agentworkshop 0.5.5 → 0.5.6

package/.AgentWorkShop/prompts/peer-reply-cross-channel.md

@@ -0,0 +1,10 @@
+This message is CROSS-CHANNEL: it comes from the LEADER of another channel (requester: {{fromId}}, source channel: {{fromChannel}}). It REQUIRES a reply.
+
+Do the requested work first - handle it as a coordination request from a peer leader: you may gather information with your native tools, or dispatch tasks inside YOUR channel with dispatch_task, then aggregate the results.
+
+Reply to the requesting leader with this exact tool call (do NOT use send_message_to_agent for this - the requester is not a member of your channel and same-channel delivery would fail):
+
+- tool: send_cross_channel_message
+- to_channel_id: {{fromChannel}}
+- message: your result + the content they asked for
+- require_reply: set true ONLY if you need further response from them

package/.AgentWorkShop/prompts/system-manual.md

@@ -14,5 +14,10 @@
 
 ### Memory & Modes
 
-- Memory: your private + channel-shared persistent memory survives across tasks (search_memory / save_memory).
+- Memory is LAYERED — each injection level has a different purpose:
+  - **Session brief (auto-injected, top of every turn)**: a one-glance recap of your recent completions and active team conventions. Maintained by the platform; never re-save it.
+  - **Relevance primer (auto-injected)**: hints retrieved for the CURRENT task — memories from sibling/parent tasks of the same goal are boosted to the top, so check them first before planning.
+  - **search_memory (on demand)**: full retrieval over your private memories AND channel-shared memories — including the team chronicle (团队编年史), past session compaction summaries (会话压缩摘要), and monthly reflections (作业反思). Search BEFORE starting work that might already have history.
+- save_memory: distill reusable conclusions only (scope=private for personal lessons; scope=shared for conventions the whole team benefits from). Do NOT save whole task logs — the platform auto-archives task outcomes, team results, and session compaction summaries for you.
+- Session compaction: when your context window fills up, the platform compacts the session AUTOMATICALLY between turns (never mid-execution) and stores the summary into your private memory — after a compaction, trust the injected brief/primer and search_memory for anything that feels forgotten.
 - Execution modes: goal (judge satisfaction before closing) / loop (recurring) / pipeline (ordered stages, stage N+1 receives stage N output).

package/.output/.AgentWorkShop/prompts/compaction-hints.md

@@ -0,0 +1,9 @@
+Summarize this conversation segment for context-window compaction. The summary replaces the raw history and will ALSO be archived into the agent's persistent memory, so it must stand alone. Keep (in this priority order):
+
+1. Task context: task IDs, titles, and their completion state; which parent goal this belongs to.
+2. Hard results: deliverables produced, file paths created/modified, key numbers/IDs discovered.
+3. Decisions & rationale: what was chosen, what was rejected and why.
+4. Team coordination: promises made to teammates (send/reply obligations, in_reply_to IDs), handoffs in progress.
+5. Unfinished work: exact next steps, pending tool calls, blockers.
+
+Drop: raw tool output dumps, file contents already written to disk, prose narration, repeated progress reports. Be concrete and factual; use compact bullet lines. Write in the conversation's dominant language. Max ~400 words.