Overview cache had no self-invalidation: once written it would return
stale lastMessageTs forever if the postImportHook failed silently.
Analysis query cache (17 chart/stats queries) was never cleared on
import, so charts showed stale data until the user triggered a manual
invalidate.
Add MAX(message.id) fingerprint to OverviewCache and introduce
getValidatedOverviewCache(), which recomputes automatically when new
messages are inserted. Update resolveOverview() in the Electron worker
to use this instead of bare getCache().
In both postImportHooks (incrementalImport, streamImport) also call
deleteSessionCache against the query cache directory so chart data is
fresh after every import.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Include assistant messages with tool content blocks in getHistoryForAgent
even when text content is empty, so multi-turn tool results are not lost
- Append Date.now() timestamp to export filenames to prevent silent overwrites
- Pass 600s timeout to exportFilterResultToFile worker call to match the
actual upper bound of large-session exports
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Extend TokenUsage types across the full data pipeline to include
cacheReadTokens and cacheWriteTokens. The ChatStatusBar tooltip now
shows cache hit information when available. This is a generic
implementation that works with any API provider reporting cache metrics,
not limited to DeepSeek.
Co-authored-by: Cursor <cursoragent@cursor.com>
DeepSeek API requires reasoning_content to be retransmitted on assistant
messages that precede tool results. Add ReplayOptions to history replay
so thinking blocks are included as ThinkingContent when the model uses
DeepSeek-format thinking, preventing 400 errors in multi-turn tool-call
conversations.
Co-authored-by: Cursor <cursoragent@cursor.com>
Message-retrieval tools mirror rawMessages inside data, and the
preprocessing pipeline rendered that array into the tool result text as
"rawMessages: [object Object], [object Object], ..." — roughly 1.7KB of
junk per call (~400-500 wasted tokens) that also polluted the truncated
result text persisted for history replay.
- formatToolResultAsText skips the rawMessages key (raw message objects
are already rendered via the formatted messages list)
- CLI tool adapter strips rawMessages from extraDetails before invoking
the pipeline, matching the existing desktop adapter behavior
Regression tests fail without the fix at both the format layer and the
CLI adapter layer.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Multi-turn agent conversations rebuilt history as plain text, dropping
all tool calls and results. Later turns could not see what the model
had actually retrieved, so it imitated tool-free answers and
hallucinated chat records instead of re-querying (worse after
incremental imports and in long conversations).
- Persist provider-issued toolCallId, truncated result text and isError
on tool content blocks (written by the shared frontend store from
stream chunks; pi-agent-core events already carry these fields)
- Replay persisted tool blocks as real toolCall/toolResult message
pairs in toPiHistoryMessages: stable ids keep requests byte-identical
across turns (prompt cache friendly), unfinished blocks are skipped
so no toolCall is left unanswered, errors map to isError, results are
capped at 4000 chars, runtime-injected underscore params are stripped
- Legacy messages without persisted tool data fall back to the old
plain-text replay; no migration needed
Verified on a real CLI web session: 25+ replayed tool pairs accepted by
the provider across turns with zero API errors.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Add optional `rows` parameter to render_chart so AI can pass pre-fetched
data arrays (e.g. from get_time_stats, member_stats) directly instead of
rewriting SQL from scratch. Update chart capability prompts and all locale
files to guide AI to check high-level tools first before falling back to SQL.
Remove debug console.debug log from agent-core thinkingLevel resolution.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
thinkingLevels was excluded from the mergeDefaults whitelist in
PreferencesManager, so user choices were silently dropped on every
save/load cycle and never written to preferences.json.
Add the field to the Preferences interface, DEFAULTS, and mergeDefaults
so the existing backendPersist → savePreferences pipeline can write and
restore it correctly. Adds regression tests covering the save/load
round-trip, multi-save merge, and legacy-file backward compatibility.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Harden the startup migration path: fail fast and show an error dialog if
schema migration cannot complete. Repair the gate retroactively for
databases that were already migrated before the gate system existed.
Raise the gate after every import path (full import, incremental import,
stream import) and after opening any database that already carries a
current-segment schema. Centralise gate-raising inside DatabaseManager so
call sites cannot accidentally skip it.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Use the packaged app version (not Electron's dev-mode 0.0.0) across all
compatibility checks. Make DatabaseManager require a RuntimeIdentity so
every gate write carries a verified version and kind.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
After incremental import via Electron IPC, the sidebar session list
showed stale message counts. Three contributing gaps:
1. The chat:incrementalImport IPC handler did not notify the renderer
to refresh the session list (unlike the API route which calls
notifySessionListChanged). Added win.webContents.send('api:importCompleted')
to trigger the Sidebar's existing onImportCompleted listener.
2. The postImportHook in worker adapters was async (due to dynamic
import()) but the caller did not await it, causing a race condition
where loadSessions() could read stale overview cache before
computeAndSetOverviewCache() finished. Converted to static import
making the hook synchronous, and updated the type to void | Promise<void>
with await at call sites in both incremental-importer and
streaming-importer.
3. The @imported handlers on group-chat and private-chat pages only
refreshed local analysis data without calling sessionStore.loadSessions().
Added the call so the sidebar updates immediately on the client side.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Enable esbuild minification for main/preload (was renderer-only by default),
reducing main process bundle by ~55%. Extract cl100k_base BPE rank table out
of the synchronous startup path via dynamic import — it now lands in its own
chunk and is loaded on first agent run instead of at app startup.
Also fix Windows CI step to use PowerShell syntax for version injection.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Unify the merge workflow (parse/conflicts/execute/clear/export-for-merge)
across CLI, Electron, and web-serve into a single set of shared routes in
@openchatlab/http-routes.
Key changes:
- Move MergeSessionCache from CLI to node-runtime shared package
- Add dual-mode parse route (multipart upload for web, JSON filePath for Electron)
- Integrate merge routes into Electron internal API server with @fastify/multipart
- Create installMergeShims() with filePath-to-handle mapping for frontend
- Remove Electron merge IPC handlers, merger module, and preload bridges (~790 lines deleted)
- Address CodeX review: CHAT_DB_TABLES in openDatabase, no chatApi patching,
outputName sanitization, resp.ok checks, @fastify/multipart in desktop deps
Co-authored-by: Cursor <cursoragent@cursor.com>
When the user selects 'auto' thinking level for an effort-based reasoning
model (Kimi, Doubao) that has thinkingLevelMap but no thinkingFormat,
the previous code fell through to `return undefined`, causing pi-agent-core
to treat it as 'off' and send reasoning_effort:'none' to the API —
silently disabling thinking.
pi-ai's clampThinkingLevel only covers EXTENDED_THINKING_LEVELS (no 'auto'),
so 'auto' cannot be forwarded verbatim. Return 'high' instead, matching
the behavior for thinkingFormat models, to at least enable reasoning.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
diguaandClaude Opus 4.8
Yui
CEQ151