_KEYS.filter((key) => key.startsWith('error.') && key !== 'error.network').map((key) => [
|
|
862
|
+
key,
|
|
863
|
+
TOOL,
|
|
864
|
+
])),
|
|
853
865
|
};
|
|
854
866
|
/** Placeholders an override for a key must keep (mechanism-critical tokens). */
|
|
855
867
|
const REQUIRED_PLACEHOLDERS = {
|
|
@@ -353,14 +353,21 @@ function startItemFix(args) {
|
|
|
353
353
|
}
|
|
354
354
|
/** A structured reply the attempt wrote that never parsed, and what went wrong with it. */
|
|
355
355
|
function unparsedReply(state, generation, profile) {
|
|
356
|
-
if (!generation.structured
|
|
356
|
+
if (!generation.structured)
|
|
357
357
|
return undefined;
|
|
358
|
-
const
|
|
358
|
+
const byTool = answersByTool(generation);
|
|
359
|
+
const text = byTool ? (state.cutOffAnswer ?? '') : collectAttemptText(state.attemptEvents);
|
|
359
360
|
if (!text.trim())
|
|
360
361
|
return undefined;
|
|
361
362
|
if (state.lastStop?.kind === 'length') {
|
|
362
|
-
return {
|
|
363
|
+
return {
|
|
364
|
+
text,
|
|
365
|
+
rejection: labelText('structured.cut_off', {}, profile.labels),
|
|
366
|
+
guidance: labelText('repair.cut_off_guidance', {}, profile.labels),
|
|
367
|
+
};
|
|
363
368
|
}
|
|
369
|
+
if (byTool)
|
|
370
|
+
return undefined;
|
|
364
371
|
const error = findLast(state.attemptEvents, (event) => event.type === 'error');
|
|
365
372
|
if (state.lastStop?.kind === 'provider_error' &&
|
|
366
373
|
error?.errorKind === 'bad_response' &&
|
|
@@ -452,6 +459,38 @@ held) {
|
|
|
452
459
|
yield failure;
|
|
453
460
|
return 'pass';
|
|
454
461
|
}
|
|
462
|
+
/**
|
|
463
|
+
* A structured reply the output limit cut off: a repair that asks for it again, shorter, while the
|
|
464
|
+
* budget lasts, else a `bad_response` failure. The budget is one repair, or `outputs.validation.maxRetries`.
|
|
465
|
+
*
|
|
466
|
+
* why: A cut-off reply holds no answer to check, with or without `outputs.validation`; without a
|
|
467
|
+
* repair the turn would end with none.
|
|
468
|
+
*/
|
|
469
|
+
function* handleCutOffReply(flow, state, profile, latestStructured,
|
|
470
|
+
/** True when the attempt's events go out at its end (`yieldBufferedAttemptEvents`). */
|
|
471
|
+
held) {
|
|
472
|
+
if (latestStructured !== undefined || flow.pendingItemFix)
|
|
473
|
+
return 'pass';
|
|
474
|
+
const unparsed = unparsedReply(state, flow.currentGen, profile);
|
|
475
|
+
if (!unparsed?.guidance)
|
|
476
|
+
return 'pass';
|
|
477
|
+
if (flow.currentAttempt < cutOffBudget(profile)) {
|
|
478
|
+
const nextRequest = buildRepairRequest(flow.currentReq, unparsed.text, unparsed.rejection, unparsed.guidance);
|
|
479
|
+
yield updateFlowForRetry(flow, state, profile, nextRequest, 'validation');
|
|
480
|
+
return 'continue';
|
|
481
|
+
}
|
|
482
|
+
const failure = toErrorEvent(new TheoremError('bad_response', 'the reply was cut off at the output limit'));
|
|
483
|
+
recordStepEvent(failure, state);
|
|
484
|
+
state.lastStop = { kind: 'provider_error' };
|
|
485
|
+
if (!held)
|
|
486
|
+
yield failure;
|
|
487
|
+
return 'pass';
|
|
488
|
+
}
|
|
489
|
+
/** How many times a structured reply cut off at the output limit is repaired. */
|
|
490
|
+
function cutOffBudget(profile) {
|
|
491
|
+
const outputs = profileTurnOutputs(profile);
|
|
492
|
+
return outputs?.structured ? Math.max(1, outputs.validation?.maxRetries ?? 0) : 0;
|
|
493
|
+
}
|
|
455
494
|
/** Puts on the attempt's structured output the fields that still failed. */
|
|
456
495
|
function markFailed(state, failed) {
|
|
457
496
|
const held = findLast(state.attemptEvents, (event) => event.type === 'structured');
|
|
@@ -476,6 +515,7 @@ async function* executeSingleAttemptCycle(args) {
|
|
|
476
515
|
// invariant: The attempt's reply spans its before_end re-entries; only a new attempt starts it again.
|
|
477
516
|
state.replyText = '';
|
|
478
517
|
state.replyMessages = [];
|
|
518
|
+
state.cutOffAnswer = undefined;
|
|
479
519
|
// invariant: The leak check's carry is the end of this attempt's reply; a retry's is read on its own.
|
|
480
520
|
state.canaryCarry = undefined;
|
|
481
521
|
let latestStructured;
|
|
@@ -545,7 +585,11 @@ async function* checkAttemptReply(flow, state, profile, system, structured) {
|
|
|
545
585
|
}
|
|
546
586
|
latestStructured = projectOutbound(state.attemptEvents).structured ?? latestStructured;
|
|
547
587
|
}
|
|
548
|
-
|
|
588
|
+
const held = Boolean(validation) || judged;
|
|
589
|
+
if ((yield* handleMissingAnswer(flow, state, profile, latestStructured, held)) === 'continue') {
|
|
590
|
+
return { status: 'continue' };
|
|
591
|
+
}
|
|
592
|
+
if ((yield* handleCutOffReply(flow, state, profile, latestStructured, held)) === 'continue') {
|
|
549
593
|
return { status: 'continue' };
|
|
550
594
|
}
|
|
551
595
|
if (validation) {
|
|
@@ -567,7 +611,7 @@ async function* checkAttemptReply(flow, state, profile, system, structured) {
|
|
|
567
611
|
async function* runAttemptsWithValidation(safe, profile, generation, system, provider, state) {
|
|
568
612
|
const { blockedReply } = resolveGuardrailPolicy(profile.guardrails);
|
|
569
613
|
// why: Each gate counts the attempts against its own cap; the turn runs while either has one left.
|
|
570
|
-
const maxRetries = Math.max(profileTurnOutputs(profile)?.validation?.maxRetries ?? 0, blockedReply.onBlock === 'retry' ? blockedReply.maxRetries : 0);
|
|
614
|
+
const maxRetries = Math.max(profileTurnOutputs(profile)?.validation?.maxRetries ?? 0, blockedReply.onBlock === 'retry' ? blockedReply.maxRetries : 0, cutOffBudget(profile));
|
|
571
615
|
const flow = {
|
|
572
616
|
currentAttempt: 0,
|
|
573
617
|
currentGen: generation,
|
|
@@ -59,6 +59,8 @@ interface StepExecutionState {
|
|
|
59
59
|
message: TurnHistoryMessage;
|
|
60
60
|
at: number;
|
|
61
61
|
}[];
|
|
62
|
+
/** The answer tool's arguments as this attempt wrote them before the output limit cut them off. */
|
|
63
|
+
cutOffAnswer?: string;
|
|
62
64
|
/** The message `input.lateResults` opened, until the run can emit it. */
|
|
63
65
|
lateResults?: string;
|
|
64
66
|
/** The canary opening the last provider call ended on, read in front of the next call's reply. */
|
|
@@ -144,12 +144,33 @@ async function* showAnswer(shown, tally, step) {
|
|
|
144
144
|
yield* shown.show(event);
|
|
145
145
|
}
|
|
146
146
|
/**
|
|
147
|
-
*
|
|
147
|
+
* An answer the output limit cut off: its drafts, kept on the attempt for the repair. The call
|
|
148
|
+
* stops at `length`, and the failure its broken arguments gave is not the turn's.
|
|
149
|
+
*
|
|
150
|
+
* why: A protocol hands over a call whose arguments do not parse as a bare `bad_response`, and
|
|
151
|
+
* drops the call; the stream then stops it as a `provider_error`. Its drafts are what the model
|
|
152
|
+
* wrote, and the cut is why it broke.
|
|
153
|
+
*/
|
|
154
|
+
function keepCutOffAnswer(tally, state, generation) {
|
|
155
|
+
if (!tally.cut || !answersByTool(generation) || !tally.answerDraft.trim())
|
|
156
|
+
return;
|
|
157
|
+
if (tally.calls.some((call) => call.name === ANSWER_TOOL && !call.failure))
|
|
158
|
+
return;
|
|
159
|
+
state.cutOffAnswer = tally.answerDraft;
|
|
160
|
+
tally.failures = tally.failures.filter((failure) => failure.errorKind !== 'bad_response');
|
|
161
|
+
tally.stop = { kind: 'length' };
|
|
162
|
+
state.lastStop = tally.stop;
|
|
163
|
+
}
|
|
164
|
+
/**
|
|
165
|
+
* A tool call's preview. The answer tool's streams as the structured reply's draft, and the call
|
|
166
|
+
* keeps it, for the repair of an answer the output limit cuts off (`keepCutOffAnswer`).
|
|
148
167
|
*
|
|
149
168
|
* why: A preview of the call that follows, which owns the arguments: not the turn's record. It streams
|
|
150
169
|
* as the reply's text does, through its own gate, and like that text is held only when delivery is buffered.
|
|
151
170
|
*/
|
|
152
|
-
async function* showToolDraft(event, shown, generation, buffer) {
|
|
171
|
+
async function* showToolDraft(event, shown, generation, buffer, tally) {
|
|
172
|
+
if (event.name === ANSWER_TOOL)
|
|
173
|
+
tally.answerDraft += event.text;
|
|
153
174
|
if (buffer.holdVisible)
|
|
154
175
|
return;
|
|
155
176
|
if (event.name === ANSWER_TOOL && answersByTool(generation))
|
|
@@ -253,7 +274,15 @@ async function* executeAutonomousStep(args, state, buffer = {
|
|
|
253
274
|
}) {
|
|
254
275
|
const { profile, generation, system, provider, signal } = args;
|
|
255
276
|
const { usage, genForStep, request, call } = startProviderCall(generation, system, state, isAnswerOnlyStep(profile, generation, state.stepCount), profile.labels);
|
|
256
|
-
const tally = {
|
|
277
|
+
const tally = {
|
|
278
|
+
calls: [],
|
|
279
|
+
reasoning: [],
|
|
280
|
+
text: '',
|
|
281
|
+
failures: [],
|
|
282
|
+
replied: false,
|
|
283
|
+
answerDraft: '',
|
|
284
|
+
cut: false,
|
|
285
|
+
};
|
|
257
286
|
const control = {
|
|
258
287
|
withholdVisible: false,
|
|
259
288
|
...(state.canaryCarry ? { canaryCarry: state.canaryCarry } : {}),
|
|
@@ -288,12 +317,13 @@ async function* executeAutonomousStep(args, state, buffer = {
|
|
|
288
317
|
}
|
|
289
318
|
if (event.type === 'done') {
|
|
290
319
|
state.providerState = event.providerState;
|
|
320
|
+
tally.cut ||= event.stop.kind === 'length';
|
|
291
321
|
tally.stop = event.stop;
|
|
292
322
|
state.lastStop = event.stop;
|
|
293
323
|
continue;
|
|
294
324
|
}
|
|
295
325
|
if (event.type === 'tool_draft') {
|
|
296
|
-
yield* showToolDraft(event, shown, generation, buffer);
|
|
326
|
+
yield* showToolDraft(event, shown, generation, buffer, tally);
|
|
297
327
|
continue;
|
|
298
328
|
}
|
|
299
329
|
// why: The model's history keeps what it wrote; only the reply the host gets is set apart.
|
|
@@ -310,6 +340,7 @@ async function* executeAutonomousStep(args, state, buffer = {
|
|
|
310
340
|
throw err;
|
|
311
341
|
}
|
|
312
342
|
yield* shown.end();
|
|
343
|
+
keepCutOffAnswer(tally, state, generation);
|
|
313
344
|
yield* showAnswer(shown, tally, { generation, state, control, buffer, call });
|
|
314
345
|
keepStreamCarry(state, control);
|
|
315
346
|
const resume = resumeOf(args.mayResume, profile, tally) ?? mediaResumeOf(state, profile, tally);
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "@theoremjs/agents",
|
|
3
|
-
"version": "0.3.4-next.
|
|
3
|
+
"version": "0.3.4-next.815.g7f5fa76",
|
|
4
4
|
"description": "A flat TypeScript agent kernel for typed profiles, deterministic turn execution, registered tools, provider adapters, guardrails, and host-injected traces.",
|
|
5
5
|
"keywords": [
|
|
6
6
|
"agent",
|