@haaaiawd/loom 2.0.1 → 2.1.0

package/design.md

@@ -137,9 +137,10 @@ The CLI has six responsibilities:
 1. Initialize the minimal project backbone.
 2. Record structured understanding and decision history safely.
 3. Scaffold a scalable design-document graph and separate professional-field dossiers.
-4. Maintain a large Work Map and a single active Task.
-5. Compile only decision-relevant context after reset.
-6. Prepare one-time Keeper and controlled Evil Eval artifacts.
+4. Maintain a large Work Map and a single active Task.
+5. Compile only decision-relevant context after reset.
+6. Adapt clarification to whether a human response channel is actually available.
+7. Prepare one-time Keeper and controlled Evil Eval artifacts.
 
 It does not orchestrate models, browse the web, generate project documents, or pretend that a CLI command
 clears model memory. Those capabilities belong to the host Agent and its tools.
@@ -164,7 +165,7 @@ generic updates cannot change status, and completion requires criterion-by-crite
 
 ## Context selection
 
-`loom context` always injects:
+`loom context` always injects:
 
 - the stable Agent protocol;
 - compact state counts and open uncertainty;
@@ -172,6 +173,20 @@ generic updates cannot change status, and completion requires criterion-by-crite
 - the active Task, if one exists;
 - the exact design, capability, source, contract, or fixture files named in that Task's `reads` list.
 
+### Environment adaptation and isolated state
+
+The host supplies human-channel availability at context time. With `available` (the default), the Agent may
+ask the human for consequential intent, preference, authority, or facts that only they hold. With
+`unavailable`, it must not fabricate a user exchange or replace an intent/authority question with web search.
+It first inspects discoverable workspace facts and permitted tools, researches only objective external facts when
+the task allows it, then records a bounded assumption and chooses a safe reversible action or blocks on
+unavailable authority.
+
+`--state-dir <outside-workspace-dir>` places LOOM's state tree in a per-run sidecar rather than the workspace.
+This is for benchmark or sandbox runners whose scored workspace must remain pristine. The sidecar remains
+semantically `.loom/` to LOOM documents and Task references, but initialization does not edit the workspace
+or add an AGENTS.md anchor. Every command in that run must receive the same state directory.
+
 When a Task is active, context also injects a short execution protocol. It tells a reset Agent to reconcile
 the Task with the current workspace and version-control state, inspect relevant tests before editing, choose
 verification according to risk and the exact `done_when` claims, persist `completed/current/next` at meaningful

package/docs/PROMPT_CATALOG.md

@@ -12,17 +12,23 @@ order as JSON, so documentation cannot quietly describe a different prompt from
 
 ## Instruction layers
 
-| Layer | Message | Injected when | Purpose |
-|---|---|---|---|
+| Layer | Message | Injected when | Purpose |
+|---|---|---|---|
 | Stable core | `layers.stable_core` | Every `loom context` | Durable collaboration judgment: user ownership, evidence classes, question threshold, corrections, reversible delegation |
+| Environment adaptation | `layers.environment_adaptation_available` or `layers.environment_adaptation_unavailable` | Every `loom context` | States whether a human can answer; routes uncertainty through inspection, permitted factual research, clarification, bounded assumption, or block |
 | Runtime adaptation | `layers.runtime_protocol` | Every `loom context` | LOOM-specific disk model, document graph, capability boundaries, Work Map, and three production loops |
 | Project state | `layers.project_state` | Every `loom context` | Dynamic status, uncertainty, assumptions, document/capability inventory, Task counts, and Keeper feedback |
 | Execution protocol | `layers.execution_protocol` | Active Task or `--task` | Recovery, workspace inspection, risk-based testing, progress persistence, exact proof, and delivery boundaries |
 | Current task | `layers.current_task` | Active Task or `--task` | Exact Task JSON followed by the exact contents of every path in `reads` |
 
 This separation is deliberate. Project-specific expertise does not inflate the stable prompt; it lives in
-field dossiers. One-off Task instructions do not become global doctrine. Host permissions and tools are not
-claimed by prompt text.
+field dossiers. One-off Task instructions do not become global doctrine. Host permissions and tools are not
+claimed by prompt text.
+
+When the human channel is unavailable, the environment message explicitly forbids fabricated user interaction and
+forbids treating web research as a substitute for user intent, preference, or authority. It preserves factual
+research when the task and host actually permit it, then requires a reversible assumption or block for what cannot
+be learned.
 
 ## Workspace and document messages
 
@@ -71,8 +77,12 @@ new digest, and another fresh Keeper.
 `evaluation.baseline_condition` gives a normal capable Agent the identical brief and ordinary tools without
 LOOM. It does not weaken the baseline or prohibit normal planning.
 
-`evaluation.loom_condition` gives the same brief, model class, tools, workspace facts, and budget, adding only
-LOOM continuity infrastructure and no extra authority.
+`evaluation.loom_condition` gives the same brief, model class, tools, workspace facts, and budget, adding only
+LOOM continuity infrastructure and no extra authority.
+
+`evaluation.unattended_baseline_condition` and `evaluation.unattended_loom_condition` are the same two
+conditions when no user response channel exists. The environment fact is equal in both arms; only the LOOM arm
+uses `loom context --human-channel unavailable`.
 
 `evaluation.blind_judge` scores anonymized output on intent fidelity, question value, whole coverage,
 professional depth, buildability, reset continuity, implementation evidence, human burden, and cost. It

package/docs/UX_FLOW.md

@@ -56,8 +56,17 @@ details that do not affect the current whole, or framework questions whose only
 | `I don't know` | Research or decide when safe; ask again only if the human owns the missing preference or authority |
 | `Skip questions and continue` | Name the concrete quality or risk loss, record skipped uncertainty, then proceed |
 
-After every answer round, the Agent re-evaluates the whole. If material uncertainty remains, it loops to another adaptive
-round. If only reversible local uncertainty remains, it continues.
+After every answer round, the Agent re-evaluates the whole. If material uncertainty remains, it loops to another adaptive
+round. If only reversible local uncertainty remains, it continues.
+
+### When no human response channel exists
+
+Some unattended, sandboxed, or benchmark runs have no human to answer. The host invokes
+`loom context --human-channel unavailable`; this does not create a synthetic user. The Agent first inspects
+discoverable workspace facts and permitted tools, researches only objective external facts when the task permits it,
+and never treats web research as a substitute for user intent, preference, or authority. It records a bounded
+assumption and selects a safe reversible action, or blocks if the missing answer controls irreversible, high-risk,
+or materially costly work.
 
 ## 3. Build project-specific capability
 
@@ -163,9 +172,13 @@ human authority govern later changes.
 
 ## Primary evaluation question
 
-The causal baseline is a normal capable Agent with no LOOM—not LOOM v1. Both conditions receive the same raw brief, model,
-tools, workspace, user oracle, budget, and reset schedule. LOOM v1 may be added as a secondary regression condition, but it does
-not answer whether the framework itself creates value over no framework.
+The causal baseline is a normal capable Agent with no LOOM—not LOOM v1. Both conditions receive the same raw brief, model,
+tools, workspace, human-channel availability, user oracle when one exists, budget, and reset schedule. LOOM v1 may be added as a secondary regression condition, but it does
+not answer whether the framework itself creates value over no framework.
+
+For an unattended external benchmark, both conditions receive `human_channel: unavailable` and no fabricated user
+oracle. A fresh LOOM sidecar sits outside the scored workspace and is passed through `--state-dir` on every LOOM
+command, so framework state cannot alter task files or hidden-test behavior.
 
 The anonymization harness must preserve runnable relative layout, record a file/digest manifest, and rerun each condition's
 declared tests and smoke command before judging. Packet-construction failures are eval failures, never evidence against a run.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@haaaiawd/loom",
-  "version": "2.0.1",
+  "version": "2.1.0",
   "description": "LOOM — invisible project understanding and continuity infrastructure for AI agents",
   "type": "module",
   "bin": {