@event4u/agent-config 7.0.1 → 7.1.0

package/docs/contracts/ai-council-config.md

@@ -33,23 +33,30 @@ fallback during the v2.4 namespace transition). This is where a single
 developer enables members, pins models, and sets caps once — and the
 council then works in **every** project they open, with no per-project
 file to check in or accidentally commit. See
-[ADR-093](../decisions/ADR-093-ai-council-config-user-global.md).
+[ADR-093](../decisions/ADR-093-ai-council-config-user-global.md) and
+[ADR-104](../decisions/ADR-104-ai-council-config-global-only.md).
 
-`scripts/ai_council/config.py:resolve_config_path` resolves the active
-file with this precedence (first match wins):
+**The council config is ALWAYS user-global. The project tree is NEVER
+searched for it.** `scripts/ai_council/config.ts:resolve_config_path`
+resolves the active file with this precedence (first match wins):
 
 1. **`$AI_COUNCIL_CONFIG`** — an explicit absolute path. Honoured even
    when the target is absent (so a typo surfaces as "create it here"
    rather than silently falling back). Primarily for tests / power users.
-2. **Project-local** `<project_root>/agents/settings/.ai-council.yml` —
-   only when a consumer project deliberately checks in its own council
-   config. This **overrides** the user-global file for that project.
-3. **User-global** `~/.event4u/agent-config/settings/.ai-council.yml` — the
+   This is an explicit path, not a project search.
+2. **User-global** `~/.event4u/agent-config/settings/.ai-council.yml` — the
    canonical default described above.
 
-When none exists, the loader reports the user-global path as the place to
-create it. `agents/templates/.ai-council.yml.example` ships the documented
-shape to copy from. The legacy `ai_council.*` block under
+There is **no project-local lookup.** A `<project_root>/agents/settings/.ai-council.yml`
+in any project (including this package's own tree) is ignored — the
+council never reads it (ADR-104, superseding the project-local override
+ADR-093 had kept). The absence of a council file *in a project* therefore
+says **nothing** about whether the council is configured; only the
+user-global file (or the `$AI_COUNCIL_CONFIG` override) decides that.
+
+When neither exists, the loader reports the user-global path as the place
+to create it. `agents/templates/.ai-council.yml.example` ships the
+documented shape to copy from. The legacy `ai_council.*` block under
 `.agent-settings.yml` remains removed (Phase 0 Step 5); a one-line
 breadcrumb there points at this contract.
 

package/docs/contracts/audit-log-v1.md

@@ -87,6 +87,13 @@ Single-line. The pretty-printed reference shape:
 Unknown trailing fields are forward-compat extensions; readers MUST
 NOT raise on them.
 
+A line MAY carry an optional `orchestration` object when the run was
+produced by the auto-dispatch orchestration layer. It is additive and
+non-breaking — `schema_version` is unchanged, no existing field is
+removed or renamed, and readers that do not understand it ignore it per
+the forward-compat rule above. Its shape lives in
+[`orchestration-telemetry.md`](../../src/agent-src/contexts/execution/orchestration-telemetry.md).
+
 ## Privacy floor
 
 Lines MUST NOT contain:

package/docs/contracts/auto-orchestration-v1.md

@@ -0,0 +1,70 @@
+---
+stability: beta
+keep-beta-until: 2026-09-23
+---
+
+# Auto-Orchestration v1
+
+**Purpose.** Pin the data shapes the automatic subagent-orchestration layer
+depends on — the host-capability manifest, the spawn brief, and the
+orchestration telemetry object — so producers and consumers agree without
+re-deriving them. Behaviour (when to dispatch, how to route, how to verify)
+lives in the context docs; this contract pins the **shapes**.
+
+**Scope.** The three JSON shapes below + the activation precondition. Does not
+define classification rules, routing policy, or the verification budget — those
+are the cited context docs.
+
+Last refreshed: 2026-06-23. Decided in
+[ADR-105](../decisions/ADR-105-automatic-subagent-orchestration.md).
+
+## Host-capability manifest
+
+Resolved once per session, cached. Safe default = all-false.
+
+```json
+{ "schema_version": 1, "subagent_spawn": false, "parallel_spawn": false, "status_polling": false, "separate_quota_pool": false }
+```
+
+Full semantics: [`host-capability-manifest`](../../src/agent-src/contexts/execution/host-capability-manifest.md).
+
+## Spawn brief
+
+Composed per delegated sub-task; knowledge is references only, capped.
+
+```json
+{ "task": "<sub-task>", "role_mode": "reviewer|null", "profile": "<id|null>", "personas": ["<id>"], "knowledge_refs": ["<id-or-path>"] }
+```
+
+Full semantics: [`subagent-spawn-contract`](../../src/agent-src/contexts/execution/subagent-spawn-contract.md).
+
+## Orchestration telemetry
+
+Optional `orchestration` object on an [`audit-log-v1`](audit-log-v1.md) line —
+additive, non-breaking, counts + ids only.
+
+```json
+{ "task_size_estimate": 0, "spawn_count": 0, "tiers": [], "token_delta": 0, "wall_clock_ms": 0, "outcome": "DONE", "verify_mode": "deterministic" }
+```
+
+Full semantics: [`orchestration-telemetry`](../../src/agent-src/contexts/execution/orchestration-telemetry.md).
+
+## Activation precondition
+
+Auto-dispatch is attempted only when:
+
+```
+subagents.enabled AND subagents.auto != off AND manifest.subagent_spawn AND task-is-delegable
+```
+
+Any failing → in-session. Full gate:
+[`auto-orchestration-activation`](../../src/agent-src/contexts/execution/auto-orchestration-activation.md).
+
+## Cross-references
+
+- [ADR-105](../decisions/ADR-105-automatic-subagent-orchestration.md) — the decisions.
+- [`subagent-routing`](../../src/agent-src/contexts/execution/subagent-routing.md) — model downshift + quota bonus.
+- [`auto-dispatch-classification`](../../src/agent-src/contexts/execution/auto-dispatch-classification.md) — delegability rules.
+- [`verify-budget`](../../src/agent-src/contexts/execution/verify-budget.md) — verification budget.
+- [`subagent-steering`](../../src/agent-src/contexts/execution/subagent-steering.md) — guardrails + kill-switch.
+- [`orchestration-benchmark-gate`](../../src/agent-src/contexts/execution/orchestration-benchmark-gate.md) — the default-flip gate.

package/docs/contracts/command-clusters.md

@@ -28,7 +28,7 @@ column 1 of this table.
 
 | Cluster | Phase | Sub-commands | Replaces |
 |---|:-:|---|---|
-| `fix` | 1 | `ci` · `pr-comments` · `pr-bot-comments` · `pr-developer-comments` · `portability` · `refs` · `seeder` | `fix-ci` · `fix-pr-comments` · `fix-pr-bot-comments` · `fix-pr-developer-comments` · `fix-portability` · `fix-references` · `fix-seeder` |
+| `fix` | 1 | `ci` · `pr-comments` · `pr-bot-comments` · `pr-developer-comments` · `portability` · `refs` · `seeder` · `comments` | `fix-ci` · `fix-pr-comments` · `fix-pr-bot-comments` · `fix-pr-developer-comments` · `fix-portability` · `fix-references` · `fix-seeder`; `comments` added 2026-06-22 — audits the **code comments** in the current branch's diff (simplify / shorten / remove), distinct from the PR-review-thread `pr-comments` sub |
 | `optimize` | 1 | `agents-dir` · `augmentignore` · `rtk` · `skills` | `optimize-augmentignore` · `optimize-rtk-filters` · `optimize-skills` · former `/optimize agents` and `/optimize agents-md` moved to the `/agents` file-family cluster 2026-05-09; `/agents prepare/audit/cleanup` collapsed into the single `/optimize agents-dir` (flags or wizard) per the agent-doc consolidation |
 | `feature` | 1 | `explore` · `plan` · `refactor` · `roadmap` · `dev` | `feature-explore` · `feature-plan` · `feature-refactor` · `feature-roadmap` · `feature-dev` |
 | `chat-history` | 2 | `show` · `import` · `learn` | `chat-history` (legacy status) — `resume` / `clear` / `checkpoint` removed in `road-to-chat-history-hook-only` (auto-adopt + structural hooks); `import` (verbatim cross-session render) and `learn` (project-improving learning extraction) added in the v4 stateless schema |

package/docs/contracts/discovery-manifest.schema.json

@@ -74,7 +74,7 @@
     },
     "workspace_id": {
       "type": "string",
-      "enum": ["engineering", "product", "finance", "founder", "gtm", "ops", "small-business", "construction", "agent-config-maintainer"]
+      "enum": ["engineering", "product", "finance", "founder", "gtm", "ops", "small-business", "construction", "legal", "agent-config-maintainer"]
     },
     "pack_id": {
       "type": "string",
@@ -84,7 +84,7 @@
         "finance-basic", "finance-advanced",
         "gtm-sales", "gtm-marketing",
         "ops-people", "founder-strategy", "small-business", "construction", "ai-video", "ai-image", "fun", "meta", "git", "frontend-design", "brand",
-        "memory", "analytics", "product-reasoning", "analysis-workbench"
+        "memory", "analytics", "product-reasoning", "analysis-workbench", "legal"
       ]
     },
     "lifecycle": {

package/docs/decisions/ADR-013-discovery-frontmatter-contract.md

@@ -93,6 +93,7 @@ install:
 | `ops` | Operations | Hiring, comp, perf, org design, runbooks. |
 | `small-business` | Small Business | Self-employed / SMB-shaped owner workflows. |
 | `construction` | Construction | Trade-business workflows (planning, quotes). |
+| `legal` | Legal | In-house legal — contract/NDA/DPA review, privacy, legal triage (EU/DE scope). |
 | `agent-config-maintainer` | Maintainer | Skills/rules/commands that maintain *this* package. |
 
 Amendments to the workspace list require an ADR-013 amendment.
@@ -120,6 +121,7 @@ Amendments to the workspace list require an ADR-013 amendment.
 | `founder-strategy` | Vision, fundraising narrative, competitive moat. (Renamed from `founder` to honour the non-overlap rule against `profile.id: founder`.) |
 | `small-business` | SMB-shaped owner workflows. |
 | `construction` | Trade-business workflows. |
+| `legal` | In-house legal — EU/DE-scoped contract/NDA/DPA review and legal triage; procedure-only, attorney-review draft output. |
 | `ai-video` | AI video pipeline (per ADR-011, the only heavyweight domain). |
 | `fun` | Non-essential social/fun workflows (prediction-pool tips, etc.). Optional under the `small-business` workspace; `experimental` trust. |
 | `meta` | Artefacts that maintain *this* package (`agent-config` itself). |
@@ -295,6 +297,23 @@ consistency governance that constrains UI). Mirrored in
 2026-06-13 (road-to-image-brand-typography Phase B). Additive, no rename;
 non-overlap with cost-profile and `profile.id` reservations holds.
 
+### 2026-06-24 — New `legal` workspace + `legal` pack
+
+Added workspace id `legal` and pack id `legal` to the closed vocabularies
+(in-house legal — EU/DE-scoped contract/NDA/DPA review and legal triage;
+procedure-only, no default legal positions; every output a draft for attorney
+review). Mirrored in
+[`src/config/discovery/workspaces.yml`](../../src/config/discovery/workspaces.yml),
+[`src/config/discovery/packs.yml`](../../src/config/discovery/packs.yml), and the
+`ADR_WORKSPACES` + `ADR_PACKS` sets in
+[`src/scripts/lint_discovery_vocabulary.ts`](../../src/scripts/lint_discovery_vocabulary.ts).
+`workspaces: [legal]`, `domain: legal`, `size_class: medium`,
+`trust_level_default: experimental`, `surface_tier: lab` (gated lab-tier until the
+Gate-2 maintenance owner + external qualification is confirmed —
+road-to-legal-pack Phase 0.2). Council-resolved across three rounds 2026-06-24
+(road-to-legal-pack). Additive, no rename; non-overlap with cost-profile and
+`profile.id` reservations holds.
+
 ## Cross-references
 
 - [ADR-007 — Agent Discovery Scopes](ADR-007-agent-discovery-scopes.md):

package/docs/decisions/ADR-020-global-only-consumer-scope.md

@@ -108,12 +108,19 @@ corrects this:
 - The global surface is **secrets** (provider keys, already global) plus the
   curated identity **whitelist** (`MERGEABLE_KEYS`) — not all of `personal.*`,
   not council config, not personas.
-- **Council configuration and personas stay project-local** (version-controlled).
-  The `MERGEABLE_KEYS` whitelist is a **security trust boundary**: an untrusted
-  project repo must not be able to escalate `ai_council` / arbitrary `personal.*`
-  into user scope. Widening the whitelist now requires a threat model.
-- The key name is **`ai_council`** (canonical, in `agents/settings/.ai-council.yml`),
-  never `agent_council`.
+- **Council config is NOT on the `MERGEABLE_KEYS` whitelist.** _(Placement
+  amended: per [ADR-093](ADR-093-ai-council-config-user-global.md) →
+  [ADR-104](ADR-104-ai-council-config-global-only.md) the council config
+  is now **user-global** at `~/.event4u/agent-config/settings/.ai-council.yml`,
+  never project-local — but it is its own dedicated `.ai-council.yml`, not a
+  merged `.agent-settings.yml` key, so it stays off this whitelist.)_ Personas
+  stay project-local (version-controlled). The `MERGEABLE_KEYS` whitelist is a
+  **security trust boundary**: an untrusted project repo must not be able to
+  escalate arbitrary `personal.*` into user scope. Widening the whitelist now
+  requires a threat model.
+- The council config key name is **`ai_council`** (canonical), never
+  `agent_council`. It lives in the user-global `.ai-council.yml` (ADR-104),
+  not under `.agent-settings.yml`.
 
 This amendment narrows ADR-020's wording; the global-only **consumer install**
 decision (distributed content writes only to `~/.event4u/agent-config/`) is

package/docs/decisions/ADR-049-configuration-trust-boundary.md

@@ -13,7 +13,17 @@ type: decision
 
 ## Status
 
-**Accepted** · 2026-06-03. Locks the AI-council convergence
+**Accepted** · 2026-06-03. **Council-config *placement* superseded by
+[ADR-093](ADR-093-ai-council-config-user-global.md) → [ADR-104](ADR-104-ai-council-config-global-only.md)**
+(council config now lives user-global at
+`~/.event4u/agent-config/settings/.ai-council.yml`, never project-local).
+The **trust-boundary decision below still stands**: `ai_council` is not on
+the user-global *settings merge whitelist* (`.agent-settings.yml` keys) —
+the council reads its own dedicated `.ai-council.yml`, not a merged
+settings key. Wherever this ADR says council config "lives project-local",
+read "lives user-global per ADR-104".
+
+Locks the AI-council convergence
 (anthropic/claude-sonnet-4-5 + openai/gpt-4o, design mode + peer-review,
 2026-06-03) on where the source of truth for personal preferences and AI-council
 configuration lives. Refines [`ADR-020`](ADR-020-global-only-consumer-scope.md)

package/docs/decisions/ADR-093-ai-council-config-user-global.md

@@ -4,7 +4,7 @@ status: accepted
 date: 2026-06-13
 decision: ai-council-config-user-global
 supersedes: —
-superseded_by: —
+superseded_by: 104
 phase: ai-council-config-relocation
 type: structural
 ---
@@ -13,9 +13,15 @@ type: structural
 
 ## Status
 
-**Accepted** · 2026-06-13. Maintainer directive (standing, restated): the
-council is a per-developer facility and must be configured once, globally,
-not re-declared per project.
+**Accepted** · 2026-06-13. **Partially superseded by
+[ADR-104](ADR-104-ai-council-config-global-only.md)** (2026-06-22): the
+user-global relocation stands, but the project-local override this ADR
+*kept* (precedence rule 2 in § Decision) is removed — the council is now
+**global-only** and never searches the project tree.
+
+Maintainer directive (standing, restated): the council is a per-developer
+facility and must be configured once, globally, not re-declared per
+project.
 
 ## Context
 

package/docs/decisions/ADR-104-ai-council-config-global-only.md

@@ -0,0 +1,119 @@
+---
+adr: 104
+status: accepted
+date: 2026-06-22
+decision: ai-council-config-global-only
+supersedes: 093
+superseded_by: —
+phase: ai-council-config-relocation
+type: structural
+---
+
+# ADR-104 — The AI-council config is global-only (drop the project-local override)
+
+## Status
+
+**Accepted** · 2026-06-22. Maintainer directive (standing, restated and
+sharpened): the council is configured **once, globally** — and the agent
+must never search a project for council config, never infer "council not
+configured" from a project's missing files.
+
+Supersedes the project-local override that
+[ADR-093](ADR-093-ai-council-config-user-global.md) deliberately kept.
+
+## Context
+
+ADR-093 relocated the council config to the user-global namespace
+(`~/.event4u/agent-config/settings/.ai-council.yml`) but **kept** a
+project-local override as precedence rule 2:
+`<project_root>/agents/settings/.ai-council.yml` would win over the
+user-global file when present. ADR-093 also rejected the "global always
+wins" alternative, reasoning it would break a consumer team's
+deliberately-checked-in shared council config.
+
+In practice that hedge produced a recurring, infuriating failure mode that
+the loader's own correctness did *not* prevent — because the failure was
+in the **agent's mental model**, not the resolver:
+
+> *"Council war in diesem Projekt nicht konfiguriert (kein <!-- md-language-check: ignore -->
+> `.agent-settings.yml`/`scripts/ai_council`), daher habe ich die eine <!-- md-language-check: ignore -->
+> Architektur-Gabel selbst entschieden."* <!-- md-language-check: ignore -->
+
+An agent working in a consumer project, asked to consult the council on an
+architecture fork, looked **in the project** for council artefacts — the
+package-internal `scripts/ai_council` directory and the long-removed
+`.agent-settings.yml` `ai_council` block — found neither (they only exist
+inside the agent-config package), and concluded the council was
+unconfigured. It then decided a structural fork solo. The user's council
+*was* configured — globally — and would have answered from any project.
+
+The project-local override is the root of this confusion: as long as "the
+project might carry its own council config" is a documented truth, the
+agent treats the project tree as a place to look, and absence-in-project
+reads as not-configured. The override's intended beneficiary (a consumer
+team pinning a shared config in their repo) has never materialised and is
+not worth the recurring solo-decision failure on real work.
+
+## Decision
+
+**The council config is global-only.** `resolve_config_path` (in
+`scripts/ai_council/config.ts`) resolves with this precedence, first match
+wins:
+
+1. **`$AI_COUNCIL_CONFIG`** — an explicit absolute path (tests / power
+   users). Honoured even when absent. This is an explicit path, not a
+   project search.
+2. **User-global** `~/.event4u/agent-config/settings/.ai-council.yml`
+   (legacy `~/.config/agent-config/` read-fallback).
+
+The project-local branch is **removed**. The loader never builds or stats
+`<project_root>/agents/settings/.ai-council.yml`. `project_root` stays in
+the signature for caller stability but is not consulted for resolution.
+
+The corollary, binding on every agent-facing surface (the `ai-council`
+skill, the `/council` command family, the config contract): **the absence
+of a council file in a project says nothing about whether the council is
+configured.** An agent must never claim "council not configured" from
+missing project files (`.agent-settings.yml`, `scripts/ai_council`, a
+project `.ai-council.yml`). The CLI/resolver is the single authority — it
+reads the global file and works from every project, worktree, and CWD.
+
+## Consequences
+
+- A user-global `.ai-council.yml` with `enabled: true` works from **every**
+  project, including consumer repos and fresh worktrees, with zero
+  per-project setup — exactly the "configure once" intent of ADR-093, now
+  without the hedge that undermined it.
+- The "council not configured in this project" solo-decision failure mode
+  is structurally impossible: there is no project location to be absent.
+- `council_cli.ts` and `cmd_doctor.ts` (`doctor council-cli`) inherit the
+  change automatically — both already route through `resolve_config_path`.
+- A consumer team that genuinely wants a repo-pinned shared council config
+  uses `$AI_COUNCIL_CONFIG` (e.g. exported in their dev shell / CI) — an
+  explicit opt-in, never an implicit project-tree search.
+- `check_council_config_location.ts` keeps guarding against
+  `.agent-settings.yml` council references; its docstring is updated to
+  drop the "project-local first" description.
+
+## Alternatives
+
+- **Keep ADR-093 as-is (project-local override stays).** Rejected — this is
+  the exact configuration that produced the recurring failure; the loader
+  being correct did not save the agent's mental model.
+- **Fix only the agent-facing docs, leave the override in the loader.**
+  Rejected — as long as a project-local lookup exists, "look in the
+  project" remains a documented truth and the absence-reads-as-unconfigured
+  confusion recurs. Removing the lookup makes the instruction unambiguous.
+- **Add a project-local override behind an explicit opt-in flag.** Rejected
+  as unneeded complexity — `$AI_COUNCIL_CONFIG` already covers the rare
+  repo-pinned case explicitly.
+
+## References
+
+- [ADR-093](ADR-093-ai-council-config-user-global.md) — the user-global
+  relocation this ADR sharpens to global-only.
+- `docs/contracts/ai-council-config.md` § File location — precedence rules.
+- `src/scripts/ai_council/config.ts` — `resolve_config_path`.
+- `src/skills/ai-council/SKILL.md`, `src/domains/meta/council/default/command.md`
+  — agent-facing surfaces carrying the "never infer not-configured from
+  missing project files" corollary.

package/docs/decisions/ADR-105-automatic-subagent-orchestration.md

@@ -0,0 +1,102 @@
+---
+adr: 105
+status: accepted
+date: 2026-06-23
+decision: automatic-subagent-orchestration
+supersedes: —
+superseded_by: —
+phase: auto-subagent-orchestration
+type: structural
+---
+
+# ADR-105 — Automatic subagent orchestration: activate-don't-fragment, prove before default-on
+
+## Status
+
+**Accepted** · 2026-06-23. Resolved by a deep AI-council round
+(claude-sonnet-4-5 + gpt-4o, design mode) plus host critical-evaluation, and
+implemented by `road-to-auto-subagent-orchestration.md`.
+
+## Context
+
+The package shipped three latent subagent surfaces — the
+`subagent-orchestration` skill (7 modes), `/orchestrate` (YAML pipelines), and
+the RDP `reasoning-orchestrator` ("dispatch by default") — all **opt-in**,
+none firing automatically. The goal: make subagent use automatic and
+settings-gated, so an orchestrating model delegates delegable sub-tasks to
+cheaper / faster / quota-separate subagents, each configured for its task,
+globally toggleable in `.agent-settings.yml`.
+
+The user's stated default was **on**. The council pushed back hard: a "default
+on" whose own design concedes it *may increase* token spend inverts risk in a
+package whose canon includes token-efficiency as an Iron Law, and whose
+multi-tool portability constraint means not every host even has a subagent
+primitive.
+
+## Decision
+
+1. **Activate, don't fragment.** Auto-dispatch is a settings + host-manifest
+   gate over the **existing** `subagent-orchestration` skill and RDP
+   "dispatch by default" — not a new parallel rule. Smallest surface.
+2. **Conservative-until-proven shipped default.** The full toggle ships now
+   (`subagents.enabled`, `subagents.auto: off|ask|on`), but the **shipped**
+   default is `ask` on hosts whose capability manifest reports
+   `subagent_spawn: true`, `off` elsewhere. The default flips toward `on`
+   (host-gated) **only** behind a real benchmark proving a net token-or-time
+   win at held quality (Phase 6). `on` is the destination, reached by
+   evidence — the reconciliation of the user's goal with the council's
+   prove-first.
+3. **Quota arbitrage is a runtime-detected bonus, never load-bearing.** The
+   "separate quota pool" (e.g. a subscription's distinct Sonnet allowance) is a
+   `host-capability-manifest` flag; routing is identical (minus the bonus)
+   where unsupported. No `.md` names a vendor's billing rule.
+4. **Deterministic classification v1.** Delegability is decided by enumerated
+   structural signals (frontmatter `parallelizable:`, ordered-plan,
+   independent-slices, a size floor); ambiguity defaults to ask/no-op, never
+   speculative spawn. No per-turn LLM meta-call in v1. LLM classification is
+   deferred to v2+, gated on the benchmark.
+5. **Verification budget preserves the floors.** Trivial/read-only sub-tasks
+   verify deterministically; non-trivial mutating sub-tasks get a full
+   cross-model judge (the `subagent-orchestration` Iron Law). A required
+   verification recorded as `none` is a surfaced safety gap.
+6. **Instrumentation first; kill-switch + surfaced guardrails.** Orchestration
+   telemetry rides additively on `audit-log-v1`. The kill-switch is a
+   single settings flip (`enabled: false` / `auto: off`). Rollback thresholds
+   (token blowup, spawn-failure, verify-skip, override rate) are surfaced from
+   the audit log — there is no automatic cohort-disable (a config package runs
+   no daemon).
+7. **Graceful host degradation.** Where the manifest reports
+   `subagent_spawn: false`, the whole layer is a clean no-op (single-agent),
+   surfaced explicitly. `.md` artefacts stay project-agnostic.
+
+## Consequences
+
+- Auto-dispatch is the default **where the host and settings allow it**, and a
+  no-op everywhere else. No safety floor is lifted.
+- The user's "default on" is an evidence-gated milestone, not assumed — the
+  one decision left for the user is to authorise/run the Phase-6 benchmark and
+  approve the default flip.
+- The implementation is mostly contexts + small pure helpers + a settings
+  block + two activated skills, keeping the diff minimal.
+
+## Alternatives considered
+
+- **Default `on` immediately** (the literal ask) — rejected: inverts risk
+  before measurement, against the frugality canon.
+- **New parallel auto-dispatch rule** — rejected: fragments a surface that
+  already exists; activation is the smaller change.
+- **Quota arbitrage as a core motivation** — rejected: vendor-specific,
+  fragile, violates portability; demoted to a runtime bonus.
+- **Council-strict default `off` forever** — rejected: forecloses the user's
+  goal; the benchmark gate is the middle path.
+
+## References
+
+- The auto-subagent-orchestration roadmap (now archived) + the
+  full council Convergence / Host verdict.
+- [`docs/contracts/auto-orchestration-v1.md`](../contracts/auto-orchestration-v1.md) — the manifest + spawn + telemetry contract.
+- Context docs under `src/agent-src/contexts/execution/`:
+  `host-capability-manifest`, `orchestration-telemetry`,
+  `auto-orchestration-activation`, `auto-dispatch-classification`,
+  `subagent-routing`, `subagent-spawn-contract`, `verify-budget`,
+  `subagent-steering`, `orchestration-benchmark-gate`.

package/docs/decisions/ADR-106-recursive-verification-benchmark-gate.md

@@ -0,0 +1,126 @@
+---
+adr: 106
+status: accepted
+date: 2026-06-23
+decision: recursive-verification-benchmark-gate
+supersedes: —
+superseded_by: —
+phase: recursive-verification · Phase 0
+type: structural
+---
+
+# ADR-106 — Recursive-verification benchmark gate: capability-OR-measured-novel-discipline, not capability-only
+
+## Status
+
+Accepted (2026-06-23). Locks the Phase 0 gate definition for
+[`road-to-recursive-verification.md`](../../agents/roadmaps/archive/road-to-recursive-verification.md)
+before any live `bench:ab` spend.
+
+## Context
+
+The `recursive-verification` skill (depth-bounded `attempt → critic verdict →
+re-attempt`, default `off`) needs a benchmark gate that decides whether its
+shipped default flips to `on`/`ask` for a given host. The roadmap's initial draft
+proposed **capability-axis-first**: ship only on a capability-axis lift; a
+discipline-axis-only lift does **not** pass, on the assumption that the package's
+existing always-on rules already deliver the discipline lift, so recursion would
+merely duplicate them at ~11× cost.
+
+This gate definition was put to the AI council (manual transport disabled; live
+`council run --depth deep`, 3 rounds; members **anthropic/claude-sonnet-4-5** +
+**openai/gpt-4o**, 2026-06-23, actual cost $0.0762). Both members independently
+converged that **capability-only is the wrong gate**, for a sound methodological
+reason; one member (anthropic) also raised a misreading that the host rejected.
+
+### What the council got right (accepted)
+
+- **The redundancy assumption is unmeasured.** The gate assumed recursion's
+  discipline lift is redundant with the existing rules *without measuring it*.
+  The fix is a **three-baseline** design and an explicit **marginal** lift:
+  - `D₀` = bare host (no rules, no recursion)
+  - `D₁` = rules only (always-on rules, no recursion)
+  - `D₂` = rules + recursion
+  - **novel discipline lift = `D₂ − D₁`** — recursion's contribution *over*
+    rules-only. Redundancy is now a measured fact, not an assumption.
+- **Cheap human-preference pre-test first.** Before any expensive benchmark, show
+  ~10 `(attempt₀, attempt_final)` pairs to humans and ask "which would you pay
+  for?". If preference < 60 %, the discipline lift is economically irrelevant →
+  capability-only gate stands and the expensive discipline arm is skipped (~$15
+  to settle the whole dispute). This mirrors the package's cheap-probe-first,
+  measure-first culture.
+- **Concrete cost ceiling, not "tolerable".** The original "at a tolerable cost"
+  is circular. Replace with a measurable ceiling (token multiplier / $-per-task).
+- **Per-(host, family) cell**, not host-global — recursion may pass for a weak
+  host on one family and fail for a strong host with no headroom.
+
+### What the council got wrong (rejected by the host)
+
+- anthropic framed the gate as a *"fatal flaw because the 0.333 → 1.000 lift is
+  already recursion."* This is a **misread** of `docs/benchmark.md`: that lift is
+  the *rules'* effect (vanilla → package), measured **without** recursion;
+  recursion's marginal lift is unmeasured (that is exactly what Phase 3 runs). The
+  valid extraction is the three-baseline design above, not "capability-first is
+  fatally wrong."
+- A weighted composite metric `w_cap·Δcap + w_disc·Δdisc` (openai) is **rejected**:
+  the weights are precisely the unknown the gate exists to discover (anthropic's
+  own counter, which the host endorses).
+
+## Decision
+
+The shipped default for `verification.recursive` flips to `on`/`ask` for a
+**(host, family) cell** if and only if:
+
+```
+capability_lift(host, family) significant       (McNemar on pass-rate, or scorer
+                                                  Wilcoxon p < 0.05, ≥6 discordant
+                                                  pairs, effect ≥ 0.5)
+OR
+( novel_discipline_lift = D₂ − D₁ > ε_disc       (recursion over rules-only —
+                                                  MEASURED, never assumed)
+  AND cost_per_task ≤ ceiling                     (concrete token-multiplier / $)
+  AND human_preference_rate > 0.6 )
+ELSE off.
+```
+
+**Cheap prerequisite (runs first, gates the rest):** the ~$15 human-preference
+test. Preference < 60 % ⇒ capability-only gate stands; skip the discipline arm.
+
+**The benchmark must report, per (host, family):** `pass_rate(D₀/D₂)`,
+`discipline_score(D₀ / D₁ / D₂)`, `cost(D₂)/cost(D₀)`, and
+`human_preference(recursion vs baseline)`. Without the `D₁` (rules-only) middle
+term, redundancy cannot be detected; without human-preference, a discipline lift
+cannot be validated as economically real.
+
+The first measured cell is `(claude-haiku-4-5, scope-discipline)`; multi-family
+expansion waits until data justifies it.
+
+## Consequences
+
+- Phase 3 of the roadmap gains a **rules-only (`D₁`) arm** and a **Phase 3a-pre
+  human-preference test**; the gate language changes from "capability-first;
+  discipline-only ⇒ STOP" to the two-branch rule above.
+- An honest-null is still first-class: capability flat **and** (human-preference
+  < 60 % OR `D₂ − D₁` ≤ ε OR cost over ceiling) ⇒ ship nothing, record the null.
+- The default stays `off` until a cell passes; no global flip without its own cell.
+- Slightly more measurement up front (three baselines + a human pre-test), bought
+  cheaply ($15 pre-test gates the expensive arm) — net a more honest gate.
+
+## Alternatives considered
+
+- **Capability-only (original draft).** Rejected: assumes, without measuring,
+  that recursion's discipline lift is redundant with the rules.
+- **Weighted composite metric.** Rejected: presupposes the weights the gate
+  exists to discover.
+- **Holistic / qualitative gate (openai round 1).** Rejected: too vague to
+  operationalise into a ship/no-ship decision.
+
+## References
+
+- [`road-to-recursive-verification.md`](../../agents/roadmaps/archive/road-to-recursive-verification.md)
+- [`docs/benchmark.md`](../benchmark.md) — the existing wrapper-lift A/B evidence.
+- Council round: anthropic/claude-sonnet-4-5 + openai/gpt-4o, 2026-06-23, deep
+  tier (3 rounds), $0.0762 actual.
+- [`subagent-orchestration`](../../src/skills/subagent-orchestration/SKILL.md),
+  [`verify-budget`](../../src/agent-src/contexts/execution/verify-budget.md),
+  `orchestration-benchmark-gate` (`gateVerdict` / `resolveShippedDefault`).