Observations were 100% failing on Claude Code 2.1.109+ because the Agent
SDK emits ["--setting-sources", ""] when settingSources defaults to [].
The existing Bun-workaround filter stripped the empty string but left
the orphan --setting-sources flag, which then consumed --permission-mode
as its value, crashing the subprocess with:
Error processing --setting-sources:
Invalid setting source: --permission-mode.
Make the filter pair-aware: when an empty arg follows a --flag, drop
both so the SDK default (no setting sources) is preserved by omission.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The synthetic summary salvage feature created fake summaries from observation
data when the AI returned <observation> instead of <summary> tags. This was
overengineered — missing a summary is preferable to fabricating one from
observation fields that don't map cleanly to summary semantics.
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: reap stuck generators in reapStaleSessions (fixes#1652)
Sessions whose SDK subprocess hung would stay in the active sessions
map forever because `reapStaleSessions()` unconditionally skipped any
session with a non-null `generatorPromise`. The generator was blocked
on `for await (const msg of queryResult)` inside SDKAgent and could
never unblock itself — the idle-timeout only fires when the generator
is in `waitForMessage()`, and the orphan reaper skips processes whose
session is still in the map.
Add `MAX_GENERATOR_IDLE_MS` (5 min). When `reapStaleSessions()` sees
a session whose `generatorPromise` is set but `lastGeneratorActivity`
has not advanced in over 5 minutes, it now:
1. SIGKILLs the tracked subprocess to unblock the stuck `for await`
2. Calls `session.abortController.abort()` so the generator loop exits
3. Calls `deleteSession()` which waits up to 30 s for the generator to
finish, then cleans up supervisor-tracked children
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix: freeze time in stale-generator test and import constants from production source
- Export MAX_GENERATOR_IDLE_MS, MAX_SESSION_IDLE_MS, StaleGeneratorCandidate,
StaleGeneratorProcess, and detectStaleGenerator from SessionManager.ts so
tests no longer duplicate production constants or detection logic.
- Use setSystemTime() from bun:test to freeze Date.now() in the
"exactly at threshold" test, eliminating the flaky double-Date.now() race.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix: add session lifecycle guards to prevent runaway API spend (#1590)
Three root causes allowed 30+ subprocess accumulation over 36 hours:
1. SIGTERM-killed processes (code 143) triggered crash recovery and
immediately respawned — now detected and treated as intentional
termination (aborts controller so wasAborted=true in .finally).
2. No wall-clock limit: sessions ran for 13+ hours continuously
spending tokens — now refuses new generators after 4 hours and
drains the pending queue to prevent further spawning.
3. Duplicate --resume processes for the same session UUID — now
killed and unregistered before a new spawn is registered.
Generated by Claude Code
Vibe coded by ousamabenyounes
Co-Authored-By: Claude <noreply@anthropic.com>
* fix: use normalized errorMsg in logger.error payload and annotate SIGTERM override
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix: use persisted createdAt for wall-clock guard and bind abortController locally to prevent stale abort
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* chore: re-trigger CodeRabbit review after rate limit reset
* fix: defer process unregistration until exit and align boundary test with strict > (#1693)
- ProcessRegistry: don't unregister PID immediately after SIGTERM — let the
existing 'exit' handler clean up when the process actually exits, preventing
tracking loss for still-live processes.
- Test: align wall-clock boundary test with production's strict `>` operator
(exactly 4h is NOT terminated, only >4h is).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude <noreply@anthropic.com>
* fix: use parent project name for worktree observation writes (#1819)
Observations and sessions from git worktrees were stored under
basename(cwd) instead of the parent repo name because write paths
called getProjectName() (not worktree-aware) instead of
getProjectContext() (worktree-aware). This is the same bug as
#1081, #1317, and #1500 — it regressed because the two functions
coexist and new code reached for the simpler one.
Fix: getProjectContext() now returns parentProjectName as primary
when in a worktree, and all four write-path call sites now use
getProjectContext().primary instead of getProjectName().
Includes regression test that creates a real worktree directory
structure and asserts primary === parentProjectName.
* fix: address review nitpicks — allProjects fallback, JSDoc, write-path test
- ContextBuilder: default projects to context.allProjects for legacy
worktree-labeled record compatibility
- ProjectContext: clarify JSDoc that primary is canonical (parent repo
in worktrees)
- Tests: add write-path regression test mirroring session-init/SessionRoutes
pattern; refactor worktree fixture into beforeAll/afterAll
* refactor(project-name): rename local to cwdProjectName and dedupe allProjects
Addresses final CodeRabbit nitpick: disambiguates the local variable
from the returned `primary` field, and dedupes allProjects via Set
in case parent and cwd resolve to the same name.
---------
Co-authored-by: Ethan Hurst <ethan.hurst@outlook.com.au>
Bun's child_process.spawn() silently drops empty string arguments from
argv, unlike Node which preserves them. When the Agent SDK defaults
settingSources to [] (empty array), [].join(",") produces "" which gets
pushed as ["--setting-sources", ""]. Bun drops the "", causing
--permission-mode to be consumed as the value for --setting-sources:
Error processing --setting-sources: Invalid setting source: --permission-mode
This caused 100% observation failure (exit code 1 on every SDK subprocess
spawn), resulting in 0 observations stored across all sessions.
The fix filters empty string args before passing to spawn(), making the
behavior consistent between Node and Bun runtimes.
Fixes#1779
Related: #1660
Co-authored-by: bswnth48 <69203760+bswnth48@users.noreply.github.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
syncAndBroadcastSummary was using the raw ParsedSummary (null when salvaged)
instead of summaryForStore for the SSE broadcast, causing a crash when the
LLM returns <observation> without <summary> tags. Also removes misplaced
tree-sitter docs from mem-search/SKILL.md (belongs in smart-explore).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Fixes Issue #1312: AI sometimes returns <observation> XML tags instead of
<summary> tags during the summarize phase, despite clear instructions in
buildSummaryPrompt() requiring <summary> ONLY output.
When this occurs, parseSummary() returns null and the entire session summary
is lost. This fix detects the condition (summary missing + observations
present) and synthesizes a summary from the observation data, ensuring
session summaries are not completely lost.
The salvage mapping:
- request: observation title
- investigated: observation narrative or facts
- learned: observation facts joined
- completed: title if type is feature/bugfix
- notes: indicates this is a synthetic salvage summary
Observations are stored normally regardless of this fallback.
Co-authored-by: Sisyphus <sisyphus@openclaw>
GET /api/corpus returned a bare array, which the MCP server wrapper
(callWorkerAPI) forwards directly. MCP's tools/call validation rejects
non-object results with "expected object, received array", so the
list_corpora MCP tool was completely unusable.
Every other corpus endpoint is a POST that already returns the
{content:[...]} shape, so this is a targeted one-file fix.
Stop hook polled queueLength===0 as a proxy for summary success, but the queue
empties regardless of whether the LLM produced valid <summary> tags. Added
lastSummaryStored tracking on ActiveSession, surfaced via the /api/sessions/status
endpoint, and emit a logger.warn in the Stop hook when summaryStored===false.
Generated by Claude Code
Vibe coded by ousamabenyounes
Co-Authored-By: Claude <noreply@anthropic.com>
* feat: add knowledge agent types, store, builder, and renderer
Phase 1 of Knowledge Agents feature. Introduces corpus compilation
pipeline that filters observations from the database into portable
corpus files stored at ~/.claude-mem/corpora/.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: add corpus CRUD HTTP endpoints and wire into worker service
Phase 2 of Knowledge Agents. Adds CorpusRoutes with 5 endpoints
(build, list, get, delete, rebuild) and registers them during
worker background initialization alongside SearchRoutes.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: add KnowledgeAgent with V1 SDK prime/query/reprime
Phase 3 of Knowledge Agents. Uses Agent SDK V1 query() with
resume and disallowedTools for Q&A-only knowledge sessions.
Auto-reprimes on session expiry. Adds prime, query, and reprime
HTTP endpoints to CorpusRoutes.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: add MCP tools and skill for knowledge agents
Phase 4 of Knowledge Agents. Adds build_corpus, list_corpora,
prime_corpus, and query_corpus MCP tools delegating to worker
HTTP endpoints. Includes /knowledge-agent skill with workflow docs.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: handle SDK process exit in KnowledgeAgent, add e2e test
The Agent SDK may throw after yielding all messages when the
Claude process exits with a non-zero code. Now tolerates this
if session_id/answer were already captured. Adds comprehensive
e2e test script (31 assertions) orchestrated via tmux-cli.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: use settings model ID instead of hardcoded model in KnowledgeAgent
Reads CLAUDE_MEM_MODEL from user settings via getModelId(), matching
the existing SDKAgent pattern. No more hardcoded model assumptions.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: improve knowledge agents developer experience
Add public documentation page, rebuild/reprime MCP tools, and actionable
error messages. DX review scored knowledge agents 4/10 — core engineering
works (31/31 e2e) but the feature was invisible. This addresses
discoverability (docs, cross-links), API completeness (missing MCP tools),
and error quality (fix/example fields in error responses).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* docs: add quick start guide to knowledge agents page
Covers the three main use cases upfront: creating an agent, asking a
single question, and starting a fresh conversation with reprime. Includes
keeping-it-current section for rebuild + reprime workflow.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: address code review issues — path traversal, session safety, prompt injection
- Block path traversal in CorpusStore with alphanumeric name validation and resolved path check
- Harden system prompt against instruction injection from untrusted corpus content
- Validate question field as non-empty string in query endpoint
- Only persist session_id after successful prime (not null on failure)
- Persist refreshed session_id after query execution
- Only auto-reprime on session resume errors, not all query failures
- Add fenced code block language tags to SKILL.md
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: address remaining code review issues — e2e robustness, MCP validation, docs
- Harden e2e curl wrappers with connect-timeout, fallback to HTTP 000 on transport failure
- Use curl_post wrapper consistently for all long-running POST calls
- Add runtime name validation to all corpus MCP tool handlers
- Fix docs: soften hallucination guarantee to probabilistic claim
- Fix architecture diagram: add missing rebuild_corpus and reprime_corpus tools
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: enforce string[] type in safeParseJsonArray for corpus data integrity
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: add blank line before fenced code blocks in SKILL.md maintenance section
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Imported observations were invisible to the MCP search tool because the
FTS5 content table was not reliably updated during bulk import. The import
handler now calls rebuildObservationsFTSIndex() after inserting new
observations, ensuring the full-text search index is consistent.
A new SessionStore.rebuildObservationsFTSIndex() method encapsulates the
FTS5 rebuild command and is a no-op when the observations_fts table does
not exist (e.g. FTS5 unavailable on Windows).
- Fix migration version conflict: addSessionPlatformSourceColumn now uses v25
- Sanitize observation titles in file-context deny reason (strip newlines, limit length)
- Guard json_each() with LIKE '[%' check for legacy bare-path rows
- Guard /stream SSE endpoint with 503 before DB initialization
- Scope bun-runner signal exit handling to start subcommand only
- Normalize platformSource at route boundary in DataRoutes
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Remove duplicate TranscriptWatcher/config imports in worker-service.ts
- Use normalizePlatformSource in handleSessionInitByClaudeId for consistency
- Don't skip DB completion when session not in memory (completeByClaudeId)
- Add try-catch around fetch in useContextPreview refresh callback
- Deduplicate store.getAllProjects() call in DataRoutes
- Fix malformed comment separators in migration runner
- Fix missing closing brace and JSDoc opener (merge artifact) in migration runner
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Two bugs fixed:
1. SessionCompletionHandler called dbManager.getSessionStore() during
WorkerService construction, before DB initialization. Changed to
accept DatabaseManager and defer the call to runtime.
2. migration009 (generated_by_model, relevance_count columns) only ran
via the deprecated MigrationRunner path, never through SessionStore's
migration chain. Added addObservationModelColumns() to SessionStore
constructor. Checks column existence directly since schema_versions
may have been marked applied without the ALTER TABLE succeeding.
Also removed duplicate transcriptWatcher declaration and shutdown block
(merge artifact).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Number(x) || 10 converts 0 to 10 since 0 is falsy, making it impossible
to request zero context depth (anchor only). Replace with explicit null
check in timeline(), getContextTimeline(), getTimelineByQuery().
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- tests/servers/mcp-tool-schemas.test.ts: remove `import '../../src/servers/mcp-server.js'`
which triggered server startup side effects; test only needs to read the TS source as text
- src/services/worker/SearchManager.ts: add Number() coercion for depth_before/depth_after
in timeline(), getContextTimeline(), getTimelineByQuery() — HTTP query strings deliver
these as strings, coercion ensures they are always numbers before being passed to
filterByDepth() and getTimelineAround*()
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The per-session FileReadGate was never requested and broke the cost
savings loop — subsequent reads in the same session silently bypassed
the timeline, hiding newly created observations.
Now the timeline fires on every read that has observations, using the
hook contract's permissionDecision: "deny" with the timeline as the
reason (exit 0 + JSON) instead of exit code 2 + stderr.
- Delete FileReadGate.ts entirely
- Remove /api/file-context/gate endpoint from DataRoutes
- Switch handler from exit code 2 to permissionDecision: "deny"
- Restore permissionDecision fields to HookResult
- Eliminate one HTTP round-trip per read (no gate check needed)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Resolve relative filePath against input.cwd before statSync; early-return on ENOENT
- Replace LIKE '%path%' with exact json_each equality to prevent false matches
- Sanitize and parameterize LIMIT to prevent NaN SQL errors
- Fix day-sorting to use earliest epoch in group, not first (specificity-sorted) item
- Use exact path equality in deduplicateObservations instead of substring includes
- Scope FileReadGate by session+cwd to prevent worktree collisions
- Refresh lastAccess TTL on active sessions; throttle prune to every 50 calls
- Type params as (string | number)[] instead of any[]
- Remove unused permissionDecision fields from HookResult
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Skip gate for files under 1,500 bytes — timeline (~370 tokens) costs
more than just reading small files directly
- Deduplicate observations by memory_session_id (one per session)
- Rank by specificity: files_modified > files_read, fewer tagged files > many
- Fetch 40 candidates, dedup/score down to 15 for display
- Reduce default by-file query limit from 30 to 15
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Observer prompt now explicitly requires XML observation blocks or empty
responses — prose explanations like "Skipping" are discarded. ResponseProcessor
logs a warning when non-XML content is received. Recording focus expanded to
include concrete debugging findings.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The generated_by_model column was added to the observations table in the
Phase 0 governance schema migration but never wired into the INSERT
statements. All 3,878+ observations in production have this field NULL.
This fix threads the model ID from each agent (SDKAgent, GeminiAgent,
OpenRouterAgent) through processAgentResponse() into storeObservation(),
storeObservations(), and storeObservationsAndMarkComplete().
Unblocks Thompson Sampling RFC (#1571) which needs {obs_type}:{model}
as the bandit arm key.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Critical:
- migrations: change version 8 → 25 to avoid collision with
MigrationRunner.addObservationHierarchicalFields (uses version 8)
- SessionRoutes: remove duplicate imports that prevent compilation
Major:
- SessionRoutes: call applyTierRouting() before every generator spawn
(stale-recovery and crash-recovery paths were missing it)
- applyTierRouting: clear session.modelOverride at top before re-evaluating
to prevent stale tier from persisting across spawns
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Tier Routing:
- Inspect pending queue before starting generator
- Summarize messages → CLAUDE_MEM_TIER_SUMMARY_MODEL (e.g., Opus)
- All simple tools (Read, Glob, Grep, LS) → CLAUDE_MEM_TIER_SIMPLE_MODEL (Haiku)
- Mixed/complex → default model (no override)
- session.modelOverride in ActiveSession, used by SDKAgent.getModelId()
- peekPendingTypes() in PendingMessageStore for non-claiming inspection
- Configurable via CLAUDE_MEM_TIER_ROUTING_ENABLED (default: true)
Feedback Collection (schema only):
- New observation_feedback table via MigrationRunner (schema version 24)
- Tracks signal_type (semantic_inject_hit, search_accessed, etc.)
- Indexes on observation_id and signal_type
- Foundation for future Thompson Sampling optimization
Production data (24h tier routing test):
- 36 Haiku observations in 4 min, quality indistinguishable from Sonnet
- Estimated ~52% cost reduction on SDK Agent usage
- 835 → 6,695 feedback signals collected over 13 days
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Follow-up to PR #1568: fix stale doc comment that still said GET, and add
limit parameter validation (default 5, clamped to 1-20 range).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: semantic context injection via Chroma on every UserPromptSubmit
On each prompt, queries ChromaDB for the top-N most relevant past
observations and injects them as additionalContext. Replaces the
recency-based "last N observations" approach with relevance-based
semantic search.
Changes:
- session-init.ts: After session init, query /api/context/semantic
with user's prompt text. If results found, return as
hookSpecificOutput with hookEventName 'UserPromptSubmit'.
- SearchRoutes.ts: New GET /api/context/semantic endpoint that queries
SearchManager with format='json' and formats results as markdown.
- SettingsDefaultsManager.ts: New settings CLAUDE_MEM_SEMANTIC_INJECT
(default: true) and CLAUDE_MEM_SEMANTIC_INJECT_LIMIT (default: 5).
Key behaviors:
- Fires on every UserPromptSubmit (not just SessionStart)
- Minimum prompt length: 20 chars (skips "ok", "yes", etc.)
- Skips media-only prompts
- Graceful degradation: if worker/Chroma unavailable, no injection
- Survives /clear: re-injects on next prompt (not session-bound)
- Uses workerHttpRequest (v10.6.3 API, not raw fetch)
Production data (23 days, 3,400+ observations):
- Before: 8 most recent observations (often irrelevant to current topic)
- After: 5 most relevant observations (semantic match)
- Token cost: ~1800 → ~800-1200 per injection
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: address CodeRabbit review on PR #1568
- session-init: don't skip semantic injection when contextInjected=true
(only skip agent re-init, semantic lookup must run every prompt)
- session-init: normalize SEMANTIC_INJECT toggle via String().toLowerCase()
- semantic endpoint: change from GET to POST to avoid URL-length limits
and prompt exposure in access logs. Handler accepts both body and query
for backwards compatibility.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Alessandro Costa <alessandro@claudio.dev>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: drain orphaned pending messages on session completion (SIGTERM)
When deleteSession() aborts the SDK agent via SIGTERM, pending messages
in the queue are never processed. Without drain, they remain in
'pending' status forever — no future generator picks them up because
the session is already completed.
Adds markAllSessionMessagesAbandoned() call after deleteSession() in
completeByDbId(). This reuses the existing PendingMessageStore method
already used by worker-service.ts terminateSession().
Production evidence: 15 orphaned summarize messages found across
completed sessions (ages 3h to 3 days) before this fix. After fix:
0 orphaned messages over 23 days of operation.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: document best-effort drain limitation per CodeRabbit review #1567
Add comment noting the rare race condition when generators outlive the
30s SIGTERM timeout. Practical risk is negligible (0 orphans over 23 days).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Alessandro Costa <alessandro@claudio.dev>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
SessionEnd has a 1.5s hardcoded cap from Claude Code (CLAUDE_CODE_SESSIONEND_HOOKS_TIMEOUT_MS),
making it unsuitable for waiting on async work. Previously, the Stop hook would fire-and-forget
the summarize request, then SessionEnd would immediately call deleteSession — aborting the SDK
agent mid-summary.
Now the Stop hook (120s timeout, no cap) owns the full lifecycle:
1. Queue summarize request
2. Poll new GET /api/sessions/status endpoint until queue drains
3. Call /api/sessions/complete after summary finishes
SessionEnd is now a true fire-and-forget fallback (process.exit(0) immediately).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
completeByDbId only cleaned up in-memory state, leaving sdk_sessions rows
with status='active' and completed_at=NULL indefinitely. Ghost sessions
accumulated and exhausted the agent pool, causing 60s timeout errors.
- Add SessionStore.markSessionCompleted() to set status/completed_at/completed_at_epoch
- Call it at the start of completeByDbId before in-memory cleanup
- Inject SessionStore into SessionCompletionHandler via constructor
- Add 4 tests covering status, timestamps, isolation, and non-existent IDs
Closes#1532
Co-Authored-By: Claude <noreply@anthropic.com>
Tighten platform source persistence so legacy callers cannot silently relabel existing sessions, repair migration 24 when schema_versions drifts from the real schema, and polish the follow-up UI/error-handler review nits.
- only backfill platform_source when it is blank and raise on explicit source conflicts for an existing session
- make migration 24 verify both the sdk_sessions column and its index before treating it as applied
- expose platform_source from the functional session getters and add regression tests for source preservation and schema drift recovery
- add the required APPROVED OVERRIDE annotation for centralized HTTP error translation
- keep mobile source pills on a single horizontal row
Persist platform_source across session creation, transcript ingestion, API query paths, and viewer state so Claude and Codex data can coexist without bleeding into each other.
- add platform-source normalization helpers and persist platform_source in sdk_sessions via migration 24 with backfill and indexing
- thread platformSource through CLI hooks, transcript processing, context generation, pagination, search routes, SSE payloads, and session management
- expose source-aware project catalogs, viewer tabs, context preview selectors, and source badges for observations, prompts, and summaries
- start the transcript watcher from the worker for transcript-based clients and preserve platform source during Codex ingestion
- auto-start the worker from the MCP server for MCP-only clients and tighten stdio-driven cleanup during shutdown
- keep createSDKSession backward compatible with existing custom-title callers while allowing explicit platform source forwarding
GeminiAgent sends the full conversation history with every API call,
causing quadratic token growth per session. A 100-observation session
sends ~30M cumulative input tokens. This ports the proven truncateHistory()
sliding window from OpenRouterAgent to GeminiAgent.
- Add CLAUDE_MEM_GEMINI_MAX_CONTEXT_MESSAGES (default: 20) and
CLAUDE_MEM_GEMINI_MAX_TOKENS (default: 100000) settings
- Add truncateHistory() to GeminiAgent using shared estimateTokens()
- Always preserve at least the newest message to avoid empty API requests
- Add settings validation in SettingsRoutes (1-100 messages, 1K-1M tokens)
- Add regression tests for truncation and oversized single-prompt edge case
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: stop spinner from spinning forever due to orphaned DB messages
The activity spinner never stopped because isAnySessionProcessing() queried
ALL pending/processing messages in the database, including orphaned messages
from dead sessions that no generator would ever process.
Root cause: isAnySessionProcessing() used hasAnyPendingWork() which is a
global DB scan. Changed it to use getTotalQueueDepth() which only checks
sessions in the active in-memory Map.
Additional fixes:
- Add terminateSession() to enforce restart-or-terminate invariant
- Fix 3 zombie paths in .finally() handler that left sessions alive
- Clean up idle sessions from memory on successful completion
- Remove redundant bare isProcessing:true broadcast
- Replace inline require() with proper accessor
- Add 8 regression tests for session termination invariant
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: address review findings — idle-timeout race, double broadcast, query amplification
- Move pendingCount check before idle-timeout termination to prevent
abandoning fresh messages that arrive between idle abort and .finally()
- Move broadcastProcessingStatus() inside restart branch only — the else
branch already broadcasts via removeSessionImmediate callback
- Compute queueDepth once in broadcastProcessingStatus() and derive
isProcessing from it, eliminating redundant double iteration
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
ColorFormatter and MarkdownFormatter names obscured their actual purpose.
The formatters serve two distinct audiences: the AI agent (compressed,
token-efficient context) and the human (rich ANSI-colored terminal output).
- MarkdownFormatter → AgentFormatter (renderMarkdown* → renderAgent*)
- ColorFormatter → HumanFormatter (renderColor* → renderHuman*)
- useColors parameter → forHuman across the pipeline
- Import aliases Color/Markdown → Human/Agent
- API query param `colors=true` unchanged (backward compatible)
Pure rename refactor — no logic or behavior changes.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Add a PreToolUse gate that blocks file reads on first attempt when rich
observation history exists, presenting the timeline as feedback. Claude
then decides: use get_observations() (skip read, save tokens) or re-read
(allowed on second attempt).
- FileReadGate: in-memory session-scoped gate with 4h TTL
- POST /api/file-context/gate endpoint in worker
- stderrMessage plumbing in hook-command for exit code 2
- file-context handler uses gate to block/allow reads
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Use getProjectContext(cwd).allProjects for project scoping (same as SessionStart)
- Convert absolute file_path to relative using cwd (observations store relative paths)
- API accepts comma-separated projects param with IN() SQL filter
- Remove basename matching — use full relative path to avoid cross-file collisions
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When Claude reads a file, the PreToolUse hook queries for existing
observations about that file and injects the timeline into context
via additionalContext + permissionDecision: allow. This prevents
duplicate observations and saves tokens through active rediscovery.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Alex Newman