Carries long-running work across a context reset.

한국어 문서는 README-ko.md를 참고하세요.

In Claude Code:

/plugin marketplace add somaz94/claude-plugins
/plugin install session-continuity@somaz94

Or from your shell, without an interactive session:

claude plugin marketplace add somaz94/claude-plugins
claude plugin install session-continuity@somaz94

Work that spans days does not fit in one context window. The window fills, the session compacts or ends, and the next one starts blind — re-deriving decisions that were already made, re-reading files that were already understood, and sometimes redoing work that was already committed.

The fix is not a longer window. It is writing down the two things a fresh session actually needs: what was decided and what to do next. This plugin owns both halves.

plan-update keeps the record honest as you go. handoff reads that record and turns it into an entry point. Use the first often and the second once, and the next session starts working instead of archaeologising.

The plugin ships a PreCompact hook. When the window is about to be compacted — automatically at the limit, or because you ran /compact — it says so at that exact moment and offers /session-continuity:handoff now.

That moment is the point. Compaction summarizes older turns, so the half-finished edit and the reason behind a decision are precisely what is about to blur. Claude Code exposes no hook that fires at a fixed context percentage, so an imminent compaction is the only "filling up" signal available.

It is advisory: it never cancels a compaction, never opens a session, and never runs the handoff itself. A parse error fails open and stays silent — a nudge is a courtesy, not a boundary.

Produces a single fenced block — copy once, paste once. It runs in one of two modes.

Plan-based (the default): a plan file under ~/.claude/plans/ is the anchor. The prompt opens by telling the new session to read it, then lists the next actions, the decisions already locked in, the sub-agents worth reaching for, and any <TBD> the previous session left behind.

Mid-session: the window is filling mid-work and there may be no plan yet. The prompt leads with the in-flight state instead — branch, the file:line left half-edited, uncommitted diff, anything running in the background — plus what is already committed and must not be redone.

The agent never invents a plan path that is not on disk, and never guesses an answer the user has not given: open questions are surfaced as <needs user input> rather than resolved on your behalf.

Edits a plan file in place after a phase completes, a blocker clears, or scope changes. Progress markers flip, a one-line "what changed" entry is appended, and work products are cited concretely — a link to what shipped beats the word "completed".

Overridden decisions are appended to, never deleted. A plan's value to a future session includes the turns it did not take, so a reversed decision gets an → **overridden, YYYY-MM-DD**: <reason> trail rather than a rewrite.

It edits plan files and nothing else — no repo working tree, no CLAUDE.md, no agent definitions, no memory files.

Both write in whatever language the plan and your own messages are in. The templates ship in English because something has to be the default; they are a shape, not a mandate. A handoff read in a second language is a handoff half-read.

• Executes the work itself. handoff describes the next action; it does not take it.
• Creates a plan file. That is plan mode's job — these two maintain and read one.
• Commits, pushes, tags, or runs a build.
• Fabricates a decision, a file path, or a plan that is not on disk.

Each plugin in this marketplace is versioned and released on its own. Every change to session-continuity — with the commits scoped to this directory — is at session-continuity releases.

MIT — see LICENSE.