Instinct layer

Enterprise Edition

This capability is part of the Enterprise Edition — a proprietary overlay on the open-source core. See Editions for the full Community vs Enterprise matrix.

The instinct layer lets vornik learn from its own history. It mines the telemetry every task already produces, distils recurring patterns into confidence-scored instincts — "in situation T, action A held up" — and makes those instincts available as evidence to the parts of vornik that decide how the next task runs.

The defining rule: instincts are advisory. Nothing mutates silently. An instinct can raise the confidence of a workflow proposal an operator is already reviewing, or surface a remediation hint on a failed task — but it never edits a prompt, swaps a model, deletes a memory, or promotes a workflow on its own. Every behaviour-affecting decision still passes through the same operator and review gates it always did. With the layer disabled (the default), behaviour is unchanged.

How it learns

A background worker runs on a cadence (30 minutes by default) and reads the audit spine — the step outcomes, tool-call log, and memory-retrieval log that vornik writes for every task. Deterministic extractors turn finalised outcomes into atomic patterns; only genuinely ambiguous clusters fall through to a cheap distillation model. Each pattern is deduplicated by a canonical trigger key and scored.

Confidence and decay

An instinct's confidence is a Wilson lower bound on its success rate (so a pattern seen 3 times is trusted far less than one seen 300 times), multiplied by a recency decay with a 30-day half-life (so a pattern that stops recurring fades rather than lingering forever). An instinct moves through states as the evidence accumulates:

State	Reached when
candidate	freshly mined; not yet trusted
active	confidence ≥ 0.6 and seen at least 3 times
promoted	holds across ≥ 2 projects at confidence ≥ 0.8 (becomes a global prior)
retired	confidence falls below 0.2 (sticky — stays retired)

These thresholds are tunable (see Configuration).

What consumes instincts

Four consumers read instincts, each independently switchable and each strictly advisory:

• Failure-class playbooks — when a task fails, high-confidence remediation instincts for that failure class are surfaced on the task's detail page and offered to the lead as recovery context. The operator's recovery-approval gate is unchanged.
• Architect evidence priors — when the workflow architect proposes a change, matching instincts act as priors. They can only nudge the proposal's confidence up toward the strongest supporting evidence (never above what the model itself reported, never widening the edit's scope); a rejected proposal is written back as contrary evidence.
• Memory hygiene — instincts suggest which memory scopes to boost and which chunks are prune candidates. The consolidation sweeper surfaces these; it never deletes on their say-so, and firewall-sensitive chunks are screened out.
• Tool-budget tiers — when a planner gives no explicit complexity verdict, a learned tier can fill the gap. An explicit verdict always wins, and the configured budget caps still bound the result.

There is an optional auto_apply mode (off by default, gated by its own confidence floor) that changes only how strongly a recovery remediation is presented to the lead — it does not bypass the approval gate.

Enabling it

Run the feature doctor to diagnose prerequisites and enable:

vornikctl doctor feature enable instinct

This sets instinct.enabled plus the consumer gates and restarts the daemon during an idle window. Because it is restart-gated, the doctor refuses to apply while tasks are running and waits for an idle window. (The tool-budget and auto_apply consumers stay manually opt-in.)

Inspecting what it learned

Everything the layer knows is browsable. Use the CLI:

vornikctl instinct list --domain recovery --min-confidence 0.8
vornikctl instinct show <instinct_id>
vornikctl instinct export -o instincts.json        # filtered snapshot
vornikctl instinct retire <instinct_id>            # advisory; the row stays for audit

list filters by --domain, --scope, --project, --status, and --min-confidence. The same inventory is available as a read-only browser at /ui/admin/instincts, and failed tasks show their learned remediations inline on the task detail page. All admin surfaces require an admin-scoped key.

Configuration

The layer is configured under the instinct block. The most useful keys:

Key	Default	Meaning
instinct.enabled	false	master switch
instinct.cadence_seconds	1800	how often the miner runs
instinct.min_support	3	observations before a pattern can leave candidate
instinct.active_confidence	0.6	promotion-to-active threshold
instinct.promote_confidence	0.8	cross-project global-promotion threshold
instinct.promote_projects	2	projects a pattern must hold across to go global
instinct.retire_floor	0.2	confidence below which an instinct retires
instinct.decay_halflife_days	30	recency-decay half-life
instinct.model	chat.model	distillation model for ambiguous clusters

The four consumers are toggled under instinct.consumers: failure_playbooks, architect_priors, memory_hygiene, application_feedback, and tool_budget. The optional auto-apply mode lives under instinct.consumers.auto_apply (enabled, min_confidence, allowed_error_classes).