[Sycun]

Context compaction now keeps a retrievable continuation summary on the project blackboard and preserves independent pre-compaction evidence snapshots. The host adds the actual snapshot path and history directory to both the continuation context and blackboard body. Existing summary sections 1–7 and raw-user-input collection remain; section 8 adds only tool-evidence keywords and key progress, with no extra model call or standalone index files.

Automatic note/context-* working records are scoped to their source conversation, including its agents, planners and executors. Initial injection, post-compaction refresh and default get/list/search exclude other conversations and unknown origins. Explicit include_history=true enables historical reference within the same authorized project; source/time labels and relationship filtering keep those records from appearing as current work. Ordinary shared project facts retain their behavior, and the system checkpoint prefix is protected from tool overwrites. SQLite checks the current project binding in the same statement as the checkpoint upsert, preventing stale runs from writing after rebinds.

Generated user-role summaries and skills retain their provenance across classic/native conversion and the actual persisted-trace resume boundary. Recovery carries only allowlisted origin fields and the host's structured original-user ledger, rerenders that ledger under current budgets, and drops old system instructions. Assistant/tool text or a generated summary cannot supply human-ledger entries by imitating text tags. Internal metadata stays out of provider request JSON. Chinese ADK regression tests exercise three rounds of compaction, persistence, recovery and recompaction, plus the real database loader, multipart text recovery and the actual native before-model system-message merge. Canonical ledger metadata takes precedence over legacy text during both classic and native merging. Generated PlanExecute templates have separate allowlisted provenance and carry the human ledger from unflattened inputs before executor/replanner rewriting. Eino’s deterministic original-user section is populated from real human entries rather than template plan/tool text; role/source labels remain in working input and snapshots. Context-overflow and transient restarts preserve only the canonical human ledger while discarding old instructions. Formatting-aware budgets retain a bounded first long human input instead of dropping it completely. This cannot reconstruct provenance already discarded by older traces.

Snapshots use transcript-<first 5 MD5 hex characters>-<timestamp>.txt, with nanoseconds and timezone. Full-content comparison handles retry deduplication and short-prefix collisions; prior SHA-256 filenames remain valid. Snapshots are atomically written and retained for the conversation-artifact lifecycle. A failed compatibility transcript.txt update no longer hides a successfully saved immutable snapshot; failed snapshot creation still produces no success hint. The timestamp is first-save time, not a strict successful-compaction counter.

Checkpoint bodies replace the previous summary per conversation/agent scope; full bodies and raw snapshots remain outside automatic injection. Brief labels preferentially reuse section 8's existing evidence clues, with a legacy fallback and a 200-character cap. Index budgets include framing, omission notices and relationship overviews. A successful empty index build removes the stale old block while preserving other system context and the user ledger; database errors retain the old block. New key writes reject surrounding whitespace, while SQL scope filters handle existing Unicode-padded keys without migrating them. Loaded historical records with an unchanged ID/project/exact key remain editable, preserving relationship references. project.context_checkpoints defaults to enabled and can disable automatic blackboard checkpoints independently of snapshots.

Attack-chain analysis now labels generated summaries, skills and plan templates by their persisted provenance instead of presenting them as new human input. The host's canonical original-user ledger remains available as historical scope; old system instructions are omitted, and graph nodes still require actual tool records. A collapsible sidebar displays only metadata-marked summaries from the current conversation's latest saved trace, including exact snapshot references and full-text copy. Cached graphs also expose current context without changing nodes, counts, filters or exports. The notice distinguishes current context from an older cached graph. Summaries render as plain text; this adds no snapshot file-read API or cross-conversation blackboard query. Request ownership and sequence guards prevent stale responses, errors or polling from overwriting another view. Busy regeneration no longer deletes the cache before acquiring its lock, and cached single-node graphs return empty edge arrays.

Validation on the PR's independent source and dependencies: go test ./..., go build ./..., go test -race ./internal/agent ./internal/attackchain ./internal/app ./internal/database ./internal/handler ./internal/multiagent ./internal/project ./internal/projectprompt, node --test web/static/js/*.test.cjs, and git diff --check passed. Desktop and mobile dark-theme sidebar behavior was checked with synthetic context in a local browser fixture. Added Context memory CI directly runs those regression suites on Ubuntu, macOS and Windows. All six Context memory and Self-update checks passed for commit 2b76b313. Tests cover actual resume provenance, forged-ledger rejection, historical scope/pagination/edges/authorization, atomic project binding, snapshot retention/deduplication/collisions/failures, classic/native middleware, legacy padded keys, reduced-budget refresh and evidence-label selection. A production-helper test recovers a synthetic Cookie absent from both the 2001-line tool-output preview and summary via real reduction, snapshot persistence, resume and segmented filesystem reads in classic/native paths. Chat/approval interactions and API routes retain their behavior; the attack-chain sidebar adds the summary/retrieval entry described above. Design and boundaries are documented in docs/zh-CN/blackboard-context-memory.md.