Merge commit '94062020c767ab7ef12647860e8afed00055b7fd' into torch/provider-runtime

.torch/install-manifest.json

@@ -123,6 +123,20 @@
       "area": "qa",
       "path": "/home/user/TORCHWorktrees/3dc11932-6e3c-42aa-9a8a-ca7b8f818b54/qa",
       "branch": "torch/qa"
+    },
+    {
+      "type": "systemd-user-unit",
+      "name": "torch-3dc11932-6e3c-42aa-9a8a-ca7b8f818b54-schedules.service",
+      "path": "/home/user/.config/systemd/user/torch-3dc11932-6e3c-42aa-9a8a-ca7b8f818b54-schedules.service",
+      "sha256": "d1f6318f91e7b64847e52e415db52c66b60150b0c809fc1e0acb283d8bc0926c",
+      "configDigest": "4e21d6bf76dbcb983d984af2bf44a199ffb56d65e88a5c5f03cd678f196b329f"
+    },
+    {
+      "type": "systemd-user-unit",
+      "name": "torch-3dc11932-6e3c-42aa-9a8a-ca7b8f818b54-schedules.timer",
+      "path": "/home/user/.config/systemd/user/torch-3dc11932-6e3c-42aa-9a8a-ca7b8f818b54-schedules.timer",
+      "sha256": "7c878a828c3f2a04cc597b12de3702f2dfd2ac6ac8228099ada27c59e00151f0",
+      "configDigest": "4e21d6bf76dbcb983d984af2bf44a199ffb56d65e88a5c5f03cd678f196b329f"
     }
   ]
 }

.torch/prompts/COMMON.md

@@ -32,7 +32,10 @@ credentials. Keep pipeline work and Claude/Pi migration queued for later.
 
 Machine manager check-ins are owner-approved hourly with a maximum of 12
 automatic invocations per UTC day in subscription invocation-count mode (not a
-dollar cap). They inspect direct reports/approval waits and report exceptions;
+dollar cap). Automatic AI wake execution is currently PAUSED pending
+TASK-pilot-executor-outcomes qualification; the machine coordination timer still
+queues check-ins. Do not re-enable wakes yourself. They inspect direct
+reports/approval waits and report exceptions;
 they must not launch specialists or expand this slice. Provider-runtime turns are
 launched by the arbiter, not automatically by the manager. Do not change timers
 or budgets yourself. When blocked or finished, record evidence and end the turn;

.torch/torch.yaml

@@ -126,7 +126,7 @@
       "action": {
         "type": "manager-check-in",
         "manager_id": "session-manager",
-        "wake": { "enabled": true, "budget_mode": "invocation-count" }
+        "wake": { "enabled": false, "budget_mode": "invocation-count" }
       },
       "required_authority": [
         "owner"

docs/TODO.md

@@ -4,7 +4,10 @@
 
 TORCH is installed and healthy at `0.1.0-alpha.2`. Setup/restore and all seven
 worktree briefs have been verified. The active organization is Session Manager
-plus six specialists; all providers are offline. The default Git branch remains
+plus six specialists. A first bounded development pilot is now active through
+Session Manager, Provider Runtime and QA; the remaining domains have not started.
+See `reports/audit/first-development-pilot-2026-09-30.md` for current live state.
+The default Git branch remains
 legacy TORCH; the configured development/integration boundary is
 `rewrite/portable-agent-fleet`. No publication or default-branch promotion is
 authorized. Later sections retain historical audit snapshots, including old
@@ -54,10 +57,21 @@ dependencies, and remain unassigned for manager triage. See
   No COMBATRIG schedules, Vercel deployment, credentials, LAN exposure or automatic
   phone notifications are assumed.
 
-These are proposed development tasks, **not implemented automatic setup**. The
-existing 15-minute manager schedule is configured, but no TORCH host timer is
-installed. Common instructions still authorize only the smoke test; development
-activation, self-claim and recurring provider wakes remain separately reviewed.
+These are proposed development tasks, **not implemented automatic init/install**.
+For the first pilot the arbiter explicitly installed an hourly machine manager
+coordination timer and verified real systemd parsing/activation. Automatic AI
+wakes are paused pending the timeout/outcome guard; self-claim remains disabled.
+COMMON.md authorizes only the diagnostics development slice, not broad autonomous
+dispatch. The original 34 proposals remain the handoff; three newly observed
+pilot reliability issues have also been added to the same backlog:
+
+- [ ] `TASK-pilot-starting-presence`: publish guarded startup state before a
+  blocking provider launch and prevent duplicate manual/timer execution.
+- [ ] `TASK-pilot-systemd-rendering`: native scalar-path rendering is fixed and
+  qualified in source; finish active/failed unit evidence and false-success
+  handling instead of equating installed files with an operational timer.
+- [ ] `TASK-pilot-executor-outcomes`: zero exit plus timeout/error/signal cannot
+  produce a successful lifecycle/wake receipt; qualify before automatic wakes.
 
 ## First real self-host run (2026-09-30)
 

reports/audit/first-development-pilot-2026-09-30.md

@@ -0,0 +1,96 @@
+# First TORCH-building-TORCH development pilot
+
+## Owner authorization and live monitor
+
+The owner requested a staged self-development fleet, with the outer assistant
+as arbiter through Session Manager, then a quality release/update and a later
+Claude Code/Pi substitution test. COMMON.md records the bounded first slice:
+`TASK-conversation-runtime-diagnostics`. Other development work, remote pushes,
+landing, public release, deployment, default-main promotion and provider migration
+are not authorized by this pilot.
+
+Monitor: **http://127.0.0.1:4174/console**. The snapshot is installed mode for
+project `3dc11932-6e3c-42aa-9a8a-ca7b8f818b54`, not demo state. The original
+Console process died during an outer-daemon restart. It was restored as the
+loopback-only transient user service
+`torch-3dc11932-6e3c-42aa-9a8a-ca7b8f818b54-console-pilot.service`, outside the
+assistant process lifetime. This is not yet a reboot-persistent installed
+Console integration; automatic service provisioning remains backlog work.
+
+## Real agent evidence
+
+- Session Manager resumed the same Sol runtime and assigned the real task to
+  Provider Runtime with revision-checked proposed → ready → assigned transitions.
+- Provider Runtime started as the configured Terra model, retained runtime
+  session `01a0f3e1-9106-72c3-b8b2-2f1cc5322ccc`, queried ownership and requested
+  QA consent before editing tests. It committed production changes at
+  `c362ccd5cc8fff94c59d4a0c8f178628dc5e7596` on its own branch. That is not landed
+  or accepted release evidence.
+- QA ran real Terra coordination turns and gave exact additive-test consent for
+  `test/fleet/lifecycle.test.mjs` and `docs/TEST_INTEGRITY.md`, retaining existing
+  assertions and requiring independent exact-candidate review.
+- A daemon restart interrupted implementation. The arbiter checked owned live
+  Codex process absence, preserved dirty code/tests/generated screenshots and
+  resumed the same assignment via a transient `runtime-pilot-turn` user service.
+  No unrelated process was killed or user work reverted.
+
+Provider-reported focused checks passed, but its generic full-suite failures and
+final regression evidence still require exact-source qualification. Do not infer
+completion from a commit, active process or green controller receipt. Generated
+test screenshots remain agent work and must not be blindly committed as changed
+expectations. First launch presence lag and assigned-versus-working task state
+need review; startup badges are not independent process-liveness proof.
+
+## Machine scheduler: qualified setup, guarded AI execution
+
+The owner-approved hourly check-in policy retains a maximum of 12 automatic
+invocations per UTC day in subscription invocation-count mode, not a USD cap.
+The project-owned persistent systemd timer was explicitly installed for the
+pilot. This manual qualification does not implement automatic init/install yet.
+
+Initial live install exposed invalid `WorkingDirectory="/absolute/path"` output.
+The actual parser rejected it, while file-based status said installed and the
+timer was enabled but inactive. Minimal source repair `d636cae` emits an absolute
+scalar path with C escaping and literal specifier handling. A new native-parser
+scenario includes spaces, percent and quote characters. Fourteen scheduling tests
+and all 39 source acceptance groups PASS; no existing test was weakened.
+
+Reconciliation used the tested source CLI and preserved owned-file hashes. The
+timer is active and its first dispatch produced a durable check-in and wake
+receipt. Installed stable tool remains alpha.2; the repaired scheduler currently
+executes the source CLI, not a newly packaged stable release. A later qualified
+tool update must incorporate the repair and reconcile the launcher.
+
+The first automatic manager invocation lasted exactly the 300-second executor
+limit and was recorded succeeded/invoked. Independent model-turn completion was
+not established. A native local child trapping SIGTERM reproduced
+`status: 0`, `signal: null`, `error: ETIMEDOUT`; current startFleet ignores the
+error when exit status is zero. This confirms a false-success boundary, not
+independent proof of how that particular model turn ended. Historical receipts
+were preserved, not rewritten.
+
+Automatic AI wakes are therefore **disabled** until `TASK-pilot-executor-outcomes`
+is fixed and qualified. The timer remains active for durable machine coordination;
+the arbiter explicitly drives the bounded specialist turns. Do not claim the
+fleet is ready for unattended autonomy.
+
+## Backlog and next gates
+
+The original 34 conversation follow-ups are preserved. Three observed pilot
+issues were added through audited owner intake: starting presence, native unit
+rendering/operational status and executor timeout/outcome semantics. The manager
+backlog remains the authoritative queue, separate from this narrative report.
+
+1. Finish diagnostics regression evidence and independent QA review.
+2. Fix executor outcome/launch-presence guards and qualify one real bounded wake.
+3. Complete reviewed automatic init/install onboarding and persistent operations.
+4. Expand the fleet's scope only after a successful bounded work/verification/
+   landing/resume cycle; never directly push competing specialist branches.
+5. Build/accept/update a quality artifact with current receipts, then separately
+   test Claude/Pi handover of the same durable identities, tasks, messages and
+   checkpoints. No model/provider swap occurred in this pilot.
+
+Receipts are under `/tmp/torch-pilot-*`, including repaired full-source acceptance,
+native parser tests, scheduler installation/reconciliation, durable wake/run
+state and owner-request messages. The live dashboard and actual branch state
+take precedence over historical progress descriptions in this report.