[STEPPING] Delete gates that never fired, and show the planner what it was blind to

.gitignore

@@ -60,10 +60,12 @@ htmlcov/
 # Skill run results (large, regenerated by agent)
 .cursor/skills/*/results/
 
-# PR description scratchpad. Tools (gh CLI, codex) sometimes drop a
-# pr.md at the repo root when drafting a pull request; the
-# description belongs in the PR body, not in the tree.
+# PR description scratchpads and local review notes.
 pr.md
+pr_*.md
+Specialist.discuss.md
+specialist.discuss*.md
+specialist.info.md
 
 # F0/F1 integration: keep test baselines, dry-run merge maps, and
 # archived regression sweeps tracked.

src/hyperloom/agents/kernel/scripts/install.sh

@@ -223,6 +223,19 @@ TRACELENS_MIRROR_DIR="${TRACELENS_MIRROR_DIR:-${_open_source_root}/TraceLens-int
 # missing from env, source $REPO_ROOT/.env but protect already-set values.
 REPO_ROOT="${REPO_ROOT:-$(pwd)}"
 DOTENV="${REPO_ROOT}/.env"
+
+# Vars whose already-resolved value must survive the `. $REPO_ROOT/.env` below.
+# The dotenv is a *credentials* fallback, but it is a full env file: sourcing it
+# under `set -a` re-imports every path var it happens to carry. A stale entry —
+# or one written by a DIFFERENT checkout / workspace sharing this .env — then
+# silently overwrites what the caller resolved, and the run installs against one
+# workspace while writing its env file into another (#Hyperloom-2 / atom mixup).
+# Only non-empty pre-source values are restored, so the dotenv still fills gaps.
+_DOTENV_PROTECTED_VARS='REPO_ROOT KERNEL_AGENT_ROOT HYPERLOOM_KERNEL_AGENT_ROOT
+USER_DATA_PATH HYPERLOOM_RUNTIME_DIR KERNEL_AGENT_ENV HYPERLOOM_ROOT
+MAGPIE_PATH INFERENCEX_PATH TRACELENS_ROOT TRACELENS_INTERNAL_ROOT
+GEAK_ROOT GEAK_E2E_RUNNER PYTHONPATH'
+
 if [ -z "${ANTHROPIC_BASE_URL:-}" ] || [ -z "${ANTHROPIC_API_KEY:-}" ] \
    || [ -z "${ANTHROPIC_AUTH_TOKEN:-}" ] || [ -z "${DEEPSEEK_API_KEY:-}" ] \
    || [ -z "${DEEPSEEK_BASE_URL:-}" ]; then
@@ -232,6 +245,9 @@ if [ -z "${ANTHROPIC_BASE_URL:-}" ] || [ -z "${ANTHROPIC_API_KEY:-}" ] \
     _snap_anthropic_token="${ANTHROPIC_AUTH_TOKEN-}"
     _snap_deepseek_key="${DEEPSEEK_API_KEY-}"
     _snap_deepseek_url="${DEEPSEEK_BASE_URL-}"
+    for _v in $_DOTENV_PROTECTED_VARS; do
+      eval "_snap_prot_${_v}=\"\${${_v}-}\""
+    done
     set -a
     # shellcheck disable=SC1091
     . "$REPO_ROOT/.env"
@@ -241,6 +257,18 @@ if [ -z "${ANTHROPIC_BASE_URL:-}" ] || [ -z "${ANTHROPIC_API_KEY:-}" ] \
     [ -n "$_snap_anthropic_token" ] && export ANTHROPIC_AUTH_TOKEN="$_snap_anthropic_token"
     [ -n "$_snap_deepseek_key" ] && export DEEPSEEK_API_KEY="$_snap_deepseek_key"
     [ -n "$_snap_deepseek_url" ] && export DEEPSEEK_BASE_URL="$_snap_deepseek_url"
+    for _v in $_DOTENV_PROTECTED_VARS; do
+      eval "_snap_val=\"\${_snap_prot_${_v}-}\""
+      if [ -n "${_snap_val}" ]; then
+        eval "_cur_val=\"\${${_v}-}\""
+        if [ "${_cur_val}" != "${_snap_val}" ]; then
+          echo "[kernel-agent] .env would have overridden ${_v}=${_snap_val} with ${_cur_val}; keeping the caller's value" >&2
+        fi
+        eval "export ${_v}=\"\${_snap_prot_${_v}}\""
+      fi
+      unset "_snap_prot_${_v}"
+    done
+    unset _v _snap_val _cur_val
     unset _snap_anthropic_url _snap_anthropic_key _snap_anthropic_token
     unset _snap_deepseek_key _snap_deepseek_url
     echo "[kernel-agent] loaded credentials fallback from $REPO_ROOT/.env (env wins)"

src/hyperloom/agents/robustness/role/envelope.py

@@ -38,6 +38,8 @@ class IntentType(str, Enum):
     RESPONSE = "response"
     REVIEW_VERDICT = "review_verdict"
     KILL_TASK = "kill_task"
+    # Robustness never emits this; kept in the mirror for the contract test.
+    EXTEND_LEASE = "extend_lease"
     PRUNE_BRANCH = "prune_branch"
     ESCALATE_STRATEGY_CHANGE = "escalate_strategy_change"
     # Robustness never emits this; kept in the mirror for the contract test.
@@ -739,14 +741,15 @@ class IntentSpec:
 # Required-field map for intents robustness never emits but the upstream
 # contract test still diffs against. No builder/validator: they are here only
 # to keep :data:`PAYLOAD_REQUIRED` byte-equal with upstream ``_PAYLOAD_REQUIRED``.
-# ``SPECIALIST_DONE`` is the specialist exit envelope (PolicyGate R3 validates
-# it); ``REVIEW_VERDICT`` enforces only the structural ``target_proposal_msg_id``
-# here (verdict/verdict_map mutual exclusion lives in upstream policy).
+# ``SPECIALIST_DONE`` is the specialist exit envelope; ``REVIEW_VERDICT``
+# enforces only the structural ``target_proposal_msg_id`` here
+# (verdict/verdict_map mutual exclusion lives in upstream policy).
 _REQUIRED_ONLY: Mapping[IntentType, tuple[str, ...]] = {
     IntentType.PROPOSE_ACTION: ("action_name", "predicted_gain_pct"),
     IntentType.REQUEST: ("target_agent", "kind"),
     IntentType.RESPONSE: ("in_reply_to", "kind"),
     IntentType.REVIEW_VERDICT: ("target_proposal_msg_id",),
+    IntentType.EXTEND_LEASE: ("task_id", "extra_sec"),
     IntentType.SPECIALIST_DONE: (
         "gap_canonical_id",
         "domain",

src/hyperloom/inference_optimizer/SKILL.md

@@ -400,13 +400,18 @@ Steps for the launching agent:
    lightweight classify from the model's local `config.json` (decoder
    type, attention variant, expert counts, MTP, SWA window) and set
    `"source": "config_classify"`.
-3. **Write the convention file** — once `session_dir` exists, write the
-   profile to exactly `<session_dir>/model_arch.json`: the file named
-   `model_arch.json` in the current session directory. Do not write it to
-   `$USER_DATA_PATH/model_arch.json`; that path is shared by concurrent
-   sessions and can be overwritten by another launch. Include `model_name`
-   (required for the stale-file guard). Set it to the **clean model name**
-   (e.g. `Qwen2.5-7B-Instruct`);
+3. **Write the profile BEFORE launch and point `$HYPERLOOM_MODEL_ARCH_FILE`
+   at it.** The CLI creates `session_dir` and reads the profile in the *same
+   process*, so a file written to `<session_dir>/model_arch.json` after the
+   session dir appears always loses the race — the run seeds
+   `state.model_arch={}` and the profile never reaches any prompt. Write the
+   JSON to a launcher-owned path instead and export
+   `HYPERLOOM_MODEL_ARCH_FILE=<that path>` in the shell that spawns
+   `optimize`; the CLI copies it into `<session_dir>/model_arch.json` for
+   provenance. Use a **session-unique** filename — `$USER_DATA_PATH` is shared
+   by concurrent sessions on WekaFS, so a fixed name races other launches.
+   Include `model_name` (required for the stale-file guard). Set it to the
+   **clean model name** (e.g. `Qwen2.5-7B-Instruct`);
    the guard normalizes launch forms — flat dirs, HF repo ids, and HF hub
    cache `models--org--repo/snapshots/<hash>` paths — so do NOT use the
    snapshot commit hash. All other fields are optional; renderers drop

src/hyperloom/inference_optimizer/cli/model_gate.py

@@ -310,6 +310,37 @@ def _resolve_amd_gpu_type(explicit: str | None = None) -> str | None:
 _MLX_QUANT_MODES = frozenset({"affine", "mlx"})
 
 
+def _read_preseeded_model_arch(arch_path: Path) -> str | None:
+    """Return the pre-seeded ``$HYPERLOOM_MODEL_ARCH_FILE`` text, or ``None``.
+
+    Copies the source into ``arch_path`` so the session keeps its own copy of
+    the profile the run was seeded with. A failed copy is not fatal — the text
+    is still returned and the profile still reaches the prompts.
+
+    Args:
+        arch_path (Path): The in-session ``model_arch.json`` destination.
+
+    Returns:
+        str | None: The profile JSON text, or ``None`` when the env is unset or
+            the file is unreadable.
+    """
+    src = (os.environ.get("HYPERLOOM_MODEL_ARCH_FILE") or "").strip()
+    if not src:
+        return None
+    try:
+        text = Path(src).expanduser().read_text(encoding="utf-8")
+    except OSError as exc:
+        logging.warning("model_arch_preseed_unreadable: %s (%s)", src, exc)
+        return None
+    try:
+        arch_path.parent.mkdir(parents=True, exist_ok=True)
+        arch_path.write_text(text, encoding="utf-8")
+    except OSError as exc:
+        logging.warning("model_arch_preseed_copy_failed: %s -> %s (%s)", src, arch_path, exc)
+    logging.info("model_arch_preseeded_from: %s", src)
+    return text
+
+
 def _load_model_arch(
     workspace_root: Path,
     model_name: str,
@@ -324,6 +355,13 @@ def _load_model_arch(
     ``models--org--repo/snapshots/<hash>`` paths so a declared clean name still
     matches a commit-hash launch basename.
 
+    The session dir is created and seeded in the same CLI process, so a
+    launcher following SKILL Step 1.5 ("write it once ``session_dir`` exists")
+    can never win that race. ``$HYPERLOOM_MODEL_ARCH_FILE`` is the pre-launch
+    escape hatch: point it at a profile written *before* launch and it is used
+    when the in-session file is absent, then copied into the session dir so the
+    run keeps its own provenance copy.
+
     Args:
         workspace_root (Path): Directory containing ``model_arch.json``.
         model_name (str): The resolved model identity (display name / basename).
@@ -340,7 +378,9 @@ def _load_model_arch(
     try:
         raw = arch_path.read_text(encoding="utf-8")
     except FileNotFoundError:
-        return {}
+        raw = _read_preseeded_model_arch(arch_path)
+        if raw is None:
+            return {}
     except OSError as exc:
         logging.warning("model_arch_unreadable: %s (%s)", arch_path, exc)
         return {}

src/hyperloom/inference_optimizer/protocol/intent.py

@@ -33,7 +33,8 @@ class IntentType(str, Enum):
     REQUEST = "request"
     RESPONSE = "response"
     REVIEW_VERDICT = "review_verdict"  # Critic-only
-    KILL_TASK = "kill_task"  # Robustness-only; payload.scope must be "task"
+    KILL_TASK = "kill_task"  # payload.scope must be "task"
+    EXTEND_LEASE = "extend_lease"  # refresh a live task's lease TTL
     # Robustness-only scheduling police.
     PRUNE_BRANCH = "prune_branch"
     ESCALATE_STRATEGY_CHANGE = "escalate_strategy_change"
@@ -61,9 +62,10 @@ class Intent:
     # verdict/verdict_map mutual exclusion enforced by _validate_review_verdict_payload.
     IntentType.REVIEW_VERDICT: ("target_proposal_msg_id",),
     IntentType.KILL_TASK: ("task_id", "reason"),
+    IntentType.EXTEND_LEASE: ("task_id", "extra_sec"),
     IntentType.PRUNE_BRANCH: ("family", "reason"),
     IntentType.ESCALATE_STRATEGY_CHANGE: ("reason", "next_action_hint"),
-    # specialist exit envelope; per-variant schema enforced by PolicyGate R3.
+    # specialist exit envelope; the runner re-stamps and defaults the payload.
     IntentType.SPECIALIST_DONE: ("gap_canonical_id", "domain", "proposal_set", "empty", "summary"),
 }
 

src/hyperloom/inference_optimizer/tests/test_agent_roles_and_policy.py

@@ -59,7 +59,7 @@ def test_orchestration_permissions():
     assert IntentType.PRUNE_BRANCH in role.allowed_intents
     assert IntentType.ESCALATE_STRATEGY_CHANGE in role.allowed_intents
     assert IntentType.REVIEW_VERDICT not in role.allowed_intents
-    assert IntentType.KILL_TASK not in role.allowed_intents
+    assert IntentType.KILL_TASK in role.allowed_intents
     assert IntentType.RESPONSE not in role.allowed_intents
 
 
@@ -123,7 +123,7 @@ def test_review_verdict_critic_only():
 
 
 def test_kill_and_robustness_only_renamed():
-    assert KILL_TASK_SOURCE_ALLOWLIST == frozenset({"robustness"})
+    assert KILL_TASK_SOURCE_ALLOWLIST == frozenset({"robustness", "orchestration"})
     assert ROBUSTNESS_ONLY_SOURCE_ALLOWLIST == frozenset({"robustness"})
     assert ROBUSTNESS_ONLY_INTENTS == frozenset(
         {
@@ -527,16 +527,26 @@ def test_gate_robustness_kill_task_ok(gate):
     )
 
 
-def test_gate_orchestration_kill_task_rejected(gate):
+def test_gate_orchestration_kill_task_allowed(gate):
+    gate.validate_intent(
+        "orchestration",
+        Intent(
+            type=IntentType.KILL_TASK,
+            payload={"task_id": "t1", "reason": "stalled"},
+        ),
+    )
+
+
+def test_gate_orchestration_kill_task_process_scope_rejected(gate):
     with pytest.raises(PolicyDenied) as exc:
         gate.validate_intent(
             "orchestration",
             Intent(
                 type=IntentType.KILL_TASK,
-                payload={"task_id": "t1", "reason": "stalled"},
+                payload={"task_id": "t1", "reason": "stalled", "scope": "process"},
             ),
         )
-    assert exc.value.rule == "role"
+    assert exc.value.rule == "kill_scope"
 
 
 def test_gate_robustness_kill_task_process_scope_rejected(gate):

src/hyperloom/inference_optimizer/tests/test_common_utils.py

@@ -2429,26 +2429,6 @@ def test_llm_prompt_parse_response_edges() -> None:
     assert parse_llm_response("[]") == {"executive_summary": "", "section_narratives": {}}
 
 
-# ---------------------------------------------------------------------------
-# orchestrator.state.orchestration_memory.deterministic_memory_fallback
-# ---------------------------------------------------------------------------
-
-
-def test_deterministic_memory_fallback_bad_gain() -> None:
-    from hyperloom.orchestrator.state.orchestration_memory import deterministic_memory_fallback
-
-    state = SimpleNamespace(
-        current_best={"tput": 123.0},
-        optimization_stack=[],
-        cumulative_gain_validated="not-a-number",
-        phase="EXPLORE",
-        macro_cycle=2,
-    )
-    record = deterministic_memory_fallback(state)
-    assert "current_plan" in record
-    assert "phase=EXPLORE" in record["current_plan"]
-
-
 # ---------------------------------------------------------------------------
 # orchestrator.specialists.profile
 # ---------------------------------------------------------------------------

src/hyperloom/inference_optimizer/tests/test_coordinator_async_batch2_unit.py

@@ -800,9 +800,6 @@ class _Policy:
         def should_checkpoint(self, **kw):
             return False
 
-        def is_hard_compaction(self, n):
-            return False
-
     coord._checkpoint_policy = _Policy()
     assert await coord._maybe_checkpoint_orchestration(tick=5) is False
 
@@ -820,9 +817,6 @@ class _Policy:
         def should_checkpoint(self, **kw):
             return True
 
-        def is_hard_compaction(self, n):
-            return False
-
     coord._checkpoint_policy = _Policy()
     took = await coord._maybe_checkpoint_orchestration(tick=12, phase_changed=True)
     assert took is True
@@ -863,7 +857,6 @@ async def _run(**kw):
 def test_checkpoint_policy_context_fraction_env(tmp_path, monkeypatch) -> None:
     monkeypatch.setenv("USER_DATA_PATH", str(tmp_path))
     monkeypatch.setenv("INFERENCE_OPTIMIZER_CTX_SOFT_FRACTION", "0.5")
-    monkeypatch.setenv("INFERENCE_OPTIMIZER_CTX_HARD_FRACTION", "0.75")
     sd = make_session_dir()
     from .conftest import seed_target_analysis_marker
 
@@ -872,7 +865,6 @@ def test_checkpoint_policy_context_fraction_env(tmp_path, monkeypatch) -> None:
     backends["orchestration"].model = "claude-opus-4-8"  # type: ignore[attr-defined]
     c = Coordinator(sd, backends=backends)
     assert c._checkpoint_policy.context_token_soft == 100_000
-    assert c._checkpoint_policy.context_token_hard == 150_000
 
 
 def test_orchestration_memory_rollback_env(tmp_path, monkeypatch) -> None:
@@ -901,9 +893,6 @@ class _Policy:
         def should_checkpoint(self, **kw):
             return True
 
-        def is_hard_compaction(self, n):  # default: not hard
-            return False
-
     coord._checkpoint_policy = _Policy()
 
 
@@ -947,30 +936,6 @@ async def test_checkpoint_degenerate_three_times_emits_medium_observation(coord:
     assert any(p.get("severity") == "medium" and p.get("consecutive") == 3 for p in degraded)
 
 
-@pytest.mark.asyncio
-async def test_checkpoint_degenerate_hard_uses_fallback(coord: Coordinator) -> None:
-    import time
-
-    _make_conversational(coord, raw_text="still no JSON")
-    coord._checkpoint_enabled = True
-    coord._orchestration_seeded = True
-    coord._run_started_monotonic = time.monotonic()
-    coord.shared_state.phase = "EXPLORE"
-
-    class _Policy:
-        def should_checkpoint(self, **kw):
-            return True
-
-        def is_hard_compaction(self, n):  # near the window → force
-            return True
-
-    coord._checkpoint_policy = _Policy()
-    took = await coord._maybe_checkpoint_orchestration(tick=20)
-    assert took is True
-    assert coord._orchestration_seeded is False
-    assert coord.shared_state.orchestration_memory.get("current_plan", "").startswith("[auto]")
-
-
 @pytest.mark.asyncio
 async def test_checkpoint_failure_resets_tracker(coord: Coordinator) -> None:
     import time
@@ -1712,7 +1677,6 @@ async def test_handle_intent_routes_rare_types(coord: Coordinator, monkeypatch)
         IntentType.PRUNE_BRANCH: "_handle_prune_branch",
         IntentType.ALERT: "_handle_alert",
         IntentType.UPDATE_STATE: "_handle_update_state",
-        IntentType.SPECIALIST_DONE: "_handle_specialist_done",
     }
     for it, attr in routes.items():
 

src/hyperloom/inference_optimizer/tests/test_coordinator_runtime.py

@@ -1367,9 +1367,9 @@ async def runner(ctx):
     assert ran["called"] is True
     assert res.state == "succeeded"
     assert res.result == {"tput": 1.0}
-    assert any(
-        "vanished" in rec.message.lower() and "_transition_resilient" in rec.message for rec in caplog.records
-    ), "expected the disappearing-row warning to fire"
+    assert any("vanished" in rec.message.lower() for rec in caplog.records), (
+        "expected the disappearing-row warning to fire"
+    )
     db.close()