Add Daytona computer-use runtime verification

Author: homanp

README.md

@@ -28,6 +28,7 @@ export FIRECRAWL_API_KEY=...     # optional web search
 export PITHOS_PROVIDER=openai    # any Pi provider
 export PITHOS_MODEL=gpt-5.1
 export PITHOS_SANDBOX_MODE=docker # docker (default) or local
+export PITHOS_COMPUTER_USE=daytona # optional live computer-use backend
 ```
 
 ## Run
@@ -45,12 +46,17 @@ GH_TOKEN=... pithos run https://github.com/owner/private-repo --model gpt-5.5
 pithos run /path/to/repo --model gpt-5.5 --no-advisories
 FIRECRAWL_API_KEY=... pithos run git@github.com:owner/repo.git --model gpt-5.5 --web
 pithos run /path/to/repo --model gpt-5.5 --sandbox-mode local
+pithos run /path/to/repo --model gpt-5.5 --execute-app --computer-use daytona
 ```
 
 Static scan agents use `--sandbox-mode docker` by default. Use `--sandbox-mode
 local` only inside a disposable outer environment such as Daytona, an ephemeral
 CI worker, or a throwaway VM; local mode runs `pi` directly and does not provide
-PITHOS-managed process isolation.
+PITHOS-managed process isolation. `--computer-use daytona` keeps static review
+source-only, then uses a hosted Daytona sandbox with computer-use support for
+live verification when `--execute-app` is set. It overrides the runtime profile's
+sandbox backend for that run, so generated Docker profiles do not need to be
+edited just to try Daytona.
 
 ## Live Verification
 
@@ -91,6 +97,30 @@ pithos run /path/to/repo \
   --runtime-profile /path/to/repo/.pithos/runtime.yaml
 ```
 
+Run live verification in Daytona:
+
+```bash
+export DAYTONA_API_KEY=...
+pithos run /path/to/repo \
+  --provider google \
+  --model gemini-2.5-pro \
+  --execute-app \
+  --computer-use daytona \
+  --runtime-profile /path/to/repo/.pithos/runtime.yaml
+```
+
+Daytona profiles can declare optional computer-use settings:
+
+```yaml
+environment:
+  daytona:
+    snapshot: daytonaio/sandbox:0.6.0
+    record: true
+verification:
+  computer:
+    enabled: true
+```
+
 If setup is incomplete, PITHOS writes `verify/RUNTIME-SETUP.md` with the missing
 env vars, personas, mocks, or profile fields. Live verification does not load
 `.env.local` or other env files; secrets must come from the process environment
@@ -108,6 +138,8 @@ Artifacts are written under `results/<repo>/<timestamp>/`:
 - `verify/environment-summary.json`
 - `verify/RUNTIME-SETUP.md` when live setup is incomplete
 - `verify/runtime-summary.json`
+- `verify/<finding-id>/computer-session.json` for Daytona computer-use sessions
+- `verify/<finding-id>/recording.json` for Daytona screen recording metadata
 - `run-summary.json`
 
 ## Notes

docs/output.md

@@ -40,6 +40,11 @@ recorded in the advisory artifacts and run summary.
   the Pi agent path is used.
 - `verify/<finding-id>/agent-command.json`: command/setup log when a profile-declared
   `verification.agent_command` is used.
+- `verify/<finding-id>/computer-session.json`: Daytona sandbox/computer-use session metadata
+  when `--computer-use daytona` is used.
+- `verify/<finding-id>/computer-events.jsonl`: Daytona computer-use lifecycle events.
+- `verify/<finding-id>/recording.json`: Daytona recording metadata when recording is enabled.
+- `verify/<finding-id>/recording.mp4`: downloaded Daytona screen recording when available.
 
 By default, PITHOS does not execute the target application. Source-oracle and SQL
 checks can still confirm or disprove findings using deterministic inspection.
@@ -76,9 +81,27 @@ verification:
   execute_app: true
 ```
 
-`environment.sandbox` accepts `docker` or `local`. Docker is isolated and remains
-the default; local runs setup commands and live verification agents directly in
-the current environment, so use it only inside a disposable outer sandbox.
+`--sandbox-mode` accepts `docker` or `local`. Docker is isolated and remains the
+default; local runs setup commands and live verification agents directly in the
+current environment, so use it only inside a disposable outer sandbox.
+`--computer-use daytona` keeps static review source-only, then uses a hosted
+Daytona sandbox for live verification when `--execute-app` is set. Profiles may
+also set `environment.sandbox: daytona` for checked-in runtime configuration.
+Passing `--computer-use daytona` overrides the profile sandbox backend for that
+run.
+
+Daytona requires `DAYTONA_API_KEY` and can enable computer-use recording:
+
+```yaml
+environment:
+  sandbox: daytona
+  daytona:
+    snapshot: daytonaio/sandbox:0.6.0
+    record: true
+verification:
+  computer:
+    enabled: true
+```
 
 Secret values should come from the shell, CI secret store, or Cursor-style
 environment secrets. The profile should name variables, not contain secret

pithos/cli.py

@@ -31,6 +31,7 @@
 
 DEFAULT_PROVIDER = "azure-openai-responses"
 DEFAULT_SANDBOX_MODE = "docker"
+COMPUTER_USE_BACKENDS = ("daytona",)
 AZURE_AUTH_MSG = (
     "error: no Azure OpenAI auth found. Set:\n"
     "  AZURE_OPENAI_API_KEY                  (required)\n"
@@ -163,6 +164,12 @@ def main(argv: list[str] | None = None) -> int:
             "outer environment"
         ),
     )
+    p_run.add_argument(
+        "--computer-use",
+        choices=COMPUTER_USE_BACKENDS,
+        default=os.environ.get("PITHOS_COMPUTER_USE"),
+        help="Enable a computer-use backend for live verification (currently: daytona)",
+    )
     p_run.add_argument("--votes", type=int, default=DEFAULT_REPO_VOTES)
     p_run.add_argument("--max-findings", type=int, default=12)
     p_run.add_argument(
@@ -278,6 +285,13 @@ def _cmd_run(args: argparse.Namespace) -> int:
             f"error: invalid sandbox mode {args.sandbox_mode!r}; choose {allowed}", file=sys.stderr
         )
         return 1
+    if args.computer_use and args.computer_use not in COMPUTER_USE_BACKENDS:
+        allowed = ", ".join(COMPUTER_USE_BACKENDS)
+        print(
+            f"error: invalid computer-use backend {args.computer_use!r}; choose {allowed}",
+            file=sys.stderr,
+        )
+        return 1
 
     pi_config_dir = sandbox.resolve_pi_config_dir(args.pi_config_dir)
     agent_env, auth_error = _resolve_agent_env(args.provider, pi_config_dir)
@@ -313,6 +327,7 @@ def _cmd_run(args: argparse.Namespace) -> int:
             "provider": args.provider,
             "model": args.model,
             "sandbox_mode": args.sandbox_mode,
+            "computer_use": args.computer_use,
             "advisories": args.advisories,
             "web": allow_web,
             "execute_app": args.execute_app,
@@ -328,6 +343,8 @@ def _cmd_run(args: argparse.Namespace) -> int:
     print(f"  provider:    {args.provider}", file=out)
     print(f"  model:       {args.model}", file=out)
     print(f"  sandbox:     {args.sandbox_mode}", file=out)
+    if args.computer_use:
+        print(f"  computer:    {args.computer_use}", file=out)
     if pi_config_dir:
         print(f"  pi config:   {pi_config_dir}", file=out)
     print(f"  advisories:  {args.advisories}", file=out)
@@ -369,6 +386,7 @@ def _cmd_run(args: argparse.Namespace) -> int:
                 agent_env=agent_env,
                 pi_config_dir=pi_config_dir,
                 sandbox_mode=args.sandbox_mode,
+                computer_use=args.computer_use,
                 env_audit_report=env_audit_report,
                 event_sink=event_sink,
             )
@@ -402,6 +420,7 @@ def _cmd_run(args: argparse.Namespace) -> int:
             agent_env=agent_env,
             pi_config_dir=pi_config_dir,
             sandbox_mode=args.sandbox_mode,
+            computer_use=args.computer_use,
             preflight_result=runtime_preflight,
             env_audit_report=env_audit_report,
             event_sink=event_sink,