@kendoo.agentdesk/agentdesk 0.15.2 → 0.15.4

package/CHANGELOG.md

@@ -0,0 +1,48 @@
+# Changelog
+
+All user-facing changes to AgentDesk. Each entry is tagged:
+
+- `[CLI]` — affects the `agentdesk` command-line tool
+- `[UI]` — affects the dashboard at agentdesk.live
+- `[Both]` — affects both
+
+Internal refactors, infrastructure changes, and architectural notes are not listed here.
+
+## [0.15.4] — 2026-04-18
+
+### Added
+- `[Both]` This changelog. Each entry is tagged by audience and scoped to user-facing changes; internal rationale lives separately in the repo.
+
+## [0.15.3] — 2026-04-18
+
+### Changed
+- `[Both]` Agents now verify outcomes with direct observation before reporting "done" / "deployed" / "fixed" / "works" — reading the command output, fetching the live endpoint, loading the page. Intermediate signals (CI green, DB row present) are no longer treated as proof of user-visible outcome.
+- `[Both]` The REVIEW phase actively audits the execution phase's claims and flags unverified ones, returning the team to EXECUTION to produce evidence.
+
+## [0.15.2] — 2026-04-18
+
+### Fixed
+- `[UI]` Project settings page was not showing saved phase-model values after refresh and on the wire returned swapped fields for phase models and update time. Saves were always persisting correctly; only the read path was broken.
+
+## [0.15.1] — 2026-04-18
+
+### Changed
+- `[UI]` Settings page now surfaces save failures inline and re-syncs state from the server after a successful save, so the values you see match what's stored.
+
+## [0.15.0] — 2026-04-18
+
+### Added
+- `[Both]` Phased mode now has five phases: Intake, Plan, Execution, **Review**, **Summary**.
+  - **Review** is a read-only completeness check — it verifies the implementation meets requirements, flags missed documentation updates or silently-deferred scope, and does not modify code. If gaps are found, the team is returned to Execution once before continuing.
+  - **Summary** owns all final tracker writes — task comments, transitioning to "In Review", session protocol. Execution no longer writes the final summary.
+- `[UI]` Project settings expose model pickers for all five phases. Review and Summary default to Haiku 4.5; all other phases keep their existing default. Every phase is overridable.
+
+## [0.14.5] — 2026-04-18
+
+### Added
+- `[CLI]` Daemon startup banner shows the running CLI version.
+
+## [0.14.4] — 2026-04-18
+
+### Changed
+- `[UI]` Phased mode is now the default when starting a session from the dashboard, so per-phase model settings apply without an extra toggle.

package/cli/agents.mjs

@@ -409,7 +409,8 @@ export function generateTeamPrompt(team, opts = {}) {
   rules.push(`1. Each agent speaks prefixed with their badge (e.g., "${team[0].badge} ...").`);
   rules.push(`2. ALL text output MUST be prefixed with the acting agent's badge.`);
   rules.push(`3. Raise concerns when you see problems — don't rubber-stamp.`);
-  let ruleNum = 4;
+  rules.push(`4. **No announcement without observation.** Before any agent claims "done", "deployed", "fixed", "passes", "works", "ready", "now shows X", or "should work" — run the check that would falsify it: execute the command and read the output, fetch the live endpoint and grep the response, load the page and inspect the DOM, verify the deployed artifact contains the new code. "The code looks right" / "the CI is green" / "the DB has the row" are intermediate signals, NOT proof of user-visible outcome. If the observation is out of reach (auth, browser state, external system), say so plainly and ask the user to confirm — don't paper over with confident language. If a fix didn't land on the first try, stop and re-diagnose from scratch: repeated near-misses mean the hypothesis is wrong, not that one more patch will do it.`);
+  let ruleNum = 5;
   for (const a of team) {
     if (a.groundRules) {
       rules.push(`${ruleNum}. ${a.groundRules}`);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@kendoo.agentdesk/agentdesk",
-  "version": "0.15.2",
+  "version": "0.15.4",
   "description": "AI team orchestrator for Claude Code — run collaborative agent sessions from your terminal",
   "type": "module",
   "bin": {
@@ -10,7 +10,8 @@
     "bin/",
     "cli/",
     "prompts/",
-    "README.md"
+    "README.md",
+    "CHANGELOG.md"
   ],
   "engines": {
     "node": ">=18"
@@ -21,7 +22,9 @@
     "server": "node server/index.mjs",
     "build": "vite build",
     "preview": "vite preview",
-    "test": "node --test tests/server.test.mjs tests/agents.test.mjs tests/homepage.test.mjs"
+    "test": "node --test tests/server.test.mjs tests/agents.test.mjs tests/homepage.test.mjs",
+    "lint:changelog": "node scripts/lint-changelog.mjs",
+    "prepublishOnly": "node scripts/lint-changelog.mjs"
   },
   "dependencies": {
     "@radix-ui/react-avatar": "^1.1.11",

package/prompts/phased.md

@@ -419,6 +419,7 @@ Re-read the session memory and the actual implementation (`git diff`, changed fi
 2. **Dennis (engineering):** Does the code match the PLAN? Are there obvious gaps — partially-implemented helpers, dead branches, TODOs left in place, error paths not wired up?
 3. **Sam (quality):** Are there hidden cross-cutting concerns the team missed — docs that should be updated (README, CLAUDE.md, guide/help text), changelog entries, config schema bumps, migration notes, dependent callers?
 4. **Bart (PR):** Is the PR description accurate? Does it reference the task? Are screenshots attached where expected?
+5. **Verification audit (Sam + Dennis):** For every claim the EXECUTION phase made ("deployed", "tests pass", "endpoint works", "migration ran"), confirm it was backed by an actual observation in the session history — a command output, a curl response, a rendered page. If any claim rests on inference ("the code looks right", "CI is green") without the matching observation, that's a finding: mark `NEEDS_MORE_WORK` and list the unverified claim.
 
 After each agent contributes, Jane calls the verdict.