continuity.md

Inspect the context behind a handoff

Use the existing lifecycle state and local transaction evidence to understand what an agent should know and what was saved. Both views are offline and read-only. They never infer permission from file content or synchronize native memory. Their output may contain private context: review its audience before sharing it.

Briefing with sources

bash scripts/contextos.sh start --format markdown
bash scripts/contextos.sh start --briefing
bash scripts/contextos.sh start --source projects/lantern/requirements.md --format markdown

The default start JSON remains the metadata inventory. --briefing, Markdown output, or explicit --source selection adds bounded excerpts, source paths, selection reasons, normalized-text hashes, and freshness. Repeated --source options select additional repository-relative Markdown files; they never cause automatic routing or link expansion. The fixed sources are routing, configured state, recent decision rows, and the latest session. Missing or unreadable sources have an explicit status. Oversized sources are unavailable; unsafe explicit paths are rejected.

Each excerpt is limited to 24 lines and 2,400 characters; decisions show the last five table rows. Read the full source when the excerpt is insufficient. Current state uses the existing 3/5/7-day policies for current priorities, weekly priorities, and blockers. Other sources use a 90-day advisory threshold when they contain a valid **Last Updated:** date. Missing dates remain unknown; the report does not substitute filesystem modification time for confirmation.

This is a source preview, not an agent read log. A recent date does not prove a claim is correct, and an old decision can still be valid. Cite actual source paths in the agent's briefing and distinguish unresolved assumptions from confirmed decisions.

Readable change history

bash scripts/contextos.sh history
bash scripts/contextos.sh history --details --path state/decisions.md
bash scripts/contextos.sh history --format json --limit 20

History lists recent local .context-os/receipts/*.json records by application time, with changed paths, before/after hashes, and self-reported runtime. With --details, it checks the matching proposal's digest and change list against the receipt before showing its recorded diff. Missing or altered proposals do not supply explanations; their status is shown. Malformed receipts produce warnings while valid records remain visible. Reports read at most 1,000 receipt files of at most 1 MiB each and return 1–100 matching entries.

Receipts and proposals are ignored local artifacts: they may be absent in another clone, manually edited, or removed. Matching digests show consistency, not authenticity. They do not authenticate a human, prove no other writes occurred, verify a fact, or recover the model's reasoning. The diff can explain a recorded rationale only when that rationale was actually saved. Use Git history for committed changes and the canonical decision file for current meaning. Do not move raw receipts into tracked session logs.

For attached application repositories, supply the same explicit --kernel-root, --context-root, and --working-root arguments used by lifecycle commands. The existing project binding must validate. Reports read ContextRoot sources and local receipts; they do not load application content as shared memory.

Try the first handoff, then use the benchmark to check constrained continuity behavior.