ur-agent 1.65.5 → 1.65.7

package/CHANGELOG.md

@@ -1,5 +1,117 @@
 # Changelog
 
+## 1.65.7
+
+- Audited all 14 technical chapters against the shipped source and generated
+  CLI. The manual now distinguishes public runtime behavior from compile-time
+  flags, compatibility-only state, and source helpers; it is included in the
+  npm package and checked for valid paths, commands, configuration, and package
+  coverage.
+- Reworked agent guidance into one compact, ordered execution contract plus
+  mode-specific code, research, debug, browser, image, video, and data
+  workflows. Structured tool calls, dependency-aware planning, bounded
+  parallelism, failure recovery, result inspection, verification, honest
+  completion, and prompt-injection boundaries remain explicit without unsafe
+  package-manager guesses or examples that merely narrate fake tool calls.
+- Hardened task execution end to end. Both Task V2 and legacy TodoWrite plans
+  participate in mutation and completion gates, task order is numeric and
+  dependency-aware, failed/skipped states are rendered honestly, and the
+  verifier refuses an overall completion while actionable work remains or task
+  state cannot be read.
+- Workflow execution now checks approval before invoking a side effect,
+  enforces verification unless explicitly advisory, records every settled
+  parallel branch, and persists progress atomically. Resume restores bounded
+  exact dependency outputs without replaying completed steps; missing or
+  oversized required output fails closed and explains how to recover.
+- Multi-agent crews and planned `/exec` runs use bounded parallel workers,
+  dependency fan-in, exact verdicts, cancellation-aware respawn, isolated retry
+  worktrees, and a shared-checkout write gate. Ambiguous side effects are not
+  replayed, per-prompt worktrees keep dependent steps together, and concurrent
+  output files cannot overwrite one another.
+- Fixed false-success and argument-parsing behavior across root and slash
+  commands. Invalid usage returns 2, failed operations return 1, child process
+  failures propagate, previews do not mutate, and claim-ledger persistence is
+  structurally validated, workspace-contained, private, atomic, and bounded.
+- Strengthened provider and tool-call interoperability for smaller/local
+  models: streamed calls keep stable ordinals, cumulative fragments are
+  deduplicated conservatively, malformed or rewritten inputs are revalidated,
+  Ollama image results and session-host overrides are preserved, and OpenAI
+  Responses transport settings now reach actual client selection.
+- Improved terminal task UX: live task/status displays use the selected
+  session model, show only actionable or active background work, preserve
+  selection as rows change, handle narrow terminals, expose blockers and every
+  terminal status, and never advertise blocked work as the next action.
+- Hardened session, export, research, notes, file-operation, worktree, and
+  sandbox paths against traversal, symlink escape, oversized/corrupt state,
+  interrupted writes, stale replay, and misleading success. Unsupported
+  workflow/monitor task implementations remain inspectable as history but are
+  no longer presented as runnable.
+- Added a typed `ur-agent/sdk` subpath for ESM and CommonJS with validated
+  subprocess inputs, deterministic JSON/NDJSON result parsing, environment
+  precedence, and nonzero-exit handling. Release gates rebuild generated
+  artifacts before publish, smoke-test both SDK module formats from the packed
+  tarball, require the technical manual, and verify runtime dependency ranges.
+
+## 1.65.6
+
+- Remote sessions no longer reconnect forever when the server repeatedly
+  accepts a WebSocket upgrade and then reports that the session is missing.
+  Repeated `connect()` calls reuse the live client, malformed permission and
+  cancellation frames fail safely, and the cancellation wire type now matches
+  the runtime schema.
+- Tool execution can no longer silently run zero tools because of a negative
+  concurrency setting. Concurrency is normalized to a bounded range, and
+  serial or parallel tool cancellation always clears in-progress state.
+- Replaced scattered and sometimes unsafe model instructions with one compact,
+  ordered execution contract shared by normal, minimal, and remote modes. It
+  preserves structured tool use, dependency-aware parallel calls, result
+  observation, recovery, verification, honest completion, and prompt-injection
+  boundaries while removing guesses such as forcing system Python package
+  installs. Ollama retains an explicit native-tool discipline and may use text
+  recovery only when the runtime says native tools are unavailable.
+- Tool calls are normalized and validated again at the final execution
+  boundary, including permission-hook rewrites. Unknown tools, duplicate IDs,
+  malformed object inputs, and unsafe read-to-write rewrites fail closed.
+  Ollama/Kimi streaming repairs supported split calls without turning invalid
+  arguments into empty objects, deduplicates cumulative frames, and preserves
+  distinct same-name calls.
+- Task planning now uses dependency-ordered lifecycle guidance, atomic
+  task-store updates, transactional dependency checks, and strict gate reads.
+  Mutations are classified at execution time, delegation cannot consume the
+  short-task allowance, and completion claims are rejected while actionable
+  tasks remain or task state cannot be verified.
+- Repeated failing calls now use bounded SHA-256 identities with per-query
+  cleanup, TTL and capacity limits, so large or sensitive tool inputs are not
+  retained. Identical failures are refused and eventually abort the turn;
+  corrected inputs are allowed immediately.
+- `ur crew` now runs independent workers in parallel, honors dependency
+  fan-in, claims work across processes, and supports bounded respawns with
+  backoff and cancellation. Automatic retries use isolated fresh worktrees;
+  ambiguous shared-workspace mutations are never replayed.
+- Hardened the upstream proxy relay against unterminated or overflowing
+  protobuf varints, half-open rejected HTTP clients, and synchronous WebSocket
+  constructor failures.
+- Terminal UI layouts now remain valid in very small windows. Search pickers,
+  fullscreen modals, dividers, highlighted code, the startup logo, and cursor
+  widths clamp or reflow instead of overflowing; empty picker focus,
+  Meta-key punctuation/input-buffer handling, prompt-footer sizing, live-model
+  status refreshes, and task progress presentation are corrected.
+- Hardened IDE state storage against symlink redirection, oversized or corrupt
+  state, and interrupted writes. The VS Code chat now treats an abnormal
+  process exit as failure even after a result frame, bounds NDJSON/stderr
+  buffering, validates webview messages, and improves keyboard focus, live
+  status, alert, and attachment accessibility.
+- The Node launcher now enforces the declared Bun minimum, including
+  prerelease ordering, rather than checking only whether a `bun` executable
+  exists.
+- Release automation now runs the validated serial Bun test command (parallel
+  workers currently collide on ephemeral test servers), grants the permissions
+  needed to download verified artifacts, keeps
+  prereleases off npm's `latest` tag, marks GitHub prereleases correctly,
+  validates every version-bearing surface, supports safe prerelease bumps, and
+  cleans package-check temporary state. Validation docs and historical
+  changelog ordering were corrected.
+
 ## 1.65.5
 
 - Syntax highlighting works again — in assistant messages, code previews and
@@ -1658,36 +1770,6 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
   `FileEditTool.call`.
 - Verified source and production bundle release checks for the `1.22.6` build.
 
-## 1.22.3
-
-### Added
-- **Executable skill directories.** A `.ur/skills/<name>/` directory containing
-  `skill.yaml` is now an executable skill that compiles into a `WorkflowSpec`.
-  Supports `instructions.md`, `scripts/`, `templates/`, and `checklists/`.
-  `src/skills/skillSpec.ts` parses, validates, and compiles skills; step prompts
-  support `$ARGUMENTS`, `$0..$N`, and `$ARGUMENTS[N]` substitution.
-- **`ur skill` CLI.** New command surface: `ur skill list`, `ur skill show <name>`,
-  `ur skill run <name> [args]`, and `ur skill init <name>`. Registered in
-  `src/commands.ts` and `src/main.tsx`.
-- **Semantic repo index.** `src/utils/codeIndex/repoIndex.ts` builds offline,
-  dependency-free indexes under `.ur/code-index/`: `repo.json` (file
-  classification + imports/importedBy), `symbols.json`, `calls.json`,
-  `tests.json`, `docs.json`, and `configs.json`. `ur code-index repo` exposes
-  `build|status|search|symbols|callers|tests|docs|configs` subcommands.
-- **`--repo` flag for code-index build/watch.** `ur code-index build --repo`
-  builds both the embedding index and the repo index; `watch --repo` refreshes
-  the repo index on file changes.
-
-### Changed
-- Reused the existing workflow engine for executable skills and extended the
-  existing `ur code-index` surface for the repo index, keeping changes additive.
-
-### Verified
-- Added `test/skillSpec.test.ts`, `test/skillCommand.test.ts`, and
-  `test/repoIndex.test.ts` covering skill parsing, compilation, CLI scaffolding,
-  repo file classification, symbol/call/test/doc/config extraction, and index
-  round-trips.
-
 ## 1.22.5
 
 ### Added
@@ -1742,6 +1824,36 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
   import updates, rollback on check failure, move, organize imports, unused
   detection, and caller mapping.
 
+## 1.22.3
+
+### Added
+- **Executable skill directories.** A `.ur/skills/<name>/` directory containing
+  `skill.yaml` is now an executable skill that compiles into a `WorkflowSpec`.
+  Supports `instructions.md`, `scripts/`, `templates/`, and `checklists/`.
+  `src/skills/skillSpec.ts` parses, validates, and compiles skills; step prompts
+  support `$ARGUMENTS`, `$0..$N`, and `$ARGUMENTS[N]` substitution.
+- **`ur skill` CLI.** New command surface: `ur skill list`, `ur skill show <name>`,
+  `ur skill run <name> [args]`, and `ur skill init <name>`. Registered in
+  `src/commands.ts` and `src/main.tsx`.
+- **Semantic repo index.** `src/utils/codeIndex/repoIndex.ts` builds offline,
+  dependency-free indexes under `.ur/code-index/`: `repo.json` (file
+  classification + imports/importedBy), `symbols.json`, `calls.json`,
+  `tests.json`, `docs.json`, and `configs.json`. `ur code-index repo` exposes
+  `build|status|search|symbols|callers|tests|docs|configs` subcommands.
+- **`--repo` flag for code-index build/watch.** `ur code-index build --repo`
+  builds both the embedding index and the repo index; `watch --repo` refreshes
+  the repo index on file changes.
+
+### Changed
+- Reused the existing workflow engine for executable skills and extended the
+  existing `ur code-index` surface for the repo index, keeping changes additive.
+
+### Verified
+- Added `test/skillSpec.test.ts`, `test/skillCommand.test.ts`, and
+  `test/repoIndex.test.ts` covering skill parsing, compilation, CLI scaffolding,
+  repo file classification, symbol/call/test/doc/config extraction, and index
+  round-trips.
+
 ## 1.22.2
 
 ### Added

package/README.md

@@ -400,7 +400,7 @@ as first-class subcommands in the shipped CLI.
 | `ur task` | Start, run, and hand off worktree-per-task sessions with optional PR creation. |
 | `ur automation` | Store and run project-local scheduled automation specs under `.ur/automations/`. |
 | `ur workflow` | Define, validate, graph, run, and resume declarative agent workflows. |
-| `ur crew` | Run a lead and worker subagent crew over a shared task board; `--decompose` auto-splits tasks with risk/tests/rollback metadata. |
+| `ur crew` | Run dependency-aware worker subagents over a shared task board; independent tasks fan out, while bounded retries require isolated worktrees. |
 | `ur pattern` | Run multi-agent collaboration patterns (PEER, DOE, concurrent, handoff, debate, parallel); `--execute` runs them as a workflow. |
 | `ur goal` | Track long-horizon objectives that persist across sessions. |
 | `ur repo-edit` | Build a repo edit index, plan AST-aware renames, preview patches, and apply with rollback. |
@@ -549,10 +549,11 @@ ur plugin disable <plugin>
 manifest and reports its declared components and capability surface, so you can
 review what a plugin touches before enabling it.
 
-The npm package includes `README.md`, `QUALITY.md`, `docs/`, `documentation/`,
-and `plugins/`, so the npm package page and installed artifact both carry the
-marketplace documentation, core plugins, community staging directory, and
-example plugin template. See [Plugin Guide](docs/plugins.md).
+The npm package includes `README.md`, `QUALITY.md`, `docs/`, `technical/`,
+`documentation/`, and `plugins/`, so the npm package page and installed
+artifact carry the audited technical manual, marketplace documentation, core
+plugins, community staging directory, and example plugin template. See
+[Plugin Guide](docs/plugins.md).
 
 UR also documents the core Cursor-style agent primitives as first-class,
 project-backed features: Agent surfaces (`ur`, `ur agents`, `ur crew`, `ur bg`),
@@ -572,6 +573,7 @@ ur spec verify checkout --kernel
 ur crew create parser-crew --goal "fix the flaky parser test" --decompose --dry-run
 ur crew plan parser-crew --goal "fix the flaky parser test" --decompose
 ur crew run parser-crew --workers 3 --decompose --dry-run
+ur crew run parser-crew --workers 3 --worktrees --max-attempts 2
 ur pattern parallel "refactor login without changing behavior" --execute --dry-run
 ur arena "implement a debounce helper" --agents 3 --judge hybrid --verify "bun test"
 ur agent-ci init default

package/RELEASE.md

@@ -54,15 +54,26 @@ rg -n "Version [0-9]|expected: [0-9]|UR-Nexus v[0-9]" README.md docs documentati
 bun test test/docsCoverage.test.ts test/docsCommands.test.ts
 ```
 
+Run the repository-owned bump script so dependency ranges cannot be changed by
+an accidental text replacement:
+
+```bash
+node scripts/version-bump.mjs <next-version>
+```
+
 Version bump checklist (all versioned release surfaces must move together):
 
 1. `package.json` `version`
 2. `bunfig.toml` `MACRO.VERSION`
-3. `documentation/index.html` version eyebrow
+3. `documentation/index.html` version eyebrow and the expected version in
+   `docs/VALIDATION.md`
 4. `extensions/vscode-ur-inline-diffs/package.json` and its lockfile version
    fields (the VSIX test requires them to match the root package version)
 5. `extensions/jetbrains-ur/build.gradle.kts` `version`
-6. Add a `CHANGELOG.md` entry, then run `bun run build` so `dist/cli.js`
+6. The `MACRO.VERSION` fallbacks in `src/commands/agent-ci/agent-ci.ts`,
+   `src/services/agents/agenticCi.ts`, and
+   `src/services/agents/featureScaffolds.ts`
+7. Add a `CHANGELOG.md` entry, then run `bun run build` so `dist/cli.js`
    embeds the new version (`bun run release:check` verifies all of this).
 
 If `npm whoami` fails, run:

package/bin/ur.js

@@ -38,7 +38,10 @@ const bun = process.env.BUN_BIN || process.env.BUN_EXECUTABLE || 'bun'
 const ollamaModel =
   process.env.OLLAMA_MODEL || process.env.UR_MODEL
 const userArgs = process.argv.slice(2)
-const requiredBun = packageMetadata.engines?.bun ?? '>=1.3.0'
+const requiredBun =
+  typeof packageMetadata.engines?.bun === 'string'
+    ? packageMetadata.engines.bun
+    : '>=1.3.0'
 
 function printBunRuntimeError(detail) {
   const attempted = bun
@@ -51,12 +54,62 @@ function printBunRuntimeError(detail) {
   console.error(
     [
       `UR-Nexus requires Bun ${requiredBun} at runtime because the published CLI bundle is built with --target bun.`,
-      `Could not execute "${attempted}" from ${source}.${detail ? ` ${detail}` : ''}`,
+      `Bun executable: "${attempted}" (resolved from ${source}).`,
+      detail,
       'Install Bun from https://bun.sh, or set BUN_BIN to the absolute path of a Bun executable.',
-    ].join('\n'),
+    ].filter(Boolean).join('\n'),
   )
 }
 
+function parseSemver(value) {
+  const match = value.match(
+    /^v?(\d+)\.(\d+)\.(\d+)(?:-([0-9A-Za-z.-]+))?(?:\+[0-9A-Za-z.-]+)?$/,
+  )
+  if (!match) return null
+  return {
+    core: [Number(match[1]), Number(match[2]), Number(match[3])],
+    prerelease: match[4]?.split('.') ?? [],
+  }
+}
+
+function compareSemver(left, right) {
+  for (let index = 0; index < 3; index += 1) {
+    if (left.core[index] !== right.core[index]) {
+      return left.core[index] > right.core[index] ? 1 : -1
+    }
+  }
+
+  if (left.prerelease.length === 0 && right.prerelease.length === 0) return 0
+  if (left.prerelease.length === 0) return 1
+  if (right.prerelease.length === 0) return -1
+
+  const length = Math.max(left.prerelease.length, right.prerelease.length)
+  for (let index = 0; index < length; index += 1) {
+    const leftPart = left.prerelease[index]
+    const rightPart = right.prerelease[index]
+    if (leftPart === undefined) return -1
+    if (rightPart === undefined) return 1
+    if (leftPart === rightPart) continue
+
+    const leftNumeric = /^\d+$/.test(leftPart)
+    const rightNumeric = /^\d+$/.test(rightPart)
+    if (leftNumeric && rightNumeric) {
+      return Number(leftPart) > Number(rightPart) ? 1 : -1
+    }
+    if (leftNumeric !== rightNumeric) return leftNumeric ? -1 : 1
+    return leftPart > rightPart ? 1 : -1
+  }
+  return 0
+}
+
+function satisfiesBunEngine(version, range) {
+  const actual = parseSemver(version)
+  const minimumMatch = range.trim().match(/^>=\s*(.+)$/)
+  const minimum = minimumMatch ? parseSemver(minimumMatch[1]) : null
+  if (!actual || !minimum) return null
+  return compareSemver(actual, minimum) >= 0
+}
+
 function assertBunAvailable() {
   const result = spawnSync(bun, ['--version'], {
     encoding: 'utf8',
@@ -64,14 +117,33 @@ function assertBunAvailable() {
   })
 
   if (result.error) {
-    printBunRuntimeError(result.error.message)
+    printBunRuntimeError(`Could not execute Bun: ${result.error.message}`)
     process.exit(1)
   }
   if (result.status !== 0) {
     const detail = (result.stderr || result.stdout || '').trim()
-    printBunRuntimeError(detail || `Exited with status ${result.status}.`)
+    printBunRuntimeError(
+      detail
+        ? `Bun --version failed: ${detail}`
+        : `Bun --version exited with status ${result.status}.`,
+    )
     process.exit(result.status ?? 1)
   }
+
+  const installedBun = result.stdout.trim()
+  const satisfies = satisfiesBunEngine(installedBun, requiredBun)
+  if (satisfies === null) {
+    printBunRuntimeError(
+      `Could not validate Bun version "${installedBun}" against engines.bun "${requiredBun}".`,
+    )
+    process.exit(1)
+  }
+  if (!satisfies) {
+    printBunRuntimeError(
+      `Found Bun ${installedBun}, which does not satisfy engines.bun "${requiredBun}".`,
+    )
+    process.exit(1)
+  }
 }
 
 const args =