recovering-session-context

Reconstruct what happened in prior agent sessions from their transcripts and surface everything raised but never resolved — unanswered questions, work promised and not done, topics dropped, partial co

SKILL.md

GitHub

Recovering Session Context

Agent sessions end badly more often than they end cleanly: interrupted mid-task, compacted, or simply overloaded until an early question is buried under later context. The transcript still holds all of it. This skill turns that transcript back into an accurate picture of what is actually outstanding.

Two modes, same machinery:

Mode Scope Ends with
Resume The tail of the single most recent session Continue the work — do not ask permission
Audit Every session for this project over a lookback window (default 3 days) Grouped findings, then an offer to track or act

They differ only in breadth and output. Locating transcripts, reconstructing turns, and deciding what counts as unresolved are identical, which is why they live here once rather than in each command that calls them.

When to use

  • Resuming after an interruption, a crash, a compaction, or a multi-day gap.
  • The user asks "what was I doing", "pick up where we left off", "what's still open", "what did we drop", "what's outstanding".
  • Before declaring a piece of work finished — to catch the "I'll also…" that never happened.
  • Auditing whether a plan was actually executed end to end, or only partly.

Not for searching the historical corpus by topic — that is cc-recall (discoverability across tens of thousands of sessions). This skill reads recent transcripts in full.

Quick reference

Need to… Read this
Find this project's transcripts, reconstruct turns, extract incomplete todos references/locating-and-reconstructing.md
Decide what counts as unresolved; format the brief or the audit references/what-counts-as-unresolved.md

Behavioral guidance

Cast a wide net, then narrow. The failure mode is reading the transcript for the task and missing everything adjacent to it. Surface candidates generously first; discard afterwards.

An unanswered question runs in both directions. Agents reliably notice the user's unanswered questions and reliably forget their own — a question the agent asked that the user never replied to is the single most common dropped thread, because the agent moved on and nothing marked it.

Resume mode continues; it does not ask. Present the brief and proceed to the next step. "Shall I continue?" wastes the turn that reconstructing the context just paid for. The one exception is a genuinely ambiguous final state — then say so in a sentence and ask exactly one focused question.

Scale the output to what you found. Three lines for a simple interrupted task. Grouped bullets only when several threads were genuinely in flight. A long report about a short session is noise.

Be specific enough to execute. "Follow up on the auth thing" is not a finding. Name the file, the decision, the question, and the session it came from — the reader should be able to act without re-reading the transcript you just read.