@lambdacurry/arbor 0.21.36 → 0.21.38

package/dist/arbor.js

@@ -19081,7 +19081,7 @@ ${ARBOR_FEEDBACK_GUIDANCE}
 
 1. ORIENT, CONFIGURE, AND AUTHORIZE. Start with get_computer(threadId) to inspect the effective inherited recipe, repository authorization, compatibility, and durable lineage. Use set_organization_computer, set_space_computer, set_topic_computer, or set_thread_computer only when the corresponding complete sparse layer must be replaced; use github_repositories then github_connect_repository to authorize repositories at the narrowest scope that should inherit them. If github_repositories looks stale, github_refresh_installation reconciles the cached inventory from GitHub. attach_computer mints durable protocol identity without starting a runtime; computer_open attaches as needed and starts or resumes the project environment. Each returned repository entry includes verified root instruction pointers when that repo has an AGENTS.md: read surfaced repository instructions before substantial edits or verification, then check for a nearer nested AGENTS.md once the target path is known. Repo-local instructions/scripts/config beat guessed formatter, test, lint, or build commands.
 2. READ THE CAPABILITIES CARD, NOT THE BACKEND. computer_open and computer_status describe the live PARENT Thread Computer and return the exact computerSessionId for parent filesystem, command, checkpoint, Preview, bundle, and stop operations. computer_status does not establish isolated Run activity; use computer_run_receipt(runId) and computer_process_read for Run-owned work. Command-line utilities, including Git and package managers, run through computer_exec; check unfamiliar binaries with command -v.
-3. ONE RUN PER TASK. A write Run defaults to an isolated execution environment with its own filesystem, processes, ports, and /tmp while preserving /workspace paths; a read Run shares the parent. Pass runId to Run-aware tools. If a response is lost, recover the durable operation/process identity from computer_run_receipt and read it before retrying. Once every operation is settled, computer_run_suspend publishes the exact workspace and releases the provider during an external wait; the next Run-aware operation restores that checkpoint under the same runId automatically. Suspension refuses active processes or unsettled outcomes and never replays unknown work, so reconcile process and Git/publication evidence first. An isolated local commit or Preview is not Git publication; push coherent work and prove it with computer_repo_work_status before computer_run_finish destroys the final placement. discard=true with outcome failed or cancelled is the explicit choice to abandon unknown or proven non-durable work without replaying it.
+3. ONE RUN PER TASK. A write Run defaults to an isolated execution environment with its own filesystem, processes, ports, and /tmp while preserving /workspace paths; a read Run shares the parent. Pass runId to Run-aware tools. If a response is lost, recover the durable operation/process identity from computer_run_receipt and read it before retrying. computer_process_read and computer_process_terminate settle a provider-absent current-generation process from exact evidence; computer_run_suspend performs that bounded settlement internally and never replays unknown work. A still-active process returns a retryable busy state; an unknown current-generation mutation fail-closes. An isolated local commit or Preview is not Git publication; push coherent work and prove it with computer_repo_work_status before computer_run_finish destroys the final placement. discard=true with outcome failed or cancelled is the explicit choice to abandon unknown or proven non-durable work without replaying it.
 4. CHECKPOINT COMPLETE PARENT UNITS. computer_checkpoint advances immutable parent lineage after an intermediate complete unit; computer_stop performs the final checkpoint before teardown. An isolated Run Preview is durable review output, not adoption or a checkpoint. The internal provider-handle publication operations are not agent tools.
 5. GIT AND GITHUB ARE ORDINARY COMMANDS AFTER AUTHORIZATION. Clone, fetch, pull, push, and gh run through computer_exec using a connected human's just-in-time ambient GitHub identity. Credentials never enter argv, workspace files, Git config, checkpoints, Events, logs, or receipts. A restored Computer deliberately keeps the checkout age of its snapshot, so fetch/pull before trusting remote state.
 6. VERIFY, PREVIEW, AND EXPORT. computer_verify captures a public URL directly without opening a project runtime. computer_publish_preview preserves immutable static output from either the parent or an active isolated write Run. computer_export moves bounded files into durable Thread Attachments; for a computer:// file Artifact asleep face, export the exact locator path .tldr and a same-Thread PNG, then use artifact set_snapshot. computer_export itself does not call set_snapshot; place returned Markdown exactly where the media belongs and cite the matching attachment ids.
@@ -19778,7 +19778,7 @@ var ACTION_DEFINITIONS = [
   {
     name: "computer_run_suspend",
     title: "Suspend a run",
-    description: "Checkpoint an active isolated WRITE Run's exact workspace, publish that Run-owned restore point, and release its provider Sandbox during an external wait; the same runId resumes automatically on its next operation. Refuses active processes or unsettled mutations and never replays unknown work.",
+    description: "Checkpoint an active isolated WRITE Run's exact workspace, publish that restore point, and release its Sandbox during an external wait; the same runId resumes on the next operation. Performs bounded process read/terminate/settlement internally: suspends when settled, returns retryable busy if a process is active, or fail-closes on unknown mutation and never replays unknown work.",
     inputSchema: {
       runId: exports_external.string().describe("the active isolated WRITE Run to suspend, run_…")
     },
@@ -19857,7 +19857,7 @@ var ACTION_DEFINITIONS = [
   {
     name: "computer_process_read",
     title: "Read a background process",
-    description: "Read one exact provider-owned process from computer_exec, including a foreground operation recovered from a Run receipt, and use nextCursor for incremental output. Returns the same processId plus compact activity, terminal result, or durable not-routed proof; only that proof makes a fresh computer_exec safe.",
+    description: "Read one exact provider-owned process from computer_exec, including a recovered Run-receipt foreground operation, and use nextCursor for incremental output. Returns processId plus activity, terminal or not-routed proof; Arbor settles the durable operation when that process is already absent, and a lost session after open retries computer_open.",
     inputSchema: {
       computerSessionId: exports_external.string().optional().describe("the computerSessionId returned by computer_open, cms_…; required unless runId is supplied"),
       runId: exports_external.string().optional().describe("the Run that owns this process, if it was started inside a Run"),
@@ -19872,7 +19872,7 @@ var ACTION_DEFINITIONS = [
   {
     name: "computer_process_terminate",
     title: "Terminate a background process",
-    description: "Terminate a background process started by computer_exec when it should no longer run. Stops the provider-owned process tree and returns the observed terminal state without requiring shell PIDs or ps/kill.",
+    description: "Terminate a background process started by computer_exec when it should no longer run, stopping the provider-owned process tree without shell PIDs or ps/kill. Repeating terminate for an already-absent or already-complete process is a no-op that settles that exact durable operation from the available evidence.",
     inputSchema: {
       computerSessionId: exports_external.string().optional().describe("the computerSessionId returned by computer_open, cms_…; required unless runId is supplied"),
       runId: exports_external.string().optional().describe("the Run that owns this process, if it was started inside a Run"),
@@ -19885,7 +19885,7 @@ var ACTION_DEFINITIONS = [
   {
     name: "computer_status",
     title: "Inspect a computer",
-    description: "Inspect the live PARENT Thread Computer's capabilities, execution readiness, processes, and profile. If its named command session is unavailable, command capabilities disappear and recovery points to computer_open; this does not report isolated Run command state, so use computer_run_receipt for Run state or get_computer for durable lineage.",
+    description: "Inspect the live PARENT Thread Computer's capabilities, execution readiness, processes, and profile. Transient session unavailability drops command capabilities and may retry computer_open once; a terminal unknown mutation is a non-retryable stop — this does not report isolated Run command state, so use computer_run_receipt for Run state or get_computer for durable lineage.",
     inputSchema: {
       computerSessionId: exports_external.string().describe("the computerSessionId returned by computer_open, cms_…")
     },

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lambdacurry/arbor",
-  "version": "0.21.36",
+  "version": "0.21.38",
   "description": "The Arbor CLI — a shared workspace for people and agents. The human + headless-agent write path over Arbor's guarded operation surface.",
   "keywords": [
     "agents",