메뉴
Resume an interrupted mid-task session — reconstructs state from context files and picks up at the right step
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Convert PDF/EPUB library to Markdown and generate Obsidian MOC notes
Hook-based compaction suggestions at logical task boundaries
Context window management — track spend, decide when to compact, preserve state
Session-start orientation — loads context, surfaces learnings, confirms registry
Quality and semantic review — catches what automated tools miss
Planner → Architect → Critic deliberation loop — produces a formally validated ADR
| name | resume |
| description | Resume an interrupted mid-task session — reconstructs state from context files and picks up at the right step |
| version | 0.1.0 |
| level | 2 |
| triggers | ["resume","pick up where we left off","continue the task","where were we","/resume"] |
| context_files | ["context/learnings.md","context/project.md","context/decisions.md"] |
| steps | [{"name":"Load State","description":"Read context/learnings.md for the most recent state entry. Identify task, step, completed work, and next action."},{"name":"Load Decisions","description":"Read context/decisions.md for any PROPOSED decisions from the interrupted session"},{"name":"Reconstruct Context","description":"Read the files relevant to the interrupted task to rebuild working context"},{"name":"State Brief","description":"Present a reconstruction brief — what was done, what was next, any open blockers"},{"name":"Confirm Resumption","description":"Ask the user to confirm the reconstructed state before proceeding"},{"name":"Execute Next Step","description":"Once confirmed, execute the specific next action from the interrupted task"}] |
Reconstructs a mid-task session from context files and picks up at the right step. Used when a session ends before a task is complete — due to context limits, interruption, or a natural break.
Without resume, a new session starts completely cold. Claude reads the conversation history (if available) or asks the user to re-explain where things are. The user re-explains. Time is wasted. Worse: if the session was compacted or the conversation history is unavailable, the state is gone. The work may be partially complete in ways that are invisible until something breaks.
Resume is the recovery mechanism for the pre-compact state capture written by token-budget and wrap-up.
[YYYY-MM-DD] state — Task: [name] | Step: [N] of [N] ([step name]) | Done: [list] | Next: [specific action] | Blocker: [if any]
Primary source: Check for context/session-memory.md. If it exists, read all 8 sections:
The session memory file is written by /wrap-up and is the authoritative state record.
Fallback source (if session-memory.md is absent): Read context/learnings.md. Find the most recent entry with category state. Extract:
If neither source exists: the session was not using the wrap-up protocol. Tell the user and ask them to describe where things are manually.
Read context/decisions.md. Find any entries with Status: PROPOSED — these are decisions made during the interrupted session that have not been implemented yet.
Read the files most relevant to the next action. Do not read everything — read what is needed to execute the next step.
Use the task name and step description from the state entry to identify which files to read.
Present the reconstruction:
Resume Brief — [task name]
Interrupted at: Step [N] of [N] — [step name]
Completed:
- [item]
- [item]
Next action: [specific action from state entry]
Open decisions: [list from decisions.md, or "none"]
Blocker: [from state entry, or "none"]
Files loaded: [list of files read for context]
Ask: "Does this match your understanding? Proceed? (yes / no — tell me what changed)"
Do not proceed to execution until confirmed. The state entry may be stale or incomplete.
Execute exactly the next action from the state entry. Not the whole remaining plan — the specific next step. After completing it, run a checkpoint.
If the state entry is unclear or the reconstructed context does not match the files (e.g., the files were modified since the state was written), surface the discrepancy explicitly:
Discrepancy detected:
State says: [what state entry says]
Files show: [what the current files show]
Which is correct? (state / files / explain)
Do not guess. Incorrect resumption can overwrite work that was done.
If an agent run shows degeneration signals, do not attempt to resume normally. Instead, trigger a clean reset:
Degeneration signals (any one is sufficient):
Reset protocol:
What to preserve across a reset:
What to discard:
Do not attempt to "salvage" a degenerated run by continuing from the last coherent step. Context pollution from a bad run will corrupt the resumed run. Reset and restart clean.
Critical: If the interrupted session tried approaches that failed, those must be documented in context/learnings.md and must not be retried during resumption.
Look for learnings entries with category mistake or approach near the state entry timestamp. These document dead ends.
Example learnings entry:
[2026-03-28] mistake | Tried using async/await for database queries but hit race conditions. Reverted to synchronous approach. DO NOT retry async approach without transaction locking.
When resuming, if a next action would re-attempt a documented dead end:
This prevents re-thrashing: repeating failed approaches because the session context was lost.
If you try an approach that fails during the resumed session:
context/learnings.md with category mistake or approachDo not try to resume from memory or conversation history. Only use context files — they are the ground truth.
Do not execute multiple steps before getting confirmation. Confirm state, then execute one step, then checkpoint.
Do not skip the confirm step even if the state looks clear. The user may have made changes between sessions.