golast

name	golast
description	Recover the previous session's last actionable request for the current workspace, but only at the very start of a fresh session. Use when the user's first substantive turn asks to continue the previous session, recover after compaction/crash, resume unfinished work, or when they send only `golast` or the skill link. Do not auto-trigger later in an already active conversation.

Resume Last

Use this skill when the user wants to continue the previous session for the current workspace. Invoke it as $golast.

If the user's message is only golast or only the skill link, treat it as an execution request: "recover the previous session for the current workspace and continue from it now".

Start-only gate

Use this skill only when the request appears at the beginning of the current session.

• Allowed: the first substantive user turn says golast, pastes the skill link, asks to continue the previous session, or asks to recover after a crash/compaction.
• Not allowed for automatic invocation: later in an already active session after substantive work has already happened in the current thread.
• In an active session, do not auto-trigger from vague phrases such as continue, 继续, 接着做, 继续 hbcore 任务, or similar wording. Prefer the current thread context instead.
• Explicit override: if the user later explicitly types golast or pastes the skill link, you may still execute this skill because that is no longer an automatic invocation.

Quick start

1. Run the bundled script:

   • Path: scripts/resume_last_request.py
   • Recommended args: --cwd <current-workspace> --format json
   • Default --scope auto: search the exact cwd, then the git root, then the most recent child-workspace thread under the current workspace.

2. Read these output fields:

   • resolved_request: the best request to continue from now
   • literal_last_user_message: the literal last user message in the previous thread
   • last_conversation_role: the role of the last meaningful conversation item near the end of the thread
   • last_conversation_content: the last meaningful conversation item near the end of the thread
   • resolved_source: whether the request came from the previous user message or the previous assistant suggestion
   • needs_more_context: whether the last exchange still looks context-dependent
   • ambiguity_hints: why the script thinks the recovered request may still be ambiguous
   • context_expanded_upward: whether the script had to keep walking upward to find an explanation
   • supporting_context: nearby earlier entries, automatically expanded upward when terms such as 三端 / 三种 / 这个方案 still look under-specified
   • assistant_message_before_last_user: the previous assistant message when the last user message was an agreement such as ok or agreed
   • previous_context_message: the previous main context message when the last user message is only a supplement such as 补充：...
   • decision_basis_message: the earlier user requirement that defines a candidate set or selection criteria, for example comparing ZLMediaKit / MediaMTX / SRS before choosing or fusing a third backend

3. Continue in the current session:

   • If the user wants execution, briefly state the recovered request and continue immediately.
   • Do not confuse "last user request" with "last conversation content". When the user asks what the previous session ended with, prefer last_conversation_content.
   • If the literal last user message is only a continuation cue such as continue, use resolved_request instead.
   • If the literal last user message is an agreement such as ok, agreed, hao, or tongyi, prefer the previous assistant action suggestion.
   • If the literal last user message starts with a supplement cue such as 补充, 另外, 还有, PS, or one more thing, merge the previous main assistant or user context into resolved_request.
   • If needs_more_context is true, read supporting_context before deciding how to continue.
   • If ambiguity_hints is non-empty, treat supporting_context as required reading even when the main request already looks actionable.
   • If decision_basis_message is non-empty, prefer it over any narrower assistant shorthand when reconstructing the true user requirement.
   • If context_expanded_upward is true, trust the expanded supporting_context over your first guess about shorthand terms. This specifically covers cases where the recovered text says things like 三端, 三种后端, 这个方案, or 按上面那样, but the concrete explanation lives further up in the previous session.
   • If the user wants only the recovered text, return only resolved_request.

Rules

• Ignore boilerplate such as # AGENTS.md instructions ..., <turn_aborted>, or a lone codex.
• Do not fabricate a previous request if nothing is found.
• If the current session has already moved past its opening turn, do not auto-use this skill unless the user explicitly invokes golast or the skill link.
• Unless the user explicitly asks for text only, resume the work instead of stopping after recovery.