fix(credentials): suppress re-seeding when a pool entry is deleted via API (#55217) (#67429)

hermes_cli/web_server.py

@@ -12498,6 +12498,7 @@ async def add_credential_pool_entry(body: CredentialPoolAdd):
         load_pool,
         PooledCredential,
         AUTH_TYPE_API_KEY,
+        CUSTOM_POOL_PREFIX,
         SOURCE_MANUAL,
     )
 
@@ -12519,6 +12520,23 @@ async def add_credential_pool_entry(body: CredentialPoolAdd):
             access_token=api_key,
         )
         pool.add_entry(entry)
+        # Re-adding a credential is an explicit re-engagement signal: lift
+        # every suppression for this provider so a source deleted earlier
+        # (via DELETE below or `hermes auth remove`) can seed again.
+        # Mirrors the `hermes auth add` behaviour in auth_commands.py.
+        if not provider.startswith(CUSTOM_POOL_PREFIX):
+            try:
+                from hermes_cli.auth import (
+                    _load_auth_store,
+                    unsuppress_credential_source,
+                )
+                suppressed = _load_auth_store().get("suppressed_sources", {})
+                for src in list(suppressed.get(provider, []) or []):
+                    unsuppress_credential_source(provider, src)
+            except Exception:
+                _log.exception("unsuppress after pool add failed (non-fatal)")
+    except HTTPException:
+        raise
     except Exception as exc:
         _log.exception("POST /api/credentials/pool failed")
         raise HTTPException(status_code=400, detail=str(exc)) from exc
@@ -12527,8 +12545,20 @@ async def add_credential_pool_entry(body: CredentialPoolAdd):
 
 @app.delete("/api/credentials/pool/{provider}/{index}")
 async def remove_credential_pool_entry(provider: str, index: int):
-    """Remove a pool entry.  ``index`` is 1-based (matches the list response)."""
+    """Remove a pool entry.  ``index`` is 1-based (matches the list response).
+
+    Removal must be sticky (#55217): ``load_pool()`` re-seeds entries from
+    their backing source (.env var, OAuth singleton file, custom-provider
+    config) on every call, so deleting only the pool row silently reverts on
+    the next dashboard refresh.  We dispatch through the same RemovalStep
+    registry the CLI ``hermes auth remove`` uses: each source cleans up its
+    external state and suppresses ``(provider, source)`` so the seeders skip
+    it.  Manual entries have no registered step — nothing external to clean,
+    no suppression needed (they aren't re-seeded).
+    """
     from agent.credential_pool import load_pool
+    from agent.credential_sources import find_removal_step
+    from hermes_cli.auth import suppress_credential_source
 
     provider = (provider or "").strip().lower()
     try:
@@ -12539,7 +12569,36 @@ async def remove_credential_pool_entry(provider: str, index: int):
         raise HTTPException(status_code=400, detail=str(exc)) from exc
     if removed is None:
         raise HTTPException(status_code=404, detail="No pool entry at that index")
-    return {"ok": True, "provider": provider, "count": len(pool.entries())}
+
+    cleaned: List[str] = []
+    hints: List[str] = []
+    step = find_removal_step(provider, removed.source or "")
+    if step is not None:
+        try:
+            result = step.remove_fn(provider, removed)
+            cleaned = list(result.cleaned)
+            hints = list(result.hints)
+            if result.suppress:
+                suppress_credential_source(provider, removed.source)
+        except Exception:
+            # Cleanup is best-effort, but suppression is the actual bug fix —
+            # without it the entry resurrects on the next load_pool().  Apply
+            # it even when source-specific cleanup blew up.
+            _log.exception(
+                "credential source cleanup failed for %s/%s; suppressing anyway",
+                provider, removed.source,
+            )
+            try:
+                suppress_credential_source(provider, removed.source)
+            except Exception:
+                _log.exception("suppress_credential_source failed")
+    return {
+        "ok": True,
+        "provider": provider,
+        "count": len(pool.entries()),
+        "cleaned": cleaned,
+        "hints": hints,
+    }
 
 
 # ---------------------------------------------------------------------------

tests/hermes_cli/test_dashboard_admin_endpoints.py

@@ -253,6 +253,104 @@ class TestCredentialPoolEndpoints:
         )
         assert r.status_code == 400
 
+    def test_env_seeded_delete_stays_deleted(self):
+        """#55217: DELETE must suppress the source or load_pool() resurrects it.
+
+        load_pool() re-seeds from ~/.hermes/.env on every call, so removing
+        just the pool row silently reverts on the next dashboard refresh.
+        The endpoint must mirror `hermes auth remove`: clean up the backing
+        source and suppress (provider, source).
+        """
+        from agent.credential_pool import load_pool
+        from hermes_cli.auth import is_source_suppressed
+        from hermes_cli.config import save_env_value
+
+        fake_key = "sk-or-" + "x" * 20  # constructed, never a real key shape
+        save_env_value("OPENROUTER_API_KEY", fake_key)
+
+        entries = load_pool("openrouter").entries()
+        assert [e.source for e in entries] == ["env:OPENROUTER_API_KEY"]
+
+        r = self.client.delete("/api/credentials/pool/openrouter/1")
+        assert r.status_code == 200
+
+        # Suppressed exactly like the CLI removal path.
+        assert is_source_suppressed("openrouter", "env:OPENROUTER_API_KEY")
+
+        # Even if the backing var comes back (shell export, another process
+        # rewriting .env), the removal must stay sticky.
+        save_env_value("OPENROUTER_API_KEY", fake_key)
+        assert load_pool("openrouter").entries() == []
+        assert self.client.get("/api/credentials/pool").json()["providers"] == []
+
+    def test_post_readd_lifts_suppression(self):
+        """Re-adding via POST is an explicit re-engagement — suppressions lift.
+
+        Mirrors `hermes auth add`, which clears every suppression for the
+        provider so a user who deleted a credential and re-adds one isn't
+        silently blocked from env re-seeding.
+        """
+        from agent.credential_pool import load_pool
+        from hermes_cli.auth import is_source_suppressed
+        from hermes_cli.config import save_env_value
+
+        fake_key = "sk-or-" + "y" * 20
+        save_env_value("OPENROUTER_API_KEY", fake_key)
+        load_pool("openrouter")
+        assert self.client.delete("/api/credentials/pool/openrouter/1").status_code == 200
+        assert is_source_suppressed("openrouter", "env:OPENROUTER_API_KEY")
+
+        r = self.client.post(
+            "/api/credentials/pool",
+            json={"provider": "openrouter", "api_key": "sk-or-" + "z" * 20},
+        )
+        assert r.status_code == 200
+        assert not is_source_suppressed("openrouter", "env:OPENROUTER_API_KEY")
+
+        # Key back in .env + suppression lifted → env entry seeds alongside
+        # the manual one.
+        save_env_value("OPENROUTER_API_KEY", fake_key)
+        sources = sorted(e.source for e in load_pool("openrouter").entries())
+        assert sources == ["env:OPENROUTER_API_KEY", "manual"]
+
+    def test_manual_delete_adds_no_suppression(self):
+        """Manual entries aren't re-seeded — CLI parity: no suppression marker."""
+        from hermes_cli.auth import _load_auth_store
+
+        self.client.post(
+            "/api/credentials/pool",
+            json={"provider": "openrouter", "api_key": "sk-or-" + "m" * 20},
+        )
+        assert self.client.delete("/api/credentials/pool/openrouter/1").status_code == 200
+        suppressed = _load_auth_store().get("suppressed_sources", {})
+        assert not suppressed.get("openrouter")
+
+        # Immediate re-add works.
+        r = self.client.post(
+            "/api/credentials/pool",
+            json={"provider": "openrouter", "api_key": "sk-or-" + "n" * 20},
+        )
+        assert r.status_code == 200 and r.json()["count"] == 1
+
+    def test_delete_does_not_clobber_other_providers(self):
+        """Deleting one provider's env entry leaves other providers' rows alone."""
+        from agent.credential_pool import load_pool
+        from hermes_cli.auth import _load_auth_store, read_credential_pool
+        from hermes_cli.config import save_env_value
+
+        self.client.post(
+            "/api/credentials/pool",
+            json={"provider": "anthropic", "api_key": "sk-ant-" + "k" * 20},
+        )
+        save_env_value("OPENROUTER_API_KEY", "sk-or-" + "q" * 20)
+        load_pool("openrouter")
+
+        assert self.client.delete("/api/credentials/pool/openrouter/1").status_code == 200
+
+        assert len(read_credential_pool("anthropic")) == 1
+        suppressed = _load_auth_store().get("suppressed_sources", {})
+        assert list(suppressed.keys()) == ["openrouter"]
+
 
 class TestMemoryEndpoints:
     @pytest.fixture(autouse=True)