agent-afk 5.307.7 → 5.307.9

package/README.md

@@ -59,6 +59,10 @@ Codex imports use `CODEX_HOME` when set (otherwise `~/.codex`) and also discover
 
 **Jev calibrated judgment.** Jev is an external cross-family judge you configure as an MCP server in `~/.afk/config/mcp.json`. When present, `afk whatif --judge auto` selects Jev to grade verification episodes; it returns calibrated probabilities and removes Anthropic self-preference bias in verdicts. Falls back to Claude when Jev is not configured. Use `--judge claude` to keep all episode data within Anthropic. Source: `src/whatif/judge/jev-connect.ts`, `docs/whatif.md`.
 
+## What it has done
+
+**Constructed SW(25,16), a math case a 2025 paper listed as open.** Given a one-sentence goal ("settle one published open question"), agent-afk agents picked the target, designed a structured SAT search, found a symmetric 25x25 weighing matrix in 760 seconds of solver time, checked it with independent exact checkers and a prior-work search, and wrote up a short proof. The author of the paper that listed the case as open confirmed it passes his verifier. Check it yourself in under a second: [`docs/results/sw-25-16`](docs/results/sw-25-16/README.md).
+
 ## How it compares
 
 | | Claude Code | Agent AFK |

package/dist/agent/providers/anthropic-direct/cache-policy.d.ts

@@ -1,4 +1,5 @@
 import type { ContentBlockParam, MessageParam } from '@anthropic-ai/sdk/resources';
+export declare const MAX_CACHE_BREAKPOINTS = 4;
 export declare function isCacheEnabled(opts?: {
     baseUrl?: string;
 }): boolean;
@@ -8,3 +9,24 @@ export declare function isExtendedCacheTtlActive(opts?: {
 }): boolean;
 export declare function withSystemBreakpoint(blocks: ContentBlockParam[], ttl: '5m' | '1h'): ContentBlockParam[];
 export declare function withMessagesBreakpoint(messages: MessageParam[], ttl: '5m' | '1h'): MessageParam[];
+export declare function countBreakpoints(params: {
+    system?: readonly ContentBlockParam[];
+    tools?: readonly CacheableToolLike[];
+    messages: readonly MessageParam[];
+}): number;
+export interface CacheableToolLike {
+    name: string;
+    cache_control?: {
+        type: string;
+    } | null;
+    [k: string]: unknown;
+}
+export declare function clampBreakpoints(params: {
+    system?: readonly ContentBlockParam[];
+    tools?: readonly CacheableToolLike[];
+    messages: readonly MessageParam[];
+}): {
+    system?: readonly ContentBlockParam[];
+    tools?: readonly CacheableToolLike[];
+    messages: readonly MessageParam[];
+};

package/dist/agent/tools/system-prompt.d.ts

@@ -13,6 +13,6 @@ export declare const WORKSPACE_SYSTEM_PROMPT = "# Shared Workspace\n\nWhen dispa
 export declare const MEMORY_SYSTEM_PROMPT = "# Cross-Session Memory\n\nYou have three tools for persisting knowledge across sessions: memory_search, memory_update, and procedure_write.\n\n## Reading memory\nOn your first turn, decide whether to call memory_search based on the request:\n- Search when the task involves ongoing work, user preferences, project conventions, or prior context \u2014 e.g. repo-specific work, multi-session projects, \"like last time\", or anything where continuity matters.\n- Skip for clearly self-contained requests \u2014 one-off questions, simple lookups, or tasks with no plausible prior context.\n- If hot memory (shown in <cross-session-memory> tags above) already covers the relevant context, skip the search.\n- Search at most once per session for general context. Search again only if new information surfaces a specific topic worth querying.\n\nUse FTS5 syntax: \"exact phrase\", term1 AND term2, prefix*.\n\n## Writing memory (memory_update)\nStore facts when you encounter:\n- User preferences or corrections (\"I prefer X\", \"don't do Y\") \u2192 category: preference\n- Key decisions with rationale (\"we chose X over Y because Z\") \u2192 category: decision\n- Non-obvious project conventions discovered during investigation \u2192 category: convention\n- Surprising learnings from debugging or exploration \u2192 category: learning\n\nDo NOT store: ephemeral task details, information derivable from code or git, speculative observations.\n\n### Hot memory vs. fact archive\n- target \"fact\" \u2192 searchable SQLite archive. **This is the default home for almost everything** \u2014 project stack, conventions, file maps, decisions, learnings. It is unbounded and searchable. When in doubt, it's a fact.\n- target \"hot\" \u2192 HOT.md, injected verbatim into EVERY future session's system prompt, on every surface. Reserve it for the few lines you'd want present in every session forever: user identity, 2\u20133 top durable preferences, and a one-line pointer to the active project (name + path) \u2014 NOT its full context. Hard ~1,500-token cap; over-cap writes are truncated from the END, so order entries most-durable first (identity), least-durable last. If something doesn't need to be in every prompt, it's a fact, not hot.\n- Use action \"supersede\" (not set + remove) when updating an existing fact \u2014 preserves history.\n- Never write to hot memory during an end-of-session reflection pass \u2014 hot entries should be written only when the user explicitly states a durable preference or identity, not inferred from task outcomes.\n\n## Procedures (procedure_write)\nSave reusable multi-step workflows the user teaches you or that you discover work well. Name in kebab-case. Searchable via memory_search.";
 export declare const MEMORY_SYSTEM_PROMPT_READONLY = "# Cross-Session Memory (child session)\n\nYou have access to memory_search and memory_update (target:\"fact\" only). Hot memory writes (target:\"hot\") and procedure_write are not available in this child session \u2014 only the parent can write to hot memory or procedures.\n\n## Reading memory\nOn your first turn, decide whether to call memory_search based on the request:\n- Search when the task involves ongoing work, user preferences, project conventions, or prior context \u2014 e.g. repo-specific work, multi-session projects, \"like last time\", or anything where continuity matters.\n- Skip for clearly self-contained requests \u2014 one-off questions, simple lookups, or tasks with no plausible prior context.\n- If hot memory (shown in <cross-session-memory> tags above) already covers the relevant context, skip the search.\n- Search at most once per session for general context. Search again only if new information surfaces a specific topic worth querying.\n\nUse FTS5 syntax: \"exact phrase\", term1 AND term2, prefix*.\n\n## Persisting facts (memory_update, target:\"fact\" only)\nStore findings when you encounter non-obvious facts worth persisting: project conventions, key decisions, surprising learnings. Use target:\"fact\" \u2014 it goes to the searchable SQLite archive. Do NOT use target:\"hot\" (blocked in child sessions).";
 export declare const MEMORY_SYSTEM_PROMPT_SEARCH_ONLY = "# Cross-Session Memory (read-only)\n\nYou have access to memory_search only. No memory-write tools are available in this session.\n\n## Reading memory\nOn your first turn, decide whether to call memory_search based on the request:\n- Search when the task involves ongoing work, user preferences, project conventions, or prior context \u2014 e.g. repo-specific work, multi-session projects, \"like last time\", or anything where continuity matters.\n- Skip for clearly self-contained requests \u2014 one-off questions, simple lookups, or tasks with no plausible prior context.\n- If hot memory (shown in <cross-session-memory> tags above) already covers the relevant context, skip the search.\n- Search at most once per session for general context. Search again only if new information surfaces a specific topic worth querying.\n\nUse FTS5 syntax: \"exact phrase\", term1 AND term2, prefix*.";
-export declare function resolveToolSystemPrompt(isSkillDispatch: boolean | undefined): string;
+export declare function resolveToolSystemPrompt(isSkillDispatch: boolean | undefined, isUnnamedWorker?: boolean | undefined): string;
 export declare function resolveMemorySystemPrompt(readOnlyMemory: boolean | undefined, readOnlyState?: boolean | undefined): string;
 export declare function resolveWorkspaceSystemPrompt(workspaceEnabled: boolean | undefined): string;

package/dist/agent/types/config-types.d.ts

@@ -107,6 +107,7 @@ export interface AgentConfig {
     phaseRole?: 'read-only' | 'read-write';
     isSkillDispatch?: boolean;
     skillDispatchName?: string;
+    isUnnamedWorker?: boolean;
     isNonInteractive?: boolean;
     autoCompact?: boolean | {
         threshold: number;