@deftai/directive-content 0.96.0 → 0.98.0

package/docs/test-boundary.md

@@ -0,0 +1,43 @@
+# Test / source boundary (`verify:test-boundary`)
+
+Refs: #3145 · Related: #1310 forward-coverage, testing layout guidance
+
+## Problem
+
+Directive accepted test harnesses, fixtures, and smoke orchestration under production-owned roots (`src/**`, `infra/**`, `Tools/**`, …) while gates stayed green. Placement guidance was prose-only.
+
+## Contract
+
+`task verify:test-boundary` / `deft verify:test-boundary` enforces a typed policy:
+
+| Field | Meaning |
+| --- | --- |
+| `sourceRoots` | Production-owned path globs |
+| `testRoots` | Allowed test roots |
+| `fixtureRoots` | Fixture roots |
+| `testFilePatterns` | Conventional test basenames (`test_*.py`, `*Tests.cs`, `*.test.ts`, `*.spec.ts`, …) |
+| `productionMayReferenceTestRoots` | Default `false` — production must not reference test/fixture roots |
+| `allow` | Narrow exceptions (`kind: exception` or `production-liveness`) |
+| `enforcementMode` | `warn` (migration/discovery) or `enforce` |
+
+## Policy sources (first wins)
+
+1. `--policy <path>`
+2. `.deft/test-boundary.policy.json`
+3. `plan.policy.testBoundary` in `xbrief/PROJECT-DEFINITION.xbrief.json`
+4. **Defaults** (conventional roots + patterns, `enforcementMode: warn`)
+
+## Migration path
+
+1. Run with defaults (warn-only): `task verify:test-boundary`
+2. Review findings; move test artifacts under declared test roots or classify production liveness/canaries in `allow`
+3. Persist reviewed policy under `.deft/test-boundary.policy.json` or `plan.policy.testBoundary` with `enforcementMode: "enforce"`
+4. Wire stays green via `task check` / consumer deposit
+
+## Production liveness carve-out
+
+Health probes, canaries, and operational evidence collectors may live under production roots when listed in `allow` with `kind: "production-liveness"` and a recorded reason. Do not use the word “smoke” alone as classification.
+
+## Remediation
+
+Failures name the path, violated boundary, and next step (move under test root, allow entry, or reclassify). See `content/contracts/test-boundary.md`.

package/main.md

@@ -58,6 +58,22 @@ Legend (from RFC2119): !=MUST, ~=SHOULD, ≉=SHOULD NOT, ⊗=MUST NOT, ?=MAY.
 
 See #634, #642. See [ADR-001](./docs/decisions/ADR-001.md) for the token-economics rationale behind this ordering (vBRIEF-as-canonical for the agentic-consumed surface).
 
+## Self-Improving, Not Self-Editing (#3164)
+
+**Stance:** Directive is **self-improving**, not **self-editing**. Improvement goes through formal gates (**propose-not-apply**).
+
+Continual-Harness-class hosts may rewrite prompts, skills, and memory mid-run. Directive does not. A running session must not mutate live operating rules in place.
+
+- ! Directive MUST NOT self-edit live operating rules mid-run (managed AGENTS.md, pinned skills, policy flags, and other constitution-tier content)
+- ! Refine and meta-loops **propose** changes; issues, PRs, and quality gates **dispose**
+- ! Learn between merges — not by mid-session rewrite of the constitution
+- ? Prose lessons (`meta/lessons.md`; Continuous Improvement below) MAY stay agent-writable. They sit at the bottom of the Rule Authority ladder and cannot override structural rules
+- ⊗ Treat mid-run self-edit of constitution, skills, or policy as the default learning model
+
+This is not timidity; it is identity. It follows from the Rule Authority ladder above and from safety via formal gates, not alignment (#1200). A framework whose value is that the gates sit outside the agent cannot let a session rewrite the gates' substrate and stay coherent.
+
+Parent epic: #3179 (self-improving under gates). Trajectory / refine constraint: #2741 — refine proposes; gates dispose. Proposer runtime (SkillOpt / skill-variant) is tracked on #2436 / #1307 and is out of scope for this stance naming.
+
 **Decision Making:**
 - ! Follow established patterns in current context
 - ~ Question assumptions and probe for clarity
@@ -71,6 +87,35 @@ See #634, #642. See [ADR-001](./docs/decisions/ADR-001.md) for the token-economi
 - ⊗ Continue executing a skill past its explicit instruction boundary — when a skill's steps are complete, stop and return to the calling context; do not drift into adjacent work (#198)
 - ! The end of a skill's final step is an exit condition — do not continue into adjacent work, even if it seems related or trivial
 
+## Dual Stop Rule (#2442)
+
+Loop engineering requires **two** stop conditions on multi-iteration autonomous work: a **success stop** (goal / AC / checker met) and a **failure or budget stop** (retries exhausted, no progress, or time/token budget). Directive already has strong success-shaped gates (`task check`, acceptance criteria, STOP on plan precondition mismatch -- #1613). This section requires the complementary failure envelope so agents escalate instead of thrashing forever.
+
+**Applies to:** multi-iteration autonomous loops -- build quality / implement-fix loops, pre-PR polish cycles, swarm repair and monitor loops, research fan-out, review fix cycles, and similar retrying work.
+
+**Does not apply to:** single-turn tasks (one shot answer, one file edit, one status probe). Not every task is a loop; do not invent iteration caps where there is no multi-step retry envelope.
+
+**Required stops on every multi-iteration loop:**
+
+1. ! **Success stop** -- goal, acceptance criteria, or checker is met; exit the loop and continue the skill or report done.
+2. ! **Failure stop** -- at least one of:
+   - **max iterations** (task-class default; e.g. a short quality-fix class vs a longer research class)
+   - **no-progress** (same outcome or same failure fingerprint N times in a row with no material change)
+   - **explicit budget** (time, tool-call, or token budget when the host exposes it)
+
+**On failure stop:**
+
+- ! Halt the loop. Do not silently continue, re-dispatch, or open a new identical attempt without an operator decision.
+- ! Emit an **operator-visible halt report** that states: (a) what was tried, (b) what is still missing or failing, (c) what human decision is needed next (scope change, unblock, override, or abandon).
+- ⊗ Keep iterating after the failure envelope is exhausted because "one more try" might work.
+- ⊗ Reset iteration counters solely by creating a new revision, swapping workers, or compacting context when the same failure class remains.
+
+**Relation to other rules:**
+
+- #1613 covers STOP when plan **preconditions** fail (reality mismatch). Dual stop covers the case where the plan is still "valid" but the agent must quit after N failed attempts, N identical no-progress outcomes, or a budget limit.
+- Skills name concrete defaults: `skills/deft-directive-build/SKILL.md` (implement / pre-PR loops), `skills/deft-directive-swarm/SKILL.md` and its Phase 4 / core-ops references (repair / monitor loops).
+- **Delivery / acceptance mechanical enforcement** (durable attempt ledger, material-progress circuit breaker, cross-revision budgets) is **#3143** — library: `packages/core/src/delivery-attempt/` (`evaluatePreDispatch`, unit ledger under `.deft/delivery-attempts/`). Docs: `content/docs/delivery-attempt.md`. #2442 is the principle + skill defaults; #3143 is the deterministic pre-dispatch gate. Route delivery/acceptance loops through that surface rather than inventing a parallel ledger.
+
 **Adaptive Teaching:**
 - ~ When a recommendation is accepted without question, be concise
 - ! When a recommendation is questioned or overridden, explain the reasoning
@@ -248,6 +293,7 @@ See [`skills/deft-directive-refinement/SKILL.md`](./content/skills/deft-directiv
 - ~ When repeated correction or better approach found, codify in `./lessons.md`
 - ? Modify `./lessons.md` without prior approval
 - ~ When using codified instruction, inform user which rule was applied
+- ! Promote constitution-tier improvements (skills, policy, managed AGENTS rules) through issue / PR / quality gate — not mid-run self-edit (see [Self-Improving, Not Self-Editing (#3164)](#self-improving-not-self-editing-3164))
 
 **Observation:**
 - ~ Think beyond immediate task

package/meta/philosophy.md

@@ -8,6 +8,16 @@ Legend (from RFC2119): !=MUST, ~=SHOULD, ≉=SHOULD NOT, ⊗=MUST NOT, ?=MAY.
 
 ---
 
+## Self-Improving, Not Self-Editing (#3164)
+
+Directive improves **across merges** through issues, PRs, and quality gates. It does not rewrite live constitution rules mid-run (**propose-not-apply**).
+
+Constitution (managed AGENTS.md, pinned skills, policy) stays gated. Playbook-tier prose (e.g. lessons) may stay agent-writable because it sits at the bottom of the Rule Authority ladder.
+
+Full stance and MUST/MAY bullets: [main.md § Self-Improving, Not Self-Editing](../../main.md#self-improving-not-self-editing-3164). Parent epic #3179; safety-via-gates #1200; trajectory/refine #2741.
+
+---
+
 ## Deterministic > Probabilistic
 
 Prefer deterministic components for repeatable actions over probabilistic ones.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@deftai/directive-content",
-  "version": "0.96.0",
+  "version": "0.98.0",
   "description": "Shippable Directive framework content in the consumer .deft/core/ layout (C1 flatten), plus the engine surfaces (.githooks/, Taskfile.yml, tasks/) the deposit wires. Python-free per #2022 Phase 3. Refs #11, #1669, #1967.",
   "license": "MIT",
   "type": "module",

package/packs/patterns/patterns-pack-0.1.json

@@ -35,6 +35,16 @@
       "path": "patterns/llm-app.md",
       "body": null
     },
+    {
+      "id": "operator-log-hygiene",
+      "title": "Operator-log hygiene (#1940)",
+      "description": "Consumer guidance for structured, operator-facing logs: six failure modes, anti-patterns, positive rules, explicit non-goals (no core schema, no default-on, not Insights/LLM telemetry). SLizard is external reference only. Checklist and optional consumer pack stub under docs/.",
+      "triggers": [
+        "operator-log-hygiene"
+      ],
+      "path": "patterns/operator-log-hygiene.md",
+      "body": null
+    },
     {
       "id": "multi-agent",
       "title": "Multi-agent identity separation pattern (#983)",