ur-agent 1.65.13 → 1.66.0

package/docs/VALIDATION.md

@@ -19,7 +19,7 @@ You need:
 
 ```sh
 ur --version
-# expected for this release: "1.65.13 (UR-Nexus)"
+# expected for this release: "1.66.0 (UR-Nexus)"
 ```
 
 ## 0.1 First-workspace model selection (1.45.4)

package/documentation/index.html

@@ -45,7 +45,7 @@
     <main id="content" class="content">
       <header class="topbar">
         <div>
-          <p class="eyebrow">Version 1.65.13</p>
+          <p class="eyebrow">Version 1.66.0</p>
           <h1>UR-Nexus Documentation</h1>
           <p class="lead">A practical, tutorial-style reference for installing, configuring, automating, extending, and operating UR-Nexus.</p>
         </div>

package/extensions/jetbrains-ur/build.gradle.kts

@@ -7,7 +7,7 @@ plugins {
 }
 
 group = "dev.urnexus"
-version = "1.65.13"
+version = "1.66.0"
 
 repositories {
     mavenCentral()

package/extensions/vscode-ur-inline-diffs/package.json

@@ -2,7 +2,7 @@
   "name": "ur-inline-diffs",
   "displayName": "UR Inline Diffs",
   "description": "Review, apply, and reject UR inline diff bundles from .ur/ide/diffs inside VS Code.",
-  "version": "1.65.13",
+  "version": "1.66.0",
   "publisher": "ur-nexus",
   "engines": {
     "vscode": "^1.92.0"

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ur-agent",
-  "version": "1.65.13",
+  "version": "1.66.0",
   "description": "UR-Nexus — autonomous engineering workflow engine (plan, execute, test, verify, document, benchmark, reproduce)",
   "type": "module",
   "packageManager": "bun@1.3.14",

package/technical/04-tools.md

@@ -74,12 +74,42 @@ constraints. Mutually independent tasks with no conflicting shared mutations
 can be delegated together, while dependent or conflicting work stays
 sequential.
 
+The proactive model lifecycle is explicit:
+`TaskCreate` → inspect successful result → `TaskUpdate(in_progress)` → inspect
+successful result → `Write`/`Edit`/mutating `Bash`/worker. Task setup and its
+dependent mutation are never one parallel batch. A feature-rich one-file build
+is non-trivial even if implementation uses one `Write`; classification follows
+the requested outcomes and verification burden, not the file or tool-call
+count. Approved-plan handoffs require task creation as their next
+state-changing action, and Ollama/Kimi receives the same ordered rule in its
+compact tool-discipline section. If earlier tasks are all terminal, the model
+must create a new cohesive outcome or reopen the relevant task before new
+workspace work.
+
+The runtime gate accepts an actionable `pending` or `in_progress` record; the
+stricter model-facing sequence keeps status truthful before work begins.
+`TodoWrite` is the equivalent single-call setup in legacy/headless pools, with
+the selected item already `in_progress`. Partial Task V2 exposure never masks
+an available `TodoWrite`. Bare/simple, REPL-simple, coordinator, custom-agent,
+and override-prompt paths retain a usable planner and capability-aware task
+contract, so the gate never instructs those modes to call a missing tool.
+If a user explicitly filters every planner from a custom tool pool, runtime
+fails closed and tells the user to enable Task V2/`TodoWrite` or explicitly
+disable the gate; it never tells the model to call a tool that is absent.
+
 Task IDs remain strings in storage and tool output. Model inputs for
 `TaskCreate` dependencies and `TaskGet`/`TaskUpdate` identifiers may also use a
 positive safe-integer JSON number; the tool boundary normalizes it to the
 canonical decimal string. Zero, negative, fractional, non-finite, Boolean, and
 precision-losing numeric IDs are rejected.
 
+The Kimi/Ollama compatibility parser applies the same task schema instead of a
+smaller parallel implementation. A clearly delimited bare `TaskCreate` object
+may contain `blocks`, `blockedBy`, `addBlocks`, or `addBlockedBy`; a bare
+`TaskUpdate` accepts the live terminal `failed` and `skipped` statuses as well
+as numeric IDs. IDs are normalized to canonical strings, while unknown fields,
+invalid IDs/statuses, ambiguous prose, and unavailable tools fail closed.
+
 Task-gate recovery names the tracking surface that is actually present:
 interactive Task V2 sessions use `TaskCreate`, while default headless sessions
 use `TodoWrite`. It never instructs a model to recover by calling a tool absent
@@ -92,6 +122,15 @@ redirection, backgrounding, sandbox overrides, and permission-time rewrites to
 mutating commands fail closed. The preview command remains a Bash side effect
 and still follows normal permission, sandbox, and plan-worker rules.
 
+Control-plane operations that establish or tear down tracking cannot depend on
+an already-actionable task: `TeamCreate`, `TeamDelete`, `TaskStop`/`KillShell`,
+and structured team shutdown/plan-response messages are narrow task-gate
+exceptions. Their own schemas, mode checks, active-member checks, and normal
+permissions remain authoritative. Loading a `Skill` and taking a desktop
+screenshot are read-only wrappers; downstream skill actions, desktop
+click/type, API/database/browser/MCP mutations, and future tools classified
+state-changing at runtime remain task-gated.
+
 Syntax verification has the same task-gate-only separation. A strictly parsed
 `node --check <single-file>` or the bounded HTML checker that reads one file,
 constructs but never invokes its first `<script>` body, and prints only a fixed

package/technical/06-configuration.md

@@ -105,6 +105,27 @@ Fan-out limits clamp rather than disable: out-of-range, negative and
 non-numeric values fall back to the default or the ceiling, so a settings file
 cannot switch the governor off.
 
+### Global app compaction controls
+
+`autoCompactEnabled` and `compaction.autoThreshold` are global app controls
+managed by `/config` or the `Config` tool and persisted in the global UR
+configuration (`~/.ur.json`, or `.config.json` in a legacy UR config home).
+They are not keys in the project `settings.json` schema.
+
+```text
+autoCompactEnabled = true
+compaction.autoThreshold = 80
+```
+
+The threshold accepts an inclusive percentage from 50 through 95. It is
+applied to the effective model context window and retains a 3,000-token manual
+compaction reserve. `UR_AUTOCOMPACT_PCT_OVERRIDE` has higher precedence for
+the current process; `DISABLE_AUTO_COMPACT` or `DISABLE_COMPACT` disables the
+proactive trigger. `/context` and the prompt notification report an
+approximate percentage left to this same resolved threshold; the active query
+may free additional context through bounded pruning immediately before the
+trigger.
+
 Rule syntax: `ToolName` (blanket) or `ToolName(specifier)` — e.g. `Bash(npm run *)`,
 `Edit(src/**)`, `mcp__server__tool`. Managed via `/permissions` UI as well.
 

package/technical/07-memory-and-context.md

@@ -104,12 +104,40 @@ rollback preserve a private copy of the full original before replacement.
 |---|---|
 | Visualize usage | `/context` (colored grid); `/files` is an ant-only command and is absent from the standard npm CLI |
 | Manual compaction | `/compact [focus instructions]` |
-| Auto-compaction | `src/services/compact` — triggers near the limit; `DISABLE_AUTO_COMPACT` env disables; PreCompact/PostCompact hooks fire |
+| Auto-compaction | `src/services/compact` — triggers at one canonical model-aware threshold; `DISABLE_AUTO_COMPACT` env disables; PreCompact/PostCompact hooks fire |
 | Context collapse | `src/services/contextCollapse` and `CtxInspect` are behind the compile-time `CONTEXT_COLLAPSE` feature, which the standard npm bundle does not include |
 | Micro-compaction | session-memory compact (`sessionMemoryCompact.ts`): force on with `ENABLE_UR_CODE_SM_COMPACT=1`, force off with `DISABLE_UR_CODE_SM_COMPACT=1`; otherwise both `tengu_session_memory` and `tengu_sm_compact` runtime gates must be on |
 | Clear | `/clear` (aliases `/reset`, `/new`) |
 | Read caps | Read tool truncates large files/lines; `/read`, `/analyze`, `/summarize` for deliberate loads |
 
+The effective context window always stays positive, including for small local
+models. By default proactive compaction reserves a scaled output buffer (up to
+20,000 tokens) plus a 13,000-token compact/manual safety band. The global
+`compaction.autoThreshold` setting can instead select 50–95% of the effective
+window while retaining a 3,000-token manual reserve;
+`UR_AUTOCOMPACT_PCT_OVERRIDE` is the explicit environment override. Trigger
+logic, prompt notifications, `/context`, and the SDK all use this same
+threshold and estimated live usage, including messages added after the last
+provider response.
+
+When proactive auto-compaction is active, the prompt notification and
+`/context` show a clamped approximate 0–100% value remaining until the actual
+trigger, with separate warning and final-error bands. The estimate uses the
+same live token counter, while the query may conservatively improve it by
+snipping or pruning immediately before the trigger. Reactive-compaction and
+context-collapse modes own their thresholds and therefore do not display a
+misleading proactive countdown. A successful compaction suppresses the stale
+pre-compact warning.
+
+Full, partial, and session-memory compaction restore the authoritative live
+task state after the boundary. Task V2 snapshots retain exact IDs, statuses,
+owners, and dependency edges; legacy TodoWrite snapshots retain order and
+status. The snapshot is bounded to 64 actionable-first records and an
+estimated 6,000-token budget. If more exist, the model is required to call
+`TaskList` before task mutation rather than guessing omitted IDs. Compact
+boundaries also permanently consume the task-gate's initial trivial-call
+allowance so compaction cannot reopen an implementation bypass.
+
 ## Repo wiki & map
 
 ```

package/technical/09-multi-agent.md

@@ -18,6 +18,13 @@ The main agent can spawn subagents. Built-in agent types
 | `Explore`, `Plan` | built-in read-only search and planning agents; registered in the standard npm bundle so plan-mode instructions never advertise missing worker types |
 
 Ordinary `Agent` subagents do not require experimental Teams/swarm mode.
+Every ordinary worker launch does require successful task setup and an
+`in_progress` parent task first; the exact built-in read-only `Explore`/`Plan`
+plan-mode path below is the only exception. Coordinator mode exposes Task V2
+tools (or `TodoWrite` in a legacy pool) instead of directing workers through an
+unsatisfiable gate. Independent tasks may launch in one worker wave only after
+their task records exist and the tasks actually launching are marked
+`in_progress`.
 Approved-plan handoff checks the actual tool pool, agent-type allowlist, live
 `Agent(type)` deny rules, and active built-in definitions. It can fan out
 independent ready tasks only when a selectable implementation worker remains.
@@ -33,6 +40,10 @@ Custom agents reusing those names, generic agents, teammates, background
 launches, custom working directories, and worktree launches remain mutating and
 task-gated. `TeamCreate` and `TeamDelete` also reject plan mode explicitly;
 team lifecycle state starts only after the plan is approved.
+Outside plan mode, team creation/deletion, structured shutdown responses, and
+emergency `TaskStop` are task-control transitions exempt from the workspace
+task gate so team bootstrap and teardown cannot deadlock. Their tool-specific
+validation still applies, including refusing deletion while members are active.
 
 In the standard bundle, `Explore` and `Plan` receive only `Glob`, `Grep`, and
 `Read`, use `dontAsk` permission mode, and have a second runtime boundary that
@@ -173,6 +184,12 @@ use legacy `TodoWrite` by default, or Task V2 when
   alone outside live plan mode are not exempt.
 - Configure the behavior at
   `tasks.requireBeforeChanges.{enabled,freeReads}`.
+- Full, partial, and session-memory compaction restore the live task board as
+  an authoritative bounded snapshot, including exact Task V2 IDs, statuses,
+  owners, and dependency edges. In-process workers do not run a second private
+  compaction algorithm: they use the same `runAgent` query-loop policy,
+  feature gates, session-memory-first behavior, threshold, circuit breaker,
+  and compact-boundary handling as the main agent.
 
 The prompt contract, plan-file structure, plan-agent output, approval handoff,
 task-tool result, and gate recovery text all reinforce the same decomposition

package/technical/README.md

@@ -1,6 +1,6 @@
 # UR-Nexus — Technical Specifications
 
-> Audited against the executable source and tests for `ur-agent` v1.65.13.
+> Audited against the executable source and tests for `ur-agent` v1.66.0.
 > Command, tool, flag, provider, and setting claims are checked against the
 > implementation rather than copied from product prose. Release validation
 > keeps this version synchronized and packages the complete `technical/`