Pipeline configuration

Pipeline V2 Config

The V2 memory pipeline lives at platform/daemon/src/pipeline/. It runs LLM-based fact extraction against incoming conversation text, then decides whether to write new memories, update existing ones, or skip. Config lives under memory.pipelineV2 in agent.yaml.

Inference selection for extraction can be routed through the top-level inference.workloads bindings. Session processing follows the memoryExtraction route and is not independently configurable. When explicit routing is enabled for default, memoryExtraction, widgetGeneration, or repair, those workloads use the shared inference control plane. Legacy extraction and synthesis fields are treated as load-time compatibility input, not separate runtime providers.

The config uses a nested structure with grouped sub-objects. Legacy flat keys (e.g. extractionModel, leaseTimeoutMs) are still supported for backward compatibility, but nested keys take precedence when both are present.

Enable the pipeline:

memory:
  pipelineV2:
    enabled: true
    shadowMode: true        # extract without writing — safe first step
    extraction:
      provider: llama-cpp
      model: qwen3:4b

Control flags

These top-level boolean fields gate major pipeline behaviors.

Field	Default	Description
enabled	true	Master switch. Pipeline does nothing when false.
shadowMode	false	Extract facts but skip writes. Useful for evaluation.
mutationsFrozen	false	Allow reads; block all writes. Overrides shadowMode.
semanticContradictionEnabled	true	Enable LLM-based semantic contradiction detection for UPDATE/DELETE proposals.
telemetryEnabled	true	Enable anonymous telemetry reporting (set false to opt out).

The relationship between shadowMode and mutationsFrozen matters: shadowMode suppresses writes from the normal extraction path only; mutationsFrozen is a harder freeze that blocks all write paths including repairs and graph updates.

Extraction (extraction)

Controls the LLM-based extraction stage. Supports multiple providers.

Field	Default	Range	Description
provider	"llama-cpp"	—	"none", "acpx", "llama-cpp", "ollama", "claude-code", "opencode", "codex", "anthropic", "openrouter", "openai-compatible", or "command"
fallbackProvider	"llama-cpp"	—	"llama-cpp", "ollama", or "none"; legacy extraction configs compile this into an inference fallback target
model	"qwen3:4b"	—	Model name for the configured provider
timeout	90000	5000-300000 ms	Extraction call timeout
minConfidence	0.7	0.0-1.0	Confidence threshold; facts below this are dropped
structuredOutput	true	—	Send JSON schema in the format field of LLM requests. Set false when the provider rejects structured output (e.g. GitHub Copilot API). The daemon also auto-detects unsupported providers at runtime and disables this transparently.
command	—	—	Command provider config (bin, args[], optional cwd, optional env) — required when provider: "command"
rateLimit.maxCallsPerHour	200 when rateLimit is set	0-10000	Max extraction-provider calls per hour; set 0 to disable rate limiting
rateLimit.burstSize	20 when rateLimit is set	1-1000	Max burst size before throttling begins
rateLimit.waitTimeoutMs	5000 when rateLimit is set	0-60000 ms	How long to wait for a token before failing with RateLimitExceededError

For provider: openai-compatible, set endpoint to the gateway’s OpenAI-compatible /v1 base URL. Remote endpoints use OPENAI_API_KEY by default when this legacy pipeline config is compiled into inference routing; explicit top-level inference.accounts.*.credentialRef can point at any stored secret or environment variable.

For safety, the intended extraction setups are:

• local llama-cpp with qwen3:4b (default)
• claude-code on a Haiku model
• codex on a gpt-5.4-mini model
• local ollama with nemotron-3-nano:4b (preferred) or qwen3:4b (deprecated — Nemotron’s superior reasoning makes Qwen3 the weaker choice going forward; expect degraded extraction quality in future updates)

Set provider: none to disable extraction entirely, which is the recommended default for VPS installs that should not make background LLM calls.

Remote API extraction can accumulate extreme fees quickly because the pipeline runs continuously in the background. Use anthropic, openrouter, openai-compatible, or remote OpenCode routes only when you explicitly want that billing behavior.

rateLimit is opt-in. If the stanza is omitted, Signet preserves the provider’s existing behavior with no throughput throttling. When configured, it applies only to remote or paid providers (acpx, claude-code, anthropic, openrouter, openai-compatible, codex, opencode). Ollama and command providers are always exempt. If you set rateLimit on an exempt provider, Signet logs a warning and passes calls through unthrottled.

An empty rateLimit: {} block is treated as disabled. Set at least one sub-field to opt in, or omit the stanza entirely to leave rate limiting off.

When a rate-limited job fails (the bucket is empty and the wait timeout expires), it is classified as non-retryable and sent directly to dead-letter status. Dead-lettered jobs are not retried when the rate-limit window resets. Choose maxCallsPerHour high enough to handle sustained ingestion bursts, or you will permanently lose extraction for memories queued during exhaustion. Dead-letter jobs are purged after 30 days by the retention worker.

When configured via YAML, burstSize is clamped to a minimum of 1. The lower-level withRateLimit() helper is more defensive: passing burstSize: 0 or maxCallsPerHour: 0 disables the wrapper entirely instead of constructing a limiter that can never acquire a token.

Rate-limiter state is in-memory only. After a daemon restart the full burstSize is available immediately (the token bucket starts full). In environments with frequent restarts (crash-loops, rolling deployments), this means the limiter cannot protect against a burst of calls right after startup. Set burstSize conservatively if your daemon restarts often under load.

When using ollama, the model must be available locally. When using claude-code, the Claude Code CLI must be on PATH. codex uses the Codex CLI as the extraction provider. Lower minConfidence to capture more facts at the cost of noise; raise it to write only high-confidence facts.

acpx is available as a setup compatibility value for installations that also have a top-level inference: block. ACPX needs harness/session config, so legacy memory.pipelineV2.extraction.provider: acpx by itself is not compiled into an implicit legacy-extraction target; keep the generated inference.targets.*.executor: acpx block or configure ACPX through top-level inference routing.

There are two command paths with different contracts. Top-level inference.targets.*.executor: command is a normal inference provider: the prompt is sent on stdin, exposed as SIGNET_PROMPT, and the model response is read from stdout.

memory.pipelineV2.extraction.provider: command is retired. It is rejected at config load time; use the canonical inference.workloads.memoryExtraction target consumed by Dreaming instead. The retired path let a subprocess write memory state outside the daemon-owned audited apply path.

Session synthesis (synthesis)

Controls the provider used by the summary-worker for session summaries. This is separate from fact extraction once explicitly configured.

If the synthesis block is omitted entirely, Signet falls back to the resolved extraction provider, model, endpoint, and timeout. When an explicit top-level inference: block exists, workload bindings decide which target handles synthesis.

Field	Default	Range	Description
enabled	true	—	Enable background session summary generation
provider	inherited from extraction when omitted	—	"none", "llama-cpp", "ollama", "claude-code", "codex", "opencode", "anthropic", "openrouter", or "openai-compatible"
model	inherited from extraction when omitted	—	Model name for the configured provider
endpoint	inherited from extraction when omitted	—	Optional base URL override for Ollama, OpenCode, OpenRouter, or OpenAI-compatible gateways
timeout	inherited from extraction when omitted	5000-300000 ms	Summary generation timeout
structuredOutput	inherited from extraction when omitted	—	Send JSON schema in the format field of LLM requests. Set false when the synthesis provider rejects structured output (e.g. GitHub Copilot API). Falls back to extraction.structuredOutput when omitted.
rateLimit.maxCallsPerHour	200 when rateLimit is set	0-10000	Max synthesis-provider calls per hour; set 0 to disable rate limiting
rateLimit.burstSize	20 when rateLimit is set	1-1000	Max burst size before throttling begins
rateLimit.waitTimeoutMs	5000 when rateLimit is set	0-60000 ms	How long to wait for a token before failing with RateLimitExceededError

Set provider: none or enabled: false to disable background session summary synthesis entirely.

synthesis.provider: command is invalid and rejected during config load.

Widget HTML generation uses a separate provider instance by default, so widget traffic does not consume the synthesis pipeline’s rateLimit bucket.

As with extraction, an empty rateLimit: {} block is treated as disabled. Set at least one sub-field to opt in.

Rate-limited synthesis jobs that fail are sent to dead-letter without retry. See the extraction rateLimit docs above for the full warning.

Claude Code background environment (claudeCode)

Applies whenever legacy pipeline extraction, synthesis, or an explicit inference route uses the claude-code provider.

Field	Default	Range	Description
allowApiKeyEnv	false	—	When false, daemon-spawned claude -p calls strip ambient ANTHROPIC_API_KEY and ANTHROPIC_AUTH_TOKEN. Set true only when background pipeline jobs should inherit those env credentials. Legacy unshipped billingMode: api-key is accepted as an alias for true; billingMode: subscription maps to false.
maxBudgetUsd	unset	0.01-1000	Optional per-invocation spend cap passed to Claude Code print mode as --max-budget-usd. Omitted by default because Claude Code documents no default limit for this flag and forcing one could change CLI behavior.
cooldownMs	300000	1000-3600000 ms	Daemon-wide Claude Code circuit cooldown opened after Claude Code reports quota, usage-limit, credit, billing, or auth failures. Calls during cooldown fail before spawning claude. Interactive and background claude-code providers in the daemon share this circuit and config snapshot.

memory:
  pipelineV2:
    claudeCode:
      allowApiKeyEnv: false
      cooldownMs: 300000

Only opt into ambient API-key/token inheritance when you intentionally want background pipeline jobs to use the Anthropic credentials already present in the daemon environment:

memory:
  pipelineV2:
    claudeCode:
      allowApiKeyEnv: true
      maxBudgetUsd: 0.25

Anthropic’s Claude Code CLI reference lists --max-budget-usd as a print-mode-only API-call budget flag: https://docs.anthropic.com/en/docs/claude-code/cli-reference. Anthropic’s Claude Code cost docs describe Claude Code charges in terms of API token consumption and subscription plan pricing separately: https://docs.anthropic.com/en/docs/claude-code/costs. Anthropic support docs also state that paid Claude subscriptions and Claude Console/API usage are separate products: https://support.anthropic.com/en/articles/9876003-i-subscribe-to-claude-pro-why-do-i-have-to-pay-separately-for-api-usage-on-console. Signet does not verify the billing account selected by a persisted claude auth login --console session; it only controls whether the daemon subprocess inherits ambient Anthropic API key/token environment variables.

Worker (worker)

The pipeline processes jobs through a queue with lease-based concurrency control.

Field	Default	Range	Description
maxRetries	3	1-10	Max retry attempts before a job goes to dead-letter
leaseTimeoutMs	300000	10000-600000 ms	Time before an uncompleted job lease expires
maxLlmConcurrency	2	1-16	Shared cap for live LLM calls across extraction, synthesis, reranking, inference streaming, and daemon route provider calls such as skills, ontology consolidation, and diagnostics greetings. SIGNET_MAX_LLM_CONCURRENCY overrides YAML when set, matching the TypeScript daemon behavior for wired provider paths.

A job that exceeds maxRetries moves to dead-letter status and is eventually purged by the retention worker. The standalone extraction worker was retired under the Dreaming cutover (#946); its former pollMs, maxLoadPerCpu, overloadBackoffMs, and threadedExtraction knobs are no longer read from configuration (legacy YAML values are ignored).

Knowledge Graph (graph)

When graph.enabled: true, the pipeline builds entity-relationship links from extracted facts and uses them to boost search relevance.

Field	Default	Range	Description
enabled	true	—	Enable knowledge graph building and querying
boostWeight	0.15	0.0-1.0	Weight applied to graph-neighbor score boost
boostTimeoutMs	500	50-5000 ms	Timeout for graph lookup during search

Hints (hints)

Prospective indexing generates hypothetical future queries at write time. These “hints” are indexed in FTS5 so memories match by anticipated cue, not just stored content. For example, a memory about “switched from PostgreSQL to SQLite” might generate hints like “database migration”, “why SQLite”, and “storage engine decision” — queries the user is likely to ask later.

Field	Default	Range	Description
enabled	true	—	Enable prospective indexing
max	5	1-20	Maximum hints generated per memory
timeout	30000	5000-120000 ms	Hint generation LLM timeout
maxTokens	256	32-1024	Max tokens for hint generation
poll	5000	1000-60000 ms	Job polling interval

memory:
  pipelineV2:
    hints:
      enabled: true
      max: 5
      timeout: 30000
      maxTokens: 256
      poll: 5000

Traversal (traversal)

Graph traversal controls how the knowledge graph is walked during retrieval. When primary: true, graph traversal produces the base candidate pool and flat search fills gaps. When primary: false, traditional hybrid search runs first with graph boost as supplementary.

Field	Default	Range	Description
enabled	true	—	Enable graph traversal
primary	true	—	Use traversal as primary retrieval strategy
maxAspectsPerEntity	20	1-50	Max aspects to collect per entity (read cap)
maxAttributesPerAspect	50	1-100	Max attributes per aspect (read cap)
maxWriteAspectsPerEntity	20	1-50	Max active aspects per entity before create_aspect is rejected
maxWriteAttributesPerAspect	50	1-100	Max active attributes per aspect before add_claim_value is rejected
maxDependencyHops	10	1-50	Max hops for dependency walking
minDependencyStrength	0.3	0.0-1.0	Minimum edge strength to follow
maxBranching	4	1-20	Max branching factor during traversal
maxTraversalPaths	50	1-500	Max paths to explore
minConfidence	0.5	0.0-1.0	Minimum confidence for results
timeoutMs	500	50-5000 ms	Traversal timeout
boostWeight	0.2	0.0-1.0	Weight for traversal boost in hybrid search
constraintBudgetChars	1000	100-10000	Character budget for constraint injection

memory:
  pipelineV2:
    traversal:
      enabled: true
      primary: true
      maxAspectsPerEntity: 20
      maxAttributesPerAspect: 50
      maxWriteAspectsPerEntity: 20
      maxWriteAttributesPerAspect: 50
      maxDependencyHops: 10
      minDependencyStrength: 0.3
      maxBranching: 4
      maxTraversalPaths: 50
      minConfidence: 0.5
      timeoutMs: 500
      boostWeight: 0.2
      constraintBudgetChars: 1000

The primary flag determines the retrieval strategy. In primary mode, entities are extracted from the query, the graph is walked to collect related memories, and flat hybrid search only runs to fill remaining slots. In supplementary mode (primary: false), the standard hybrid search runs first and traversal results are blended in using boostWeight. Primary mode is faster for entity-dense queries; supplementary mode is more conservative and better for freeform text.

Reranker (reranker)

An optional reranking pass that runs after initial retrieval. An embedding-based reranker is built in (uses cached vectors, no extra LLM calls). Optionally, reranking can call the active extraction provider model.

Field	Default	Range	Description
enabled	true	—	Enable the reranking pass
model	""	—	Model name for the reranker (empty uses embedding-based)
useExtractionModel	false	—	When true, use the extraction provider LLM for reranking and emit a synthesized summary card
topN	20	1-100	Number of candidates to pass to the reranker
timeoutMs	2000	100-30000 ms	Timeout for the reranking call

Autonomous (autonomous)

Controls autonomous maintenance, repair, and mutation behavior.

Field	Default	Description
enabled	true	Allow autonomous pipeline operations (maintenance, repair).
frozen	false	Block autonomous writes; autonomous reads still allowed.
allowUpdateDelete	true	Permit the pipeline to update or delete existing memories.
maintenanceIntervalMs	1800000	How often maintenance runs (30 min). Range: 60s-24h.
maintenanceMode	"execute"	"observe" logs issues; "execute" attempts repairs.

In "observe" mode the worker emits structured log events but makes no changes. When frozen is true, the maintenance interval never starts, though the worker’s tick() method remains callable for on-demand inspection.

Repair budgets (repair)

Repair sub-workers limit how aggressively they re-embed, re-queue, or deduplicate items to avoid overloading providers.

Field	Default	Range	Description
reembedCooldownMs	300000	10s-1h	Min time between re-embed batches
reembedHourlyBudget	10	1-1000	Max re-embed operations per hour
requeueCooldownMs	60000	5s-1h	Min time between re-queue batches
requeueHourlyBudget	50	1-1000	Max re-queue operations per hour
dedupCooldownMs	600000	10s-1h	Min time between dedup batches
dedupHourlyBudget	3	1-100	Max dedup operations per hour
dedupSemanticThreshold	0.92	0.0-1.0	Cosine similarity threshold for semantic dedup
dedupBatchSize	100	10-1000	Max candidates evaluated per dedup batch

Document ingest (documents)

Controls chunking for ingesting large documents into the memory store.

Field	Default	Range	Description
workerIntervalMs	10000	1s-300s	Poll interval for pending document jobs
chunkSize	2000	200-50000	Target chunk size in characters
chunkOverlap	200	0-10000	Overlap between adjacent chunks (chars)
maxContentBytes	10485760	1 KB-100 MB	Max document size accepted

Chunk overlap ensures context is not lost at chunk boundaries. A value of 10-15% of chunkSize is a reasonable starting point.

Guardrails (guardrails)

Content size limits applied during extraction and recall to prevent oversized content from degrading pipeline performance.

Field	Default	Range	Description
maxContentChars	500	50-100000	Max characters stored per memory
chunkTargetChars	300	50-50000	Target chunk size for content splitting
recallTruncateChars	500	50-100000	Max characters returned per memory in recall results

These limits are enforced at the pipeline level. Content exceeding maxContentChars is truncated before storage. Recall results are truncated at recallTruncateChars to keep session context budgets predictable.

Continuity (continuity)

Session checkpoint configuration for continuity recovery. Checkpoints capture periodic snapshots of session state (focus, prompts, memory activity) to aid recovery after context compaction or session restart.

Field	Default	Range	Description
enabled	true	—	Master switch for session checkpoints
promptInterval	10	1-1000	Prompts between periodic checkpoints
timeIntervalMs	900000	60s-1h	Time between periodic checkpoints (15 min default)
maxCheckpointsPerSession	50	1-500	Per-session checkpoint cap (oldest pruned)
retentionDays	7	1-90	Days before old checkpoints are hard-deleted
recoveryBudgetChars	2000	200-10000	Max characters for recovery digest

Checkpoints are triggered by five events: periodic, pre_compaction, session_end, agent, and explicit. Secrets are redacted before storage.

Sub-agents (subagents)

Controls deterministic parent-session context inherited by sub-agent sessions at session-start. This uses stored active transcripts and checkpoints; it does not make an LLM call.

Field	Default	Range	Description
inheritContext	true	—	Inject a compact parent context block when parent lineage is available
tailChars	3000	0-20000	Max transcript tail characters included from the parent session

memory:
  pipelineV2:
    subagents:
      inheritContext: true
      tailChars: 3000

Set inheritContext: false to disable automatic inherited context while leaving the explicit session_search MCP/API surface available.

Telemetry (telemetry)

Anonymous usage telemetry. On by default; set telemetryEnabled: false to opt out. Events are batched and flushed periodically. Sending requires both posthogHost and posthogApiKey. Each install gets a random anonymous id (persisted in the workspace database) used as the PostHog distinct_id, so installs stay countable without being identifiable.

See TELEMETRY.md — the single reference for the event catalog, privacy contract, the open JSONL audit log, runtime opt-out (SIGNET_TELEMETRY_OPTOUT=1), and how to query the data.

Disclosure (issue #1026): signet setup tells users telemetry is on by default and asks whether to disable it. Declining writes telemetryEnabled: false; non-interactive/CI setups keep the default (enabled).

Runtime opt-out: setting SIGNET_TELEMETRY_OPTOUT=1 in the daemon’s environment disables telemetry without touching config — the same knob the install ping honors. CI runners, containers, and scripted environments should set it so automated daemon boots don’t count as installs.

Open telemetry log: every recorded event is appended as one JSON line to <agentsDir>/.daemon/telemetry/events.jsonl — the single inspectable audit surface for exactly what was sent (daemon events and CLI command.invoked lines). CLI command events are also queued in the workspace database and flushed to PostHog in bounded, best-effort batches without awaiting the command. The CLI and daemon use the same persisted install id. No memory content, code, file paths, or personal data are ever included.

Lifecycle events: daemon.started (version, platform, uptime), command.invoked (command name only, never arguments), error.occurred (sanitized crash report — truncated message with user paths stripped, top stack frames with home directories removed, uptime, and rate-limited EventLoopLag reports with measured lag), version.upgraded (from, to).

Development fleet marker: setting SIGNET_TELEMETRY_ENV=dev keeps operator development checkouts in the dataset while making them filterable. The daemon adds deployment: dev to its local and PostHog events, while the CLI adds it to its JSONL-only command events. The native install ping applies the marker and -dev version suffix when the environment is present. The daemon and CLI bun run dev scripts set this marker automatically. The marker does not disable telemetry; use SIGNET_TELEMETRY_OPTOUT=1 when development or automated events should be silenced entirely.

Field	Default	Range	Description
posthogHost	https://us.i.posthog.com	—	PostHog instance URL (empty disables)
posthogApiKey	phc_mLsvJmbmp6e9UarrX9Cq5QtTjVNiiphM9mvi5Xnddd8Q	—	PostHog project API key. Public ingest key by design; shared by daemon and CLI
flushIntervalMs	60000	5s-10min	Time between event flushes
flushBatchSize	50	1-500	Max events per flush batch
retentionDays	90	1-365	Days before local telemetry data is purged
memorySearchQaEnabled	false	boolean	Capture local-only recall QA rows with query text and result snapshots

memorySearchQaEnabled is separate from anonymous telemetry. It writes a local review ledger to SQLite and intentionally includes recall query text and recalled result content, so it is exposed only through analytics-gated endpoints and is never sent to PostHog.

Embedding tracker (embeddingTracker)

Background polling loop that detects stale or missing embeddings and refreshes them in small batches. Runs alongside the extraction pipeline.

Field	Default	Range	Description
enabled	true	—	Master switch
pollMs	5000	1s-60s	Polling interval between refresh cycles
batchSize	8	1-20	Max embeddings refreshed per cycle

The tracker detects embeddings that are missing, have a stale content hash, or were produced by a different model than the currently configured one. It uses setTimeout chains for natural backpressure.