joshidikshant · joshidikshant · Oct 3, 2026 · Sep 5, 2026 · Sep 6, 2026 · Sep 6, 2026
diff --git a/.claude-plugin/marketplace.json b/.claude-plugin/marketplace.json
@@ -5,14 +5,14 @@
     },
     "metadata": {
         "description": "DevSquad - Engineering Manager that coordinates AI coding agents",
-        "version": "0.10.0"
+        "version": "0.11.0"
     },
     "plugins": [
         {
             "name": "devsquad",
             "source": "./plugin",
             "description": "Engineering Manager that coordinates AI coding agents through enforced delegation",
-            "version": "0.10.0"
+            "version": "0.11.0"
         }
     ]
-}
+}
diff --git a/.env.example b/.env.example
@@ -0,0 +1,4 @@
+# Copy to .env, set permissions to 600, and fill the key locally.
+# Never paste credentials into chat or commit them.
+# The Jev probe reads this only with --env-file .env; runtime routing stays off.
+TYPESAFE_API_KEY=
diff --git a/.github/workflows/offline.yml b/.github/workflows/offline.yml
@@ -0,0 +1,68 @@
+name: Offline compatibility
+
+on:
+  push:
+  pull_request:
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+jobs:
+  legacy-bash:
+    name: Legacy Bash 3.2 (macOS)
+    runs-on: macos-latest
+    steps:
+      - uses: actions/checkout@v4
+      - name: Verify system Bash floor
+        run: test "$(/bin/bash -c 'printf "%s.%s" "${BASH_VERSINFO[0]}" "${BASH_VERSINFO[1]}"')" = "3.2"
+      - name: Run legacy and packaging contracts
+        run: /bin/bash test/run.sh
+
+  core:
+    name: Core Python ${{ matrix.python }} (macOS)
+    runs-on: macos-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        python: ["3.11", "3.14"]
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+      - uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python }}
+          cache: pip
+      - name: Install build-only test tooling
+        run: python -m pip install "setuptools>=68" wheel
+      - name: Verify generated command and schema reference
+        run: python scripts/generate-core-reference.py --check
+      - name: Run dependency-free core, package and native-protocol fixtures
+        env:
+          PYTHONDONTWRITEBYTECODE: "1"
+          PYTHONPATH: plugin/core/src
+          PYTHONWARNINGS: error::ResourceWarning
+        run: python scripts/run-core-tests.py --failfast
+
+  optional-mcp:
+    name: Optional MCP 2.2.0 (provider-offline)
+    runs-on: macos-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.11"
+          cache: pip
+      - name: Install exact optional transport lock
+        run: python -m pip install -r plugin/core/requirements-mcp.lock
+      - name: Run MCP SDK and stdio compatibility fixtures
+        env:
+          PYTHONDONTWRITEBYTECODE: "1"
+          PYTHONPATH: plugin/core/src
+          PYTHONWARNINGS: error::ResourceWarning
+        run: python test/core/test_mcp.py -q
+
+# Test bodies make no provider or application network calls. Dependency setup
+# resolves pinned public packages; live provider/app smoke runs remain explicit,
+# bounded receipts outside CI.
diff --git a/.gitignore b/.gitignore
@@ -4,8 +4,18 @@
 .claude/
 .agent/
 
+# Local credentials; commit only the blank example.
+.env
+.env.*
+!.env.example
+
 # Dependencies
 node_modules/
+__pycache__/
+*.py[cod]
+.pytest_cache/
+*.egg-info/
+plugin/core/build/
 
 # Logs
 *.log

diff --git a/AGENTS.md b/AGENTS.md
@@ -0,0 +1,31 @@
+# Working on DevSquad
+
+Read [CONTRIBUTING.md](CONTRIBUTING.md) before changing code. Run
+`bash test/run.sh` before every commit; core changes also need the relevant
+offline Python tests. Preserve the existing Bash 3.2 and optional-jq contracts.
+
+## Continuing the engineering-team build
+
+This build is already underway. After a usage limit, interrupted session or
+handoff, read these files before starting work:
+
+1. [RESUME.md](docs/plans/engineering-team/RESUME.md): latest checkpoint,
+   verified results, open findings and exact next action.
+2. [backlog.json](docs/plans/engineering-team/backlog.json): milestone status
+   and evidence.
+3. [SOL-HANDOFF.md](docs/plans/engineering-team/SOL-HANDOFF.md): the full
+   authorized delivery scope and acceptance criteria.
+
+Compare the recovery note with `git status` and recent commits; retain work
+newer than the note. Continue on `codex/engineering-team` unless the user
+directs otherwise. Do not restart the architecture exercise or discard
+implementation to return to `main`.
+
+Checkpoint coherent partial work and update RESUME.md before long probes or
+handoffs. Record failures and incomplete gates truthfully. Finish a pause with
+a clean tree; never use a stash as the recovery mechanism. Keep raw provider
+diagnostics and credentials outside tracked evidence.
+
+Usage limits do not authorize buying credits, redeeming reset credits, using
+a paid API fallback or changing global AI settings. The saved artifacts must
+allow the next session to resume without relying on conversation memory.
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -4,6 +4,39 @@ All notable changes to DevSquad are documented here.
 
 > **Note:** Project was renumbered from 2.x to 0.x semver in Feb 2026 to reflect pre-stable status. Entries below have been renumbered accordingly.
 
+## [0.11.0] — 2026-10-02
+
+### Added
+- Standalone dependency-free Python core (`squad` 0.1.0) with durable saved
+  runs, isolated candidates, bounded Claude implementation, independent Codex
+  review, candidate-bound checks and fenced host acceptance.
+- Guided setup, readiness, review/fix, status/result, finish, cancel and safe
+  resume commands; optional pinned MCP transport for the same local ledger.
+- Catalog-backed profiles, shared subscription-pool capacity fences, explicit
+  outcome/trial reports and guarded qualification/promotion/rollback.
+- Reversible content-addressed installation and guarded schema-17 upgrade.
+  The wheel packages every runtime schema, including the experimental Council
+  contract. Old schema-16 clients cannot mutate the upgraded ledger.
+
+### Fixed
+- Interrupted guided completion now saves exact intent and recovers only
+  matching authority, retaining prior expiry rejection history.
+- Native probes fail closed without process identity. Bounded natural exit
+  preserves version/auth exit status before owned cleanup; controlled Python
+  fixtures cannot invalidate a pristine candidate with undeclared bytecode.
+- Bash 3.2 compatibility, optional-jq behavior and legacy wrapper error
+  contracts remain supported.
+
+### Limitations
+- Council is experimental and native-unavailable; automatic invocation is OFF.
+  Its fixture comparison is inconclusive, not a quality or savings claim.
+- Jev routing remains OFF. Provider operations depend on installed harness
+  capabilities and subscription authentication; paid API fallback is not used.
+- Claude CLI handoff and Claude-to-Codex delivery have live receipts. Claude
+  desktop Code-tab proof is separate and unverified. Antigravity means the
+  `agy` CLI, not its IDE; Grok/Antigravity MCP status does not prove automatic
+  implementation/review roles. See the runtime guide for supported boundaries.
+
 ## [0.10.0] — 2026-07-07
 
 ### Maintainability & scalability review

diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
@@ -6,6 +6,9 @@
 git clone https://github.com/joshidikshant/devsquad.git
 cd devsquad
 bash test/run.sh    # no network, no real CLIs required — should be all green
+PYTHONWARNINGS=error::ResourceWarning PYTHONPATH=plugin/core/src \
+  python3 scripts/run-core-tests.py
+python3 scripts/generate-core-reference.py --check
 ```
 
 The test suite (`test/run.sh`) is the contract. It runs offline against fake
@@ -54,3 +57,10 @@ Bump the version in `plugin/.claude-plugin/plugin.json` and both entries in
 `.claude-plugin/marketplace.json`, add a `CHANGELOG.md` entry, then after
 pushing run `claude plugin update devsquad@devsquad-marketplace`. Never point
 hook commands at a versioned cache dir — that freezes hooks at install time.
+
+Before a release, also run a fresh standalone install into temporary
+`DEVSQUAD_INSTALL_ROOT` and `DEVSQUAD_BIN_DIR` locations. The tracked installer
+tests cover Claude-free installation, idempotence, actual payload drift,
+plugin contents and an active run surviving a release switch. The operator
+commands and supported/deferred surface boundaries live in
+[docs/RUNTIME-GUIDE.md](docs/RUNTIME-GUIDE.md).
diff --git a/README.md b/README.md
@@ -2,11 +2,13 @@
 
 # DevSquad
 
-### Your AI coding agent ignores your rules. Hooks don't.
+### A local engineering team with saved work and verifiable results
 
-**DevSquad turns Claude Code into an engineering manager that _physically intercepts_ tool calls and routes the grunt work to Gemini, Codex, and Grok — then runs a live A/B test on whether that even helps.**
+**DevSquad coordinates bounded implementation, independent review, tests and
+host acceptance across your installed AI harnesses. Runs and evidence survive
+closing a client. The standalone runtime does not require Claude's plugin.**
 
-[![tests](https://img.shields.io/badge/tests-177%20passing-brightgreen)](test/)
+[![tests](https://github.com/joshidikshant/devsquad/actions/workflows/offline.yml/badge.svg)](https://github.com/joshidikshant/devsquad/actions/workflows/offline.yml)
 [![bash](https://img.shields.io/badge/bash-3.2%2B-blue)](CONTRIBUTING.md)
 [![jq](https://img.shields.io/badge/jq-optional-blue)](CONTRIBUTING.md)
 [![license](https://img.shields.io/badge/license-MIT-black)](LICENSE)
@@ -22,6 +24,48 @@
 
 ---
 
+## Standalone runtime quickstart
+
+Requires a Unix-like host, Python 3.11+ and existing provider CLI subscription
+logins. Installation is local and does not download Python dependencies:
+
+```bash
+git clone https://github.com/joshidikshant/devsquad.git
+cd devsquad
+./install.sh --core-only
+export PATH="$HOME/.local/bin:$PATH"
+squad doctor
+squad review --base main --dry-run
+```
+
+In the project you want to work on, commit the inputs, inspect the dry run,
+then use `squad review --base main --wait` or
+`squad fix "the bounded issue" --write-path src --wait`. The saved run stops
+at a host handoff with checks and the exact finish command; assess the evidence
+before accepting. `squad status`, `squad result RUN_ID`, `squad cancel RUN_ID`
+and `squad resume RUN_ID` operate on the same saved work.
+
+For Codex/Claude/Antigravity CLI/Grok MCP setup, safe updates, supported
+operations and troubleshooting, read the [runtime guide](docs/RUNTIME-GUIDE.md).
+The optional MCP transport needs an explicitly prepared pinned wheelhouse;
+it is not silently downloaded. Existing paid accounts do not imply that every
+harness supports every worker role. No paid API fallback is enabled.
+
+Council is native-unavailable and automatic use is OFF. Jev routing is OFF;
+no production quality, cost savings or Plus-window savings are claimed.
+Claude CLI handoff and implementation → independent Codex review → tests have
+live receipts; Claude desktop Code-tab proof remains unverified. Antigravity
+means `agy`, not the IDE.
+
+Continuing development after an interruption? Read the
+[recovery checkpoint](docs/plans/engineering-team/RESUME.md) and
+[implementation plan](docs/plans/engineering-team/START-HERE.md).
+
+## Legacy Claude plugin
+
+The sections below describe the separately supported Claude hook plugin,
+not automatic worker-role coverage in the standalone runtime.
+
 ## The 30-second version
 
 You told Claude to delegate the boring stuff. It nodded. Then it read 40 files itself, blew through its context window, and you paid for every token.

diff --git a/docs/ARCHITECTURE.md b/docs/ARCHITECTURE.md
@@ -2,6 +2,14 @@
 
 One page for future maintainers (including future Claude sessions).
 
+> **September 2026:** The diagram below describes the legacy Claude plugin;
+> provider names, deployment details and measurements are historical. The
+> [current assessment](audits/2026-09-06-engineering-team-assessment.md) records
+> verified gaps. The proposed shared runner is specified in
+> [ADR-002](adr/ADR-002-surface-independent-engineering-team.md), with a
+> [coding-agent execution packet](plans/engineering-team/START-HERE.md).
+> That new runtime is planned, not implemented.
+
 ## End-to-end flow
 
 ```mermaid
@@ -118,16 +126,26 @@ Enforced by `test/test_wrapper_contract.sh` (offline, fake CLI binaries):
 
 ## Deployment modes (the drift trap)
 
-- **User mode**: `install.sh` registers global hooks pointing at the
-  marketplace clone (`~/.claude/plugins/marketplaces/devsquad-marketplace/plugin`).
-  Refresh with `claude plugin marketplace update devsquad-marketplace` +
-  `claude plugin update devsquad@devsquad-marketplace` after each release.
-- **Dev mode** (this machine): global hooks point at the source checkout, so
-  hooks run at HEAD. Agents/commands/skills STILL load from the installed
-  plugin — after pushing, update the plugin or subagents run stale code.
-- Never let hook commands reference a versioned cache dir
-  (`plugins/cache/.../<version>/`): that froze production at 0.3.0 for
-  five months while fixes accumulated unreleased.
+- **Standalone mode**: `scripts/install-core.sh` creates an immutable release
+  below `~/.devsquad/releases/`, atomically selects it through
+  `~/.devsquad/current` and keeps `~/.local/bin/squad` stable. It needs no
+  Claude installation. `--status --json` compares source, plugin and actual
+  installed payload digests.
+- **Claude plugin mode**: `install.sh` installs the standalone runtime first
+  and, when Claude is available, installs or updates the marketplace plugin.
+  `plugin/hooks/hooks.json` is the hook registration source. The installer no
+  longer writes a duplicate set into global Claude settings.
+- **Dev mode**: run legacy hooks from this source checkout only when explicitly
+  testing changes at HEAD. Agents/commands/skills may still come from an
+  installed plugin, so doctor and install-status output must be checked before
+  attributing behavior to the source tree.
+- Never point a stable launcher or manual hook at a versioned Claude cache
+  directory (`plugins/cache/.../<version>/`). Content-addressed standalone
+  releases are retained for active runs; the selector, not a running process,
+  moves during updates.
+
+See [the runtime guide](RUNTIME-GUIDE.md) for install, operation, recovery and
+the honestly blocked live-surface matrix.
 
 ## State (per project, `.devsquad/`, self-gitignored)