project-lambda-zero · echel0nn · Jul 19, 2026 · Jul 19, 2026 · Jul 19, 2026 · Jul 20, 2026
diff --git a/CHANGELOG.md b/CHANGELOG.md
diff --git a/docs/CONFIG_REGISTRY.md b/docs/CONFIG_REGISTRY.md
@@ -225,6 +225,12 @@ The `platform` namespace is registered with `PlatformConfigSchema`. All fields:
 | `arq_max_tries` | int | `3` | Maximum retry attempts for failed jobs |
 | `arq_keep_result_s` | int | `3600` | Job result retention in Redis |
 | `progress_stream_maxlen` | int | `1000` | Max events per Redis progress stream |
+| `llm_default_model` | str | `antigravity/claude-opus-4-6-thinking` | Default LLM model id (provider/model) when no per-task override is set |
+| `llm_base_url` | str | `https://openrouter.ai/api/v1` | LLM provider API base URL |
+| `llm_default_max_tokens` | int | `4096` | Default max completion tokens |
+| `llm_default_temperature` | float | `0.0` | Default sampling temperature (deterministic) |
+| `llm_tool_timeout_s` | float | `300.0` | Wall-clock ceiling for a single tool-calling step |
+| `llm_kill_switch` | bool | `False` | Operator circuit breaker; short-circuits every LLM call when true |
 | `llm_pipeline_classify_default` | bool | `True` | LLM pipeline classify stage default-on |
 | `llm_pipeline_validate_default` | bool | `True` | LLM pipeline validate stage default-on |
 | `llm_pipeline_gate_default` | bool | `True` | LLM pipeline gate stage default-on |
@@ -240,6 +246,7 @@ The `platform` namespace is registered with `PlatformConfigSchema`. All fields:
 | `llm_human_consultant_hourly_rate` | float | `150.0` | USD/hr used for human-equivalent cost projections |
 | `data_posture_mode` | str | `"standard"` | `transparent` / `standard` / `paranoid` |
 | `data_direction_default` | str | `"bidirectional"` | `inbound` / `local_only` / `bidirectional` |
+| `knowledge_embedding_model` | str | `"bge-m3"` | Selects the KnowledgeService embedding provider: `bge-m3` (1024-dim, default) or `all-MiniLM-L6-v2` (384-dim, zero-padded to the 1024 column) |
 
 Per-task-type overrides land under namespaced keys (e.g.
 `llm_pipeline_classify_<task_type>`, `llm_pipeline_classify_fail_mode_<task_type>`,
@@ -248,6 +255,44 @@ Per-task-type overrides land under namespaced keys (e.g.
 
 ---
 
+## Dynamic key families
+
+Some configuration is per-task-type or per-team, so the key space is open (for
+example `llm_model_{task_type}` or `llm_monthly_budget_usd_{team_id}`) and
+cannot be enumerated as a fixed set of schema fields. A namespace schema
+declares typed dynamic-key families in a `__dynamic_families__` class
+attribute; each family carries a prefix, a `value_type`, and a default. When a
+key does not match an exact static field, ConfigRegistry resolves it to the
+longest-matching family, so the key is settable through `PUT /config` and cast
+on read exactly as a static field is. A key that matches no static field and
+no family is rejected by `set`.
+
+The `platform` namespace declares the following families:
+
+| Prefix | Type | Purpose |
+|--------|------|---------|
+| `llm_model_` | str | Per-task-type model id |
+| `llm_max_tokens_` | int | Per-task-type max output tokens |
+| `llm_temperature_` | float | Per-task-type sampling temperature |
+| `llm_max_tool_steps_` | int | Per-task-type tool-call loop cap |
+| `llm_tool_timeout_s_` | float | Per-task-type per-tool timeout (seconds) |
+| `llm_data_direction_` | str | Per-task-type data-direction constraint |
+| `llm_budget_max_total_tokens_` | int | Per-task-type token budget ceiling |
+| `llm_monthly_budget_usd_` | float | Per-team monthly budget ceiling (USD) |
+| `llm_pipeline_gate_high_threshold_` | float | Per-task-type HIGH-confidence gate threshold |
+| `llm_pipeline_gate_medium_threshold_` | float | Per-task-type MEDIUM-confidence gate threshold |
+| `llm_pipeline_gate_reject_threshold_` | float | Per-task-type REJECT-confidence gate threshold |
+| `llm_pipeline_gate_consensus_strategy_` | str | Per-task-type consensus strategy (`same_model_high_temp` or `cross_model`) |
+| `llm_pipeline_gate_consensus_model_` | str | Per-task-type consensus model id for `cross_model` |
+| `llm_pipeline_gate_consensus_retries_` | int | Per-task-type consensus retry count |
+| `llm_pipeline_verify_threshold_` | float | Per-task-type cross-model verification threshold |
+| `llm_pipeline_verify_model_` | str | Per-task-type verifier model id |
+| `llm_pipeline_pre_call_steps_` | str | Per-task-type pre-call pipeline step order (comma-separated) |
+| `llm_pipeline_post_call_steps_` | str | Per-task-type post-call pipeline step order (comma-separated) |
+| `llm_pipeline_` | str | Generic pipeline step enable / fail-mode override (bool or `open` / `closed`; callers coerce) |
+
+---
+
 ## Module Config Schemas
 
 Modules register their own config schemas under their own namespace. For example,
@@ -278,6 +323,47 @@ Or via env var:
 export AILA_VULNERABILITY_SOME_FIELD=new_value
 ```
 
+### VR namespace (`vr`)
+
+Registered by the VR module via `VRConfigSchema`. All keys are settable via
+`PUT /config/vr/{key}` and read through ConfigRegistry.
+
+| Key | Type | Purpose |
+|-----|------|---------|
+| `llm_model` | str | LLM model id used by every VR agent (empty falls back to the platform default) |
+| `nday_max_turns` | int | Turn cap for the n-day researcher agent |
+| `nday_tool_time_seconds` | float | Per-tool timeout for the n-day researcher |
+| `poc_max_attempts` | int | Max PoC generation attempts per finding |
+| `poc_reliability_target` | str | Required PoC reliability class |
+| `poc_timeout_seconds` | float | Wall-clock ceiling for a single PoC run |
+| `poc_memory_limit_mb` | int | Memory ceiling for a single PoC run (MB) |
+| `ssh_command_timeout_seconds` | float | Timeout for individual SSH commands on the analysis host |
+| `audit_mcp_url` | str | Base URL for the audit-mcp indexer bridge |
+| `ida_headless_url` | str | Base URL for the ida-headless-mcp bridge |
+| `android_mcp_url` | str | Base URL for the android-mcp bridge |
+| `max_branches_per_investigation` | int | Hard cap on active branches per investigation |
+| `claim_verifier_auto_promote_floor` | float | Minimum verifier confidence to auto-promote a claim |
+| `investigation_total_turn_cap` | int | Aggregate turn ceiling across all branches of one investigation |
+| `zombie_task_heartbeat_min` | int | Minutes without a heartbeat before a running task is treated as zombie |
+| `cursor_cleanup_batch` | int | Rows per batch when the cursor reaper deletes crashed cursors |
+| `stale_branch_frozen_min` | int | Minutes without progress before a branch is marked `frozen` |
+| `stale_branch_halted_min` | int | Minutes without progress before a branch is marked `halted` |
+| `ingestion_poll_timeout_s` | float | Wall-clock ceiling for target-ingestion poll loops |
+
+### Forensics namespace (`forensics`)
+
+Registered by the forensics module via `ForensicsConfigSchema`. All keys are
+read through ConfigRegistry.
+
+| Key | Type | Purpose |
+|-----|------|---------|
+| `llm_model` | str | LLM model id for every forensics agent (empty falls back to the platform default) |
+| `ssh_command_timeout_seconds` | float | Timeout for individual SSH commands on the analyzer machine |
+| `script_execution_timeout_seconds` | float | Timeout for agent-generated script execution |
+| `collection_timeout_seconds` | float | Timeout for the full artifact collection pipeline |
+| `freeflow_max_attempts` | int | Max script-execution attempts per free-flow investigation |
+| `freeflow_max_cost_usd` | float | Hard per-investigation LLM spend ceiling in USD; `0.0` disables the ceiling |
+
 ---
 
 ## ConfigRegistry vs Settings

diff --git a/docs/DATABASE_MIGRATIONS.md b/docs/DATABASE_MIGRATIONS.md
@@ -40,7 +40,21 @@ src/aila/
       064_vr_branch_persona_voice_not_null.py
       065_task_records_input_hash_unique.py
       066_task_records_status_check.py
-      067_workflow_state_cursor_archived_state.py    # current head
+      067_workflow_state_cursor_archived_state.py
+      068_malware_module_tables.py
+      069_malware_audit_fixes.py
+      070_malware_project_slim_and_ack.py
+      071_platform_journal.py
+      072_malware_observation_dedup.py
+      073_scheduled_report_team_scope.py
+      074_automation_tz_disable.py
+      075_hot_column_indexes.py
+      076_oidc_default_team.py
+      077_knowledge_embedding_1024.py
+      078_notification_created_index.py
+      079_refresh_token_client_meta.py
+      080_platform_schema_reconcile.py
+      081_vr_schema_reconcile.py                     # current head
 ```
 
 `alembic.ini` and the `alembic/` directory live under `src/aila/`, not at the repo root. All Alembic commands must run from `src/aila/`:
@@ -86,29 +100,29 @@ Where `{NNN}` is the next sequential number (zero-padded to 3 digits). Check the
 
 ```bash
 ls src/aila/alembic/versions/*.py | tail -1
-# 067_workflow_state_cursor_archived_state.py -> next is 068
+# 081_vr_schema_reconcile.py -> next is 082
 ```
 
 ### Step 3: Write the migration
 
 Use this template:
 
 ```python
-"""063 -- add priority column to task records.
+"""082 -- add priority column to task records.
 
 Adds a nullable priority field so operators can influence queue ordering.
 
-Revision ID: 063_taskrecord_priority
-Revises: 067_workflow_state_cursor_archived_state
-Create Date: 2026-06-10
+Revision ID: 082_taskrecord_priority
+Revises: 081_vr_schema_reconcile
+Create Date: 2026-07-22
 """
 from __future__ import annotations
 
 import sqlalchemy as sa
 from alembic import op
 
-revision = "063_taskrecord_priority"
-down_revision = "067_workflow_state_cursor_archived_state"
+revision = "082_taskrecord_priority"
+down_revision = "081_vr_schema_reconcile"
 branch_labels = None
 depends_on = None
 
@@ -168,6 +182,20 @@ Examples from the codebase:
 | `065_task_records_input_hash_unique.py` | Platform column + partial UNIQUE |
 | `066_task_records_status_check.py` | Platform CHECK constraint |
 | `067_workflow_state_cursor_archived_state.py` | Platform column addition |
+| `068_malware_module_tables.py` | Malware module initial tables (17 tables) |
+| `069_malware_audit_fixes.py` | Module column additions + findings table rebuild |
+| `070_malware_project_slim_and_ack.py` | Module column drops + acked_at column |
+| `071_platform_journal.py` | Platform append-only hash-chained journal + dead-letter |
+| `072_malware_observation_dedup.py` | Module column + partial UNIQUE index |
+| `073_scheduled_report_team_scope.py` | Platform team_id column + backfill |
+| `074_automation_tz_disable.py` | Platform cron timezone + disable reason columns |
+| `075_hot_column_indexes.py` | Platform indexes on hot query columns |
+| `076_oidc_default_team.py` | Platform column addition (OIDC default team) |
+| `077_knowledge_embedding_1024.py` | Platform column type widening to vector(1024) |
+| `078_notification_created_index.py` | Platform composite index on notifications |
+| `079_refresh_token_client_meta.py` | Platform column additions (auth client metadata) |
+| `080_platform_schema_reconcile.py` | Platform schema drift reconciliation (types + constraint promotion) |
+| `081_vr_schema_reconcile.py` | Module UNIQUE constraint renames on VR tables |
 
 The `revision` string inside the file must match the filename (without `.py`).
 

diff --git a/docs/DB_SCHEMA.md b/docs/DB_SCHEMA.md
@@ -4,13 +4,16 @@ All SQLModel tables used by AILA, organized by ownership (platform vs module).
 
 PostgreSQL 16 with the `pgvector` extension is the only supported backend. Tables
 use SQLModel (SQLAlchemy Core) DDL; `sa_column=Column(Text)` carries large-text
-columns; `pgvector` carries 384-dim embedding columns. asyncpg is the runtime
+columns; `pgvector` carries the 1024-dim knowledge embedding column produced by
+BGE-M3 (`BAAI/bge-m3`). The `all-MiniLM-L6-v2` fallback (384-dim) is zero-padded
+to 1024 by `KnowledgeService.embed` when it is selected via the
+`knowledge_embedding_model` config key. asyncpg is the runtime
 driver; Alembic swaps to psycopg automatically via `src/aila/alembic/env.py`.
 
 Two creation paths coexist:
 - Platform + module tables that predate the Alembic baseline (`001_baseline_stamp`)
   are created on first boot by `make db-init`, which runs `SQLModel.metadata.create_all()`
-  then stamps `alembic_version` at the current head (`067_workflow_state_cursor_archived_state`).
+  then stamps `alembic_version` at the current head (`081_vr_schema_reconcile`).
 - Every schema change since then ships as an Alembic revision under
   `src/aila/alembic/versions/`. See [`DATABASE_MIGRATIONS.md`](DATABASE_MIGRATIONS.md).
 
@@ -197,7 +200,7 @@ Vector-indexed knowledge entry for agent retrieval.
 | id | int | PK, auto | Row identity |
 | namespace | str | indexed | Agent class name |
 | content | Text | | Knowledge text |
-| embedding | LargeBinary | | 384-dim float32 BLOB |
+| embedding | Vector(1024) | pgvector, HNSW cosine index | BGE-M3 1024-dim vector; MiniLM 384-dim output zero-padded to fit when the fallback provider is selected |
 | entry_metadata | Text | default="{}" | JSON metadata |
 | dedup_key | Text/null | indexed, UQ(namespace,dedup_key) | Deduplication key |
 | created_at | datetime | default=utc_now | Creation timestamp |
@@ -570,7 +573,7 @@ Owned by `aila.modules.vr.db_models`. Created and evolved by migrations 040 + 04
 
 ## Table Summary
 
-Counts reflect the Alembic head `067_workflow_state_cursor_archived_state` (2026-06-21).
+Counts reflect the Alembic head `081_vr_schema_reconcile` (2026-07-22).
 
 | Group | Owner | Tables |
 |---|---|---|
@@ -590,4 +593,4 @@ The `hello_world` module ships as a reference and does not own any DB tables.
 ---
 
 *Generated from source models in `src/aila/storage/db_models.py`, `src/aila/platform/tasks/models.py`, `src/aila/platform/llm/cost_record.py`, `src/aila/platform/llm/idempotency_cache.py`, `src/aila/platform/automation/models.py`, and the per-module `db_models/` packages under `src/aila/modules/<module>/`.*
-*Last updated: 2026-06-21 (Alembic head `067_workflow_state_cursor_archived_state`).*
+*Last updated: 2026-07-22 (Alembic head `081_vr_schema_reconcile`).*
diff --git a/docs/ENV_VARS.md b/docs/ENV_VARS.md
@@ -152,9 +152,10 @@ through PowerShell.
 
 ### AILA_PLATFORM_LLM_DEFAULT_MODEL / _BASE_URL / _DEFAULT_MAX_TOKENS
 
+- **Defaults:** `antigravity/claude-opus-4-6-thinking`, `https://openrouter.ai/api/v1`, `4096` (from `PlatformConfigSchema`).
 - **Used in:** `src/aila/platform/llm/config.py` (`LLMConfigResolver`)
-- **Resolution:** ConfigRegistry env-var override path. These keys live under the `platform` namespace but are NOT in `PlatformConfigSchema` -- the env var sets the value, the DB row carries persistent overrides, and the resolver bakes in its own fallbacks (`openai/gpt-4o-mini`, `https://openrouter.ai/api/v1`, `4096`) when nothing matches.
-- **Production guidance:** Set per deployment to pin the default model/provider for every task type. Per-task-type overrides land under `AILA_PLATFORM_LLM_MODEL_<TASK_TYPE>` and `AILA_PLATFORM_LLM_MAX_TOKENS_<TASK_TYPE>`.
+- **Resolution:** Declared as static fields on `PlatformConfigSchema` under the `platform` namespace. Reads go through the standard ConfigRegistry chain (env var, then in-memory cache, then the persisted DB row, then the schema default). The env var wins over the DB, so setting `AILA_PLATFORM_LLM_DEFAULT_MODEL` overrides any value stored through `PUT /config/platform/llm_default_model`.
+- **Production guidance:** Set per deployment to pin the default model, provider base URL, and completion ceiling for every task type. Runtime edits without a restart go through `PUT /config/platform/llm_default_model` (and its `_base_url` / `_default_max_tokens` siblings). Per-task-type overrides land under `AILA_PLATFORM_LLM_MODEL_<TASK_TYPE>` and `AILA_PLATFORM_LLM_MAX_TOKENS_<TASK_TYPE>`, or via the same `PUT /config` endpoint against the matching dynamic-key family.
 
 ### AILA_LLM_TIMEOUT_SECONDS
 

diff --git a/docs/LLM_INTEGRATION.md b/docs/LLM_INTEGRATION.md
@@ -714,20 +714,30 @@ All LLM configuration lives in the `platform` namespace of ConfigRegistry. Chang
 | `llm_default_temperature` | `0.0` | Default sampling temperature (deterministic). |
 | `llm_temperature_{task_type}` | *(unset)* | Per-task-type temperature override. |
 | `llm_max_tool_steps_{task_type}` | `0` | Max tool-calling loop iterations per task type. `0` = tool calling disabled. |
+| `llm_tool_timeout_s` | `300.0` | Default wall-clock ceiling (seconds) for a single tool-calling step. |
+| `llm_tool_timeout_s_{task_type}` | *(unset)* | Per-task-type per-tool timeout override (seconds). |
 
 ### Pipeline Step Toggling
 
 | Key | Default | Description |
 |-----|---------|-------------|
 | `llm_pipeline_{step}_{task_type}` | `true` | Enable/disable a pipeline step for a task type. Steps: `classify`, `validate`, `gate`, `seal`. |
 | `llm_pipeline_{step}_fail_mode_{task_type}` | `open` | Fail mode for a pipeline step. `"open"` = log and continue. `"closed"` = raise `LLMError`. |
+| `llm_pipeline_pre_call_steps_{task_type}` | *(unset)* | Comma-separated pre-call step-order override for a task type (e.g. `classify,validate`). |
+| `llm_pipeline_post_call_steps_{task_type}` | *(unset)* | Comma-separated post-call step-order override for a task type (e.g. `gate,seal`). |
 
 ### Classification
 
 | Key | Default | Description |
 |-----|---------|-------------|
 | `llm_pipeline_classify_restricted_behavior_{task_type}` | `fail` | RESTRICTED data behavior. `"fail"` = raise `ClassificationBlockedError`. `"redact"` = replace sensitive tokens with `[REDACTED-*]` tags and continue. |
 
+### Data Direction
+
+| Key | Default | Description |
+|-----|---------|-------------|
+| `llm_data_direction_{task_type}` | *(falls back to `data_direction_default`)* | Per-task-type data-direction constraint. Values: `inbound`, `local_only`, `bidirectional`. Used by the classify stage to reject calls that would send data in a forbidden direction. |
+
 ### Confidence Gating
 
 | Key | Default | Description |
@@ -752,6 +762,7 @@ All LLM configuration lives in the `platform` namespace of ConfigRegistry. Chang
 | Key | Default | Description |
 |-----|---------|-------------|
 | `llm_budget_max_total_tokens_{task_type}` | `0` | Max total tokens per run for a task type. `0` = unlimited (no enforcement). |
+| `llm_monthly_budget_usd_{team_id}` | *(unset)* | Per-team monthly spend ceiling in USD. Evaluated against the running sum of `LLMCostRecord.cost_usd` rows for the team over the current calendar month; a call that would cross the ceiling is refused before the API request. |
 
 ### Common operations